Skip to content

journal: the fly drives the WaveShare (#6) - #261

Merged
superposition merged 3 commits into
mainfrom
journal/fly-drives-waveshare
Sep 12, 2026
Merged

superposition merged 3 commits into
mainfrom
journal/fly-drives-waveshare

Conversation

@superposition

@superposition superposition commented Sep 12, 2026 •

Copy link
Copy Markdown
Owner
agent: JournalReframeFlyDrives
branch: journal/fly-drives-waveshare
commit: 5560926
state: review
next: rebase onto the main carrying docs/mermaid-diagrams, then re-run the gates and ask the editors to re-read
blocked_on: holding for the docs/mermaid-diagrams PR (convention amendment) to merge
evidence: docs/figures/the-fly-brain-on-the-robot/ (drive-commands.svg, dataset-gate.svg, wave-numbers.json), 'C:\Users\ericm\.omp\agent\sessions\-qualia\2026-09-11T16-05-44-028Z_01a09137-bbdc-7554-abb8-c297e3492ccc\local\qualia-swarm-context.md'

Reframes the EPIC-06 entry (docs/figures/the-fly-brain-on-the-robot/README.md, #6) so its subject is the
robot and the fly driving it, not the hardware the work ran on.

Voice: the first-person account is operator-requested (as recorded for the merged #259); the piece is
"we" throughout except one "I", which is the operator's own judgement about the motor path's safety
properties.

  • The robot: a WaveShare UGV that leash owns end to end (drive on /dev/ttyTHS1, LD06 lidar, MJPEG
    camera, IMU), with a motor path that is safety-shaped by construction — DriveCommand::WheelSpeeds resolved
    from a NavGoal, the link claimed at zero speed first, ACTION_SAFETY_DISARMED / ACTION_SAFETY_TRANSPORT_ERROR
    instead of silence.
  • The fly: the released male-CNS connectome as a spiking controller producing left/right/hold from the
    robot's own camera, with each run's own command counts (first board run 197 L / 99 R / 4 hold of 300; the
    fixed-reader re-run 98 / 196 / 6; host 130 / 66 / 4 of 200). crates/fly-circuit appears only as what it
    is — a rate model with invented dynamics, observe-only, never on the motor path.
  • The learning: transitions -> dataset -> JEPA train_batch, then the honest frontier — no plasticity in
    the network, and no checkpoint passes D-027 (1,499 valid transitions vs 50,000 / 12 / 3 / 3).
  • Figures: drive-commands.svg replaces step-rates.svg and is re-counted from the committed trace CSVs
    (the generator refuses on disagreement); dataset-gate.svg stays; the loop diagram is a Mermaid flowchart
    block in the entry itself, per the operator. No logo, mark or hero render anywhere.

Gates at d2287e4: qualia-gates figures -> figures: OK (11 entries, 53 figures, 400 KiB budget);
qualia-gates provenance -> provenance: OK.

Base: rebuilt from a clean base (parent = current main, 4b36422) after accuracy caught a rewound
rebase; the branch now differs from main by exactly the ten files of this entry, and figures/provenance
pass at this head.

Refs #6

@superposition

Copy link
Copy Markdown
Owner Author
role: style
verdict: request-changes
notes: 4

Re-read of the live head b087e992 (the PR body's braid names c432a29; the branch has moved two
commits since — 41aa898, b087e99, both in the entry only). The four style legs mostly hold; two of
my notes are about the Mermaid drawing, which is the one departure this PR makes from the figures
convention.

Style checklist (docs/journal-review.md, copied verbatim)

  1. The site's form: **The claim.** first, then ## What we tried, ## Evidence,
    ## What this does not establish. — pass. # The fly brain on the robot (1) →
    **The claim.** (3) → ## What we tried (34) → ## What we changed (129) → ## Evidence (162) →
    ## What this does not establish (310). All four template H2s present, in order. Machine check:
    qualia-gates figures → figures: OK (11 entries, 53 figures, 400 KiB budget), exit 0;
    journal --entry → journal-gate: entry OK, no <!-- ASK: -->.
  2. Short paragraphs, no filler. — pass, with note 4. Claim 193 words; no filler register, no
    second person, no marketing. The glossary paragraph is 234 words (l. 15) and sits in the claim slot.
  3. Front matter valid for the site's build. — not satisfiable in this tree. No --- block, as in
    the merged entry; the closing bullet states the site step as a fact. Publish-step work.
  4. Figures and band referenced by absolute URL. — pass for the committed figures, note 1 for the
    drawing.
    The two charts are the absolute
    https://raw.githubusercontent.com/superposition/qualia/main/docs/figures/the-fly-brain-on-the-robot/<name>.svg
    form (ll. 222, 234), with their .png companions in the table. The loop diagram is an inline
    fenced mermaid block (l. 171), not a figure URL — fine as a form, but see notes 1 and 2.

Notes

  1. The entry misstates the figures convention, and the directory no longer satisfies rule 1 as
    written.
    The Regenerating paragraph (ll. 291–295) says: "docs/figures/README.md names it [a fenced
    mermaid block] as the drawing form alongside a _house.py schematic, a render, or a Blender
    asset." It does not. Rule 1 (l. 33–35) names exactly two forms: "a schematic drawn in _house.py's
    box/arrow vocabulary, or a Blender render from steps 42–43". So docs/figures/the-fly-brain-on-the-robot/
    now ships two charts and no drawing by the convention's own definition. The Mermaid choice can stay —
    the root README.md carries four mermaid blocks and docs/architecture/*.mmd are the repository's
    flow form, so it is a defensible drawing — but the fix belongs in this PR: either add a fenced
    mermaid block to rule 1's list (one line in docs/figures/README.md, which is what makes the
    choice legitimate and the sentence true), or commit a _house.py schematic/render and correct the
    sentence. As it stands the sentence is false against a file the entry cites twice.
  2. The figure table makes README.md a declared figure. Row 3 (l. 219) is
    | `README.md` (a Mermaid `flowchart` block, inline) | diagram, not a committed SVG | … |.
    docs/figures/README.md says the table's File cell is "the committed path in backticks" and the
    first table is "one row per file" of the entry's pictures; this row's file is the entry itself. It is
    also the only reason figures passes rule 1 with no committed drawing — the checker sees kind
    diagram and a path that exists. That is honest labelling, not a hidden pass, but it means the
    machine-green result for rule 1 currently rests on the entry README rather than on an asset. Note 1's
    fix resolves this too; if the convention keeps Mermaid out of rule 1, the row should go and a real
    drawing should take its place.
  3. The journal row's shape changed. README.md:244 is now
    | the fly brain on the robot | [EPIC-06](…) — the fly driving the WaveShare UGV |, while all ten
    other rows are | entry | epic links | with nothing after the links. The tagline duplicates the
    entry's subject and is the only prose in the table; check_journal still passes (the epic link is
    absolute and the min-epic order holds), so this is a shape consistency call rather than a gate
    failure. I would drop the trailing clause, or add one to every row — not one of eleven.
  4. Three first-person-singular uses, where the benchmark entries have none. "I care about those
    properties more than the control law" (l. 55), "instead of by constants I chose" (l. 103), "I am not
    calling that a behaviour" (l. 228). The merged journal: rewrite the fly brain on the robot for a reader #259 revision kept exactly one "I" and I accepted it
    as operator-requested; the two published entries I compare against (the-ladder,
    the-connectome-as-a-prior) use none, and the house first person is "we", sparingly. If the author's
    voice is again operator-requested here, one line in the PR body recording that closes this note; if
    not, these three become "we". (Related, non-blocking: the 234-word glossary paragraph at l. 15 is
    above the 162/176-word longest paragraph of those two entries — trimming it would also shorten the
    claim slot, which already carries the 193-word claim plus a 28-word notation pointer before
    ## What we tried.)

What passes, so the revision can be narrow

Claim first and the four H2s in order; the two committed charts by absolute URL with their .png
companions and per-run captions; the Evidence/Data tables with units on every quantity; sources
named on every claim I sampled (docs/evidence/…, decisions, issue numbers); British spellings
("behaviour", "coloured"); no second person (you/your = 0); no emoji (the file's non-ASCII is em dash,
×, en dash, §, …, →, ↔, ≥ only); no draft marker (TODO|FIXME|TBD|placeholder|ASK:|WIP
returns nothing); figures and the journal entry leg both green. The loop diagram is a real flow and
the entry explains why it is Mermaid rather than a third SVG — the explanation just needs to match the
convention.

What I could not verify

Whether the operator's instruction was specifically "a Mermaid block" or "no third SVG figure" (I did
not see it; I infer from the PR body); the rendered Mermaid output (I read the block source, not the
render); and front-matter validity (no site tree in this repository).

@superposition

Copy link
Copy Markdown
Owner Author
role: teaching
verdict: request-changes
notes: 4

Teaching verdict — PR #261, d2287e4

One consolidated teaching comment: Main dispatched two teaching reviewers to this PR, and this is my verdict for the role (the gate takes the last comment per role, so it supersedes any parallel teaching comment posted for this head).

Checklist (docs/journal-review.md, role teaching)

  1. The entry explains the idea before the result. Met, and the reframe earns it. The claim establishes the robot, the leash's ownership of it and why that ownership makes a brain safe to attach before the connectome or any number; the glossary follows; the sections then keep robot → fly → learning. The new sentence separating the connectome's read-out from qualia-drive's WheelSpeeds is a real repair — before it, the entry read as if the fly's output already reached the motors.
  2. At least one chart and at least one diagram or 3D render, each captioned so a reader can interpret it without the text. Not met as shipped: the drive-commands chart attaches its run labels to the wrong bars (N1), and the Mermaid loop diagram draws a path the entry's own prose says does not exist (N2). The dataset-gate.svg chart is correct and well captioned. The Mermaid-as-drawing question is a convention dependency the entry now states honestly (rule 1 today names a _house.py schematic or a Blender render); the teaching item only needs a captioned diagram, and the Mermaid is one, so that dependency is not what fails this item.
  3. Jargon defined on first use. Mostly met, and the glossary is the strongest part of the rewrite (neuron, synapse, connectome, spiking, membrane potential, spike, refractory, tick, leash, UGV, MCAP, JEPA, promotion gate, plasticity, neuromodulation, plus the codes key). Residual: deadman (N4).
  4. The analogy named where one exists. Met: "the wiring diagram" for the connectome, and "a brain safe to attach" for the point of the leash's ownership.
  5. A reader who stops after the first paragraph still learns something true. Not met as written. The claim's own wording is careful ("produced the steering commands that would reach the wheels … stopped at a trace file rather than a wheel"), but one of its three distributions is the first board run, which the entry itself disowns twice later (N3); a reader who stops there takes a number the entry does not stand behind.

Notes

N1 — drive-commands.svg labels the wrong runs (blocking, item 2). make_figures.py:98 draws for pos, entry in enumerate(reversed(runs)), but :134 sets ax.set_yticklabels([entry["run"] for entry in runs]) from the unreversed list and :135 inverts the axis. The committed SVG (sha256 d6ec8061…, byte-identical at c432a29, d5bbe87 and d2287e4) therefore puts "board loop, first run" over the host run's bar — 130/66/4, "196 of 200 ticks commanded", 112.7 ticks/s — and "host loop, the board's camera" over the first board run's — 197/99/4, "296 of 300", 31.3 ticks/s. I rendered the committed SVG and read the labels back; it says the first board loop ran at 112.7 ticks/s, the one number that run cannot have. Fix the pairing (label from reversed(runs) too, or drop the reversal), regenerate, and check the label↔bar pairs — the caption and the Data table have the runs right, so only the drawing moves.

N2 — the loop diagram teaches a conversion the entry says does not exist (blocking, items 1/2). The Mermaid has READ -->|steering| DRIVE, and its caption reads "the connectome's spikes become a left/right/hold decision, and qualia-drive turns that into a wheel command on a link the leash owns". Three paragraphs earlier the entry now says the opposite: "That left/right/hold is the connectome's own read-out, not yet the robot's motor path: qualia-drive writes DriveCommand::WheelSpeeds from a NavGoal, and nothing yet converts the fly's command into one." The prose is the correct account; the diagram and caption need to agree with it — e.g. READ -. "left/right/hold, not yet converted" .-> DRIVE, with a caption that says the drive runner's WheelSpeeds path exists but is fed by a NavGoal. While there, reconsider DRIVE -->|WheelSpeeds| MOTORS as a solid edge: the limits say no wheel turned and the leash holds the port, so the solid arrow currently reads as an exercised path.

N3 — the first board run is drawn as data and disowned in the same entry (item 5). The claim, the chart, the Data table ("Command distribution") and the chart caption all carry the first board run's 197/99/4, while the failures say "Because that first run's numbers came off the broken reader, the entry quotes both runs and labels which is which" and the limits say "the first run's numbers came from a reader later found to be broken". T55's own record is narrower than that: docs/evidence/T55/connectome-runner/board-bench.log's header says its loop section is "a FAILED read-back … not the loop's stdout", the recording is "recoverable byte-exactly from fe8c4f4" (60,681,968 B = 8 + 300*20 + 4*15,168,990), and board-loop-trace.csv is the committed per-tick trace the chart re-counts. So the first run's trace is real and the read-back of its recording is what failed. Say that precisely — then either keep both runs labelled ("trace real; its recording was never read back") or drop the first run's bar. As it stands a reader cannot tell whether the chart's first bar is evidence or the thing the failures section warns about.

N4 — deadman is undefined where the entry leans on it (item 3). "its own deadman at 400 ms (LEASH_DEADMAN_MS)" carries the claim's safety argument — the leash can stop the motors — and the term recurs in the Mermaid ("deadman 400 ms") and the Data table. Half a clause ("a heartbeat the leash must receive or it stops the motors") closes it. Optional: "the shared arena" (one use in the motor-path paragraph) and MJPEG (the camera's stream format) are the only other terms that arrive unglossed.

What a reader would still ask

  • What turns the fly's left/right/hold into a WheelSpeeds? Named as the gap by the new paragraph, and the limits say the motor authority belongs to the leash, so it is a leash decision rather than a runner flag. Answered.
  • Why a fly? Answered: the bet on structure as the controller, the five-type/nine-edge stand-in it replaces, and the ventral nerve cord's motor neurons as the path from a frame to a decision.
  • How far is the learning loop? Answered and quantified: 1,499 valid of 50,000, 1 session of 12, 1 environment of 3, 1 condition of 3.

What I checked

The entry at d2287e4 read end to end, plus the c432a29 → d2287e4 diffs; make_figures.py and wave-numbers.json for N1, and the committed drive-commands.svg rendered with its labels read back; the Mermaid block and caption against the prose for N2; board-bench.log, the T55 README's file list and the #236 braids for N3; first-occurrence search for deadman / shared arena / MJPEG for N4. I did not run the generator or the gates — the author's own gate output is quoted in the entry, and the numbers are the accuracy editor's leg.

agent: Journal258Teaching
branch: journal/fly-drives-waveshare
commit: d2287e4
state: teaching request-changes (4 notes)
next: author fixes N1 (chart label pairing) and N2 (diagram/caption vs prose), and states N3 (first run: trace real, recording not read back) and N4 (deadman); teaching re-reads after the rebase
blocked_on: none
evidence: this comment; make_figures.py:98,134,135; drive-commands.svg rendered (sha256 d6ec8061…); docs/evidence/T55/connectome-runner/board-bench.log header; the entry's Mermaid block and caption

@superposition

Copy link
Copy Markdown
Owner Author
role: accuracy
verdict: request-changes
notes: 3

Checklist (copied verbatim from docs/journal-review.md §accuracy) — read on d2287e4

  1. Every number traced to a source in the entry itself. — holds for the prose and the Data table; the three notes below are one figure that encodes the wrong row labels (note 1) and two sentences about the provenance of numbers (notes 2–3), not untraced values.
  2. Units on every quantity. — holds: rates carry ticks/s with their regime (free-running, closed loop, host loop), the serial facts carry baud and ms, the floor carries its five numbers, and the chart labels say "ticks of one closed-loop run".
  3. Any value measured in a different session or on different hardware labelled as such. — holds: the Orin NX row is the robot's machine, the RTX 4090 and i9-14900KF rows are marked — supporting, the 60 Hz row is — bench only, and the two board runs are labelled first run / fixed-reader re-run everywhere.
  4. Every figure's caption says what it encodes. — fails for drive-commands.svg: the caption's run→count mapping is right, but the chart's row labels are swapped (note 1). The Mermaid drawing's caption is also out of step with the prose it sits beside (note 3). dataset-gate.svg is unchanged from the version I traced and its caption still holds.
  5. No claim about a private repository. — holds; every source named is in this repo.

The three things you asked me to check, and what I found.

The command counts, re-counted from the CSVs, not read from the prose. I counted the command column of each committed trace:

$ awk -F, 'NR>1 {c[$6]++} END {for (k in c) print k, c[k]}' <trace>
host-loop-gpu-trace.csv       -1: 130, +1: 66, 0: 4   (200 ticks) → 196 non-hold
board-loop-trace.csv          -1: 197, +1: 99, 0: 4   (300 ticks) → 296 non-hold
board-loop-trace-rerun.csv    -1:  98, +1: 196, 0: 6  (300 ticks) → 294 non-hold

Every split in the claim, the caption, the Data table and wave-numbers.json's command_runs is exact. The generator refuses on a mismatch and re-runs green, so the recording and the traces agree by construction. Luminance too: the host run's 0.4613–0.4679 and the first board run's 0.4662–0.4739 are the min/max of those two CSVs' second column.

The robot and leash facts, against the tree. runners/drive/src/main.rs: DEFAULT_PORT "/dev/ttyTHS1", DEFAULT_BAUD 115_200, DEFAULT_TICK_MS 100, POSE_STALE_NS/LIDAR_STALE_NS 1.5e9, MAX_WHEEL_CMD 0.34, MAX_TURN_CMD 0.20, FRONT_STOP_DIST_M 0.48, MIN_PROGRESS_M 0.10, PROGRESS_TIMEOUT_SECS 5, ESTOP_FILE, speed_line() → {"T":1,"L":…,"R":…} newline-terminated, parse_armed() accepts exactly "1" | "true" | "TRUE" | "yes" | "YES", the zero-speed claim at startup under if config.armed, and ACTION_SAFETY_DISARMED / ACTION_SAFETY_TRANSPORT_ERROR in action_safety_flags(). arbitrate() is the precedence order the entry states: estop → stale_pose → stale_lidar → idle → obstacle_turn → progress_stall → goal_reached/tracking. D-025 and T58/T59 carry the leash facts the entry uses (LEASH_PROFILE=waveshare-ugv, :8000, deadman 400 ms, /dev/ttyACM0 LD06 / /dev/video0 camera / /dev/ttyTHS1 drive, errno 16, fd=3 on the camera, exit 2). The read-out rule is crates/connectome-cns/src/main.rs:661-674: command = sign(rate_r − rate_l) against DECISION_MARGIN 0.001, throttle the total rate — and the decoder's descending_neuron/vnc_motor superclasses are the release's, as the entry says. frame 83464099482600 ids are not strictly ascending; board-loop.log:439 really is the 300 ticks in 9.358 s (32.1 ticks/s), 15085686 spikes, 294 non-hold commands line; crates/jepa-model/src/train.rs:169 really is train_batch. The D-027 floor (50,000 / 12 / 3 / 3, each ≥ 5 %, 33.4× = 50,000/1,499) is as D-027 states, and the trainer transcript in the entry is D-027's.

The 4090/Nano material. It is only in the Data table's two rows marked — supporting (RTX 4090, i9-14900KF) plus the — bench only 60 Hz row, and the entry says so in "What is deliberately not here". No 4090 or host-CPU number appears in the claim, the figures or the captions.

Reproduced at this head: qualia-gates figures → figures: OK (11 entries, 53 figures, 400 KiB budget); qualia-gates provenance → 152 authored file(s) … 0 identical, 27 EOL-identical, 7 code runs over 20, 0 prose runs over 2 and provenance: OK; both charts regenerated from this head's generator are byte-identical to the committed blobs.

Notes.

  1. drive-commands.svg draws each run's counts under the wrong row label (top and bottom are swapped); the caption and the Data table are right. make_figures.py:98 iterates enumerate(reversed(runs)) — so pos 0 is the host run — but :134 labels the ticks from runs in order (board loop, first run, board loop, fixed-reader re-run, host loop, the board's camera), and :135's invert_yaxis() puts y=0 at the top. I reproduced the geometry in matplotlib: after invert_yaxis() the y-limits are (2.54, −0.54), the host bar sits at y=0 (top) and the first-run bar at y=2 (bottom), while the tick at y=0 is labelled board loop, first run. Read off the committed PNG, top to bottom: the row named "board loop, first run" carries 130 left / 66 right and 196 of 200 ticks commanded · 112.7 ticks/s · 1.775 s (that is the host run), and the row named "host loop, the board's camera" carries 197 / 99 and 296 of 300 ticks commanded · 31.3 ticks/s · 9.589 s (that is the first board run); only the middle row is correct. The chart's own segment numbers, the caption and the Data table all hold the right values, so the figure contradicts its caption. Fix: reverse the label list as well (or drop the reversed()) and regenerate — the caption's mapping then matches the picture, and no number changes.

  2. The two sentences that say the first board run's numbers came off the broken reader give the wrong mechanism (README.md:85, :338). "Because that first run's numbers came off the broken reader, the entry quotes both runs and labels which is which" and "in different lighting and with different reader code in front of the stream; the first run's numbers came from a reader later found to be broken". The loop computes command per tick from the in-memory spike ids and then writes the trace and the spikes file (crates/connectome-cns/src/main.rs:661-690); the pre-71c9900 defect was in the read path, and board-bench.log's --- loop --- section holds only malformed: spikes: frame 83464099482600 ids are not strictly ascending. So the first run's 197/99/4 come from its own committed trace CSV, not from the broken reader — which is why re-counting them is meaningful at all. What is true, and what the entry should say, is T55's own line: that run's transcript is superseded because its spikes read-back failed, while its command counts are the run's own recorded output. As written, the second sentence also implies the reader changed the command splits, which the code path rules out; the honest difference between the two board runs is the lighting (0.4662–0.4739 vs 0.3802–0.3842) and the network's own dynamics.

  3. The Mermaid drawing's caption still makes the conversion claim the prose has just withdrawn (README.md:210-212 vs :69-71). The prose now says "That left/right/hold is the connectome's own read-out, not yet the robot's motor path: qualia-drive writes DriveCommand::WheelSpeeds from a NavGoal, and nothing yet converts the fly's command into one." The caption says "the connectome's spikes become a left/right/hold decision, and qualia-drive turns that into a wheel command on a link the leash owns" — the very conversion the prose says does not exist — and the diagram draws it as a solid READ -->|steering| DRIVE edge. I checked the tree: no code maps the loop's −1/0/+1 into a NavGoal or a WheelSpeeds (every WheelSpeeds construction in runners/drive comes from NavGoal/control_to_goal). Make the edge dashed and the caption say the left/right/hold stops at the trace, as the limits bullet already does — otherwise the drawing and the prose disagree about whether the fly reaches the drive runner.

(For Main, not an entry note: this branch's merge-base with main is 0cb90a5, so it is still behind the base 42e5a5f; merging without a rebase would revert #260's gate revision and the other base commits. The author is holding for the Mermaid convention PR to rebase, which is the right move.)

@superposition

Copy link
Copy Markdown
Owner Author
role: teaching
verdict: request-changes
notes: 8

Supersedes the earlier teaching comment at this head (issuecomment-5646304556). This is the
consolidated teaching verdict: it folds that comment's four notes in with credit (its N1 and N3 are my
notes 1 and 7; its N2 and N4 are notes 2 and 5). Nothing below restates its list as a parallel one.

Teaching checklist (docs/journal-review.md, copied verbatim)

  1. The entry explains the idea before the result. — pass. The claim (l. 3) states the machine and
    the fly before any rate; ## What we tried (l. 36) gives the leash and the motor path (l. 46)
    before the connectome as controller (l. 60), and the learning loop (l. 102) before the refusal that
    closes it (l. 120).
  2. At least one chart and at least one diagram or 3D render, each captioned so a reader can
    interpret it without the text. — fail, notes 1 and 2. Both legs exist in substance: the two
    charts carry axes, legends, per-bar counts and a source footer, the Mermaid flowchart is captioned
    (l. 210) and every edge is labelled, and the figure table (l. 217) says what each encodes. But
    drive-commands.svg labels two of its three runs with the wrong run (note 1), and the diagram
    draws the fly's read-out feeding qualia-drive as if that path existed (note 2). A reader
    interpreting either without the text is led to a false reading.
  3. Jargon defined on first use. — fail, note 5. All eleven terms in the operator's list carry a
    definition by l. 31 — but four are first used before that paragraph, two of them in the claim
    itself, and two more terms are never defined at all.
  4. The analogy named where one exists. — fail, note 6. The entry's own service name is one and it
    is unspent.
  5. A reader who stops after the first paragraph still learns something true. — pass. The claim now
    ends "The wiring runs. It does not learn, and its commands stopped at a trace file rather than a
    wheel." (l. 13) — the state and the limit in one breath.
agent: Ed261Teaching
branch: journal/fly-drives-waveshare
commit: d2287e4
state: review
next: writer fixes the two blocking defects (notes 1, 2) plus the six editorial notes; I re-read
blocked_on: none
evidence: docs/figures/the-fly-brain-on-the-robot/{README.md,make_figures.py,drive-commands.png}, make_figures.py:133-134, entry l. 200/209, docs/evidence/T55/connectome-runner/README.md, docs/figures/README.md @ main 42e5a5f

The operator's test — pass

I judged the opening as a reader who knows nothing about this project. It now answers all three
questions: what the machine is (a WaveShare UGV; the leash owns its motors, LD06 lidar, camera,
inertial unit), what drives it (the released male-CNS connectome running as a spiking network on
the robot's own computer, camera → encoder → connectome → read-out → left/right/hold), and what is
not yet there
("It does not learn, and its commands stopped at a trace file rather than a wheel",
l. 13; "The loop ran; it did not drive", l. 313). The reframe does what the operator asked: the old
claim at origin/main opened on "11.840 ms on the Orin NX iGPU of Pinkie … and 1.381 ms/tick on
the development RTX 4090
" — the hardware the work ran on — while the new one opens on the robot and
the fly, and the 4090 survives only as a "supporting" Data row (l. 260) and a pointer to the entry
the-4090-and-the-nano.

One wording risk, not a blocker: "produced the steering commands that would reach the wheels"
(l. 9-10) is a counterfactual a fast reader can take as fact; the next sentence is doing the work of
correcting it.

Notes

1. drive-commands.svg labels two of its three runs wrong (blocking). make_figures.py:133 draws
for pos, entry in enumerate(reversed(runs)), but l. 134 labels
ax.set_yticklabels([entry["run"] for entry in runs]) from the unreversed list, whose order is
board loop, first run / board loop, fixed-reader re-run / host loop, the board's camera. So the
top row draws the host run's 130/66/4, "196 of 200", 112.7 ticks/s under the label "board loop, first
run"
, and the bottom row draws the first board run's 197/99/4, "296 of 300", 31.3 ticks/s under
"host loop, the board's camera"; only the middle row is right. A vision read of the committed
drive-commands.png returns exactly that pairing, independently of the code. This is the chart the
entry uses to show what the fly commanded, so as committed it teaches the wrong run's numbers. Fix:
label from the same reversed iteration (or drop reversed and reverse the order once). Credit to
Journal258Teaching; I reproduced it from the generator and from the rendered PNG.

2. The loop diagram draws the missing link as if it existed (blocking). The Mermaid block has
READ -->|steering| DRIVE (l. 200) and DRIVE -->|WheelSpeeds| MOTORS (l. 201), and the diagram's
own caption says "the connectome's spikes become a left/right/hold decision, and qualia-drive turns
that into a wheel command on a link the leash owns" (l. 210-212). The prose three paragraphs earlier
now says the opposite: "That left/right/hold is the connectome's own read-out, not yet the robot's
motor path: qualia-drive writes DriveCommand::WheelSpeeds from a NavGoal, and nothing yet
converts the fly's command into one" (l. 69-71). The prose fix is right; the diagram and its caption
still show the oversell the writer just removed, and the diagram is the artefact a reader looks at
first. Draw the missing conversion the way the learning loop is drawn — a dashed edge from the
read-out to qualia-drive, labelled "not built: T67 (#262)" — and make the caption agree with
l. 69-71. Credit to Journal258Teaching.

3. The failures section has one item that teaches nothing about driving or learning, and one that
stops a step short.
The section's own frame is "The ones worth keeping are the ones that taught
something about driving and learning" (l. 75).

  • Cut or fold "A silent frame that fired node 0" (l. 87-93). It is the console's spike reader
    accepting an EOF frame with ids: [0]; its rule — "an end-of-file that reads as data is worse than
    an error" (l. 92) — is a decoder rule about the instrument's reader, and the console figure is
    already demoted to two supporting rows and bounded in the limits section. It is the same class as
    the item above it (a reader decoded bytes it should have refused), so one clause inside that item
    keeps the lesson and loses the process report.
  • Keep "A reader skipped four bytes too many" (l. 77-85): it is the reason the entry quotes two board
    runs, which is what a reader needs to trust the evidence.
  • Sharpen "A bottleneck that was never probed" (l. 95-100). It does teach something about the loop —
    the probe found 80.5 snapshots/s at the leash against 32.1 ticks/s consumed, so the loop was not
    camera-bound — but it stops before the reader's next question: docs/evidence/T55/… says the board
    loop is 31.2 ms/tick end to end, ~19 ms above the 11.840 ms bare step, and the entry never says
    where that remainder goes. One clause naming the unmeasured remainder turns the item from "we
    withdrew a claim" into "we found where the bottleneck is not, and have not found where it is".

4. The learning story names the pipeline and the block, but never says what learning would mean.
The transitions → MCAP → eleven gates → train_batch → promotion floor path is clear, the D-027
refusal is quoted whole (l. 120-121), and "Nothing here learns" (l. 321) is honest. What is missing is
the middle: the diagram's dashed edge says the gate "would retune the fly" (l. 206), but nothing says
by what signal a world model that predicts the next latent changes what the read-out emits, and
nothing says what you would have to see for "the fly learned to drive" to be a true sentence — a
checkpoint that passes the floor and a drive run whose commands move with it, or whatever the real
observable is. One or two sentences in the learning paragraph (l. 102) would give the closing line
"making it learn to drive is the work now" (l. 128) a picture behind it. This is the note I would most
want answered, because the operator chose this entry as the fly learning to drive.

5. Four terms are used before the gloss paragraph, two of them in the claim; two are never
defined.
The claim's numbers are unreadable without the definition that follows them: tick first
appears at l. 10 ("296 non-hold ticks out of 300") and is defined at l. 22; UGV at l. 3, defined at
l. 24; neuron and synapse at l. 7-8, defined at l. 15. Add inline glosses where they land — "a
WaveShare UGV (a small wheeled rover)", "ticks (simulation steps)" — so the opening is self-contained.
deadman (l. 42, and "deadman 400 ms" in the flowchart at l. 181) and the shared arena (l. 49) are
never defined anywhere — one clause each at first use. (The earlier comment's N4 is the deadman half
of this note.)

6. The analogy is present but unnamed. The service is called the leash and the entry's whole
safety spine is that one owner holds the motor link; the definition is functional ("the board-side
process that owns the hardware", l. 23) but never says the name is the image — a leash is what stops an
animal bolting, and the 400 ms deadman, the zero-speed link claim and the disarm codes are the leash's
job. One clause at first use makes the ownership point stick, and it costs nothing.

7. The entry draws the first board run and tells the reader to distrust it, without saying which
number is untrustworthy.
l. 84-85 says "that first run's numbers came off the broken reader", and
l. 338 repeats "the first run's numbers came from a reader later found to be broken" — yet
drive-commands.svg and the Data table both draw and quote 197/99/4 of 300 from that run's trace CSV.
docs/evidence/T55/connectome-runner/README.md §What was run describes board-bench.log as
"historical … an Orin build/verify/bench excerpt whose loop section is a failed read-back by the
binary built before 71c9900", superseded by board-loop.log, and lists board-loop-trace.csv as
"the earlier board loop's per-tick trace" — while the generator re-counts the trace, not the reader.
So a reader cannot tell whether 197/99/4 is real or corrupt. Either say precisely what about the first
run was untrustworthy (its replay/read-back, not necessarily its trace), or drop the disowning and
keep the honest "the two board runs are two runs, not a distribution". The causal claim itself is
accuracy's to rule on
— I am flagging the reader's problem and the source I compared, not
adjudicating the number. (The earlier comment's N3.)

8. "What is deliberately not here" (l. 275-279) tells the reader about the logo/mark. The
operator's instruction was that they do not want to read about the logo/mark, and a published reader
learns nothing from a list of absent things. Cut the logo sentence; if the render-band rule needs a
word, keep one clause pointing at the-4090-and-the-nano. The style review's note 2 covers the
related README.md-as-figure row in the table.

What a reader would still ask

  • How would a world model retune the connectome? (note 4) — the intended coupling, in a sentence.
  • What would count as evidence that the fly learned to drive? (note 4) — the observable, named
    before the floor is met rather than after.
  • Where does the closed-loop tick actually go? 31.2 ms/tick end to end against an 11.840 ms bare
    step, with the camera serving 80.5 snapshots/s (note 3).
  • What connects the trace to the wheels? The entry says nothing could open /dev/ttyTHS1 because
    the leash holds it, "by design" (l. 317) — a reader asks what the intended route is, and the diagram
    currently answers "this solid arrow" (note 2). T67 (T67 — The fly's wheel commands reach the wheels: a leash-drive transport, with a deadman #262) is that route; a forward pointer costs one
    clause.
  • What does the robot look like, and how big is it? The entry names the model and its sensors but
    carries no picture of the machine (correctly, given the render rule); "a small wheeled rover" in the
    claim would place it.

How I checked, and what I did not

I read the entry as it exists at the head, not the diff:
git show origin/journal/fly-drives-waveshare:docs/figures/the-fly-brain-on-the-robot/README.md
(d2287e4; the PR body's braid still names c432a29, so the sha matters). I diffed that file against
origin/main's to see what the reframe moved — the old claim opened on the Orin NX and 4090 rates —
and against b087e99/d2287e4 to see the style answers. Notes 1 and 2 come from make_figures.py
lines 133-134 and the Mermaid block at l. 173-206 with its caption at l. 210-212, checked against the
new prose at l. 69-71; I also read wave-numbers.json to confirm the run order the labels are drawn
against, and a vision read of the committed drive-commands.png confirms the swap. Note 7 is a
comparison of the entry's two sentences with T55's log list — I did not open board-loop-trace.csv or
re-derive whether the trace or the reader produced the first run's counts; that is the accuracy
editor's leg, and I say so in the note. I did not re-run the generator or the gates: the writer reports
figures: OK (11 entries, 53 figures, 400 KiB budget) and provenance: OK at this head. The
docs/mermaid-diagrams branch exists (fb1e05c), and the entry's own sentence now says the convention
"is being amended" (l. 295-297) — honest — so checklist item 2's diagram-form question is the style
review's to close, not mine.

@superposition

Copy link
Copy Markdown
Owner Author
role: style
verdict: request-changes
notes: 5

Re-read of the live head d2287e4, not the PR body's c432a29: the branch has moved four times
since that body was written (41aa898, b087e99, then the author's two answer commits d5bbe87,
d2287e4). Only docs/figures/the-fly-brain-on-the-robot/README.md moved in those four, so the root
README, the two charts and the wave-numbers.json are the same files the earlier style comment read.
All four checklist legs pass; three of my five notes are the one form question this PR leaves open —
the Mermaid drawing — judged here against the parallel convention amendment, which the earlier style
comment could not see.

Style checklist (docs/journal-review.md §style, copied verbatim)

  1. The site's form: **The claim.** first, then ## What we tried, ## Evidence,
    ## What this does not establish. — pass. # The fly brain on the robot (l. 1) →
    **The claim.** (l. 3) → ## What we tried (36) → ## What we changed (131) → ## Evidence (164)
    → ## What this does not establish (311), with ### Data (245) and ### Regenerating (281) under
    Evidence: the claim first and all four template H2s present, in order, the claim one paragraph with
    its numbers inline. Machine-checked on this head's tree:
    qualia-gates journal --entry …/the-fly-brain-on-the-robot/README.md →
    journal-gate: entry OK (roles not checked: pass --pr <n>); qualia-gates figures --root <head tree>
    → figures: OK (11 entries, 53 figures, 400 KiB budget), exit 0. See notes 1–2: the drawing this
    form leans on is not yet a form the convention sanctions.
  2. Short paragraphs, no filler. — pass, with notes 3 and 4. No filler register, no second person
    (you|your = 0), no marketing vocabulary (very|really|simply|seamless|powerful|cutting-edge… = 0),
    and the British spellings hold ("behaviour", "coloured", "labelled"). "artifact" is this
    repository's own dominant spelling (git grep -c artifact returns 44 files against 5 for
    "artefact"), so it is not a correction to make.
  3. Front matter valid for the site's build. — not satisfiable in this tree, as with the three
    merged entries: no --- block, and the entry says so itself in its last bullet ("the site's front
    matter and live URL are produced from them when the entry moves to the site's _posts/"). Publish
    step, not this PR.
  4. Figures and band referenced by absolute URL. — pass for the committed figures and for the band;
    notes 1–2 for the drawing.
    Both chart embeds are the absolute
    https://raw.githubusercontent.com/superposition/qualia/main/docs/figures/the-fly-brain-on-the-robot/<name>.svg
    form (ll. 224, 236), each with its .png companion declared in the table (ll. 219–220), and there
    is no relative figure reference anywhere (a search for a non-http .svg|.png|.webp|.glb
    reference returns nothing). The band is omitted, and that omission is stated, not merely
    missing
    — the check below.

The logo/mark absence, and the tables

The entry states the omission itself (ll. 275–279): "What is deliberately not here. This entry has
no logo, no mark and no hero render of one — not in the text, not in the figures, not in the Data
table — so the render band of docs/figures/README.md rule 2 is omitted on purpose and
docs/figures/the-fly-brain-on-the-robot/ holds two charts and no render." I checked the absence
rather than the assertion: git grep -oi 'psi|logo|mark|hero' over drive-commands.svg,
dataset-gate.svg and wave-numbers.json returns nothing; git ls-tree of the directory at this head
is README.md, the two .svg + .png pairs, make_figures.py and wave-numbers.json — no
hero.webp, no .glb, no mark asset — and in the entry the words "logo" and "mark" occur only inside
that statement. The band in the older entries is a render of the mark (the-ladder's hero.webp is
"the extruded psi mark"), so the omission is the operator's instruction landing, not a missing asset.

Tables: the figure table (ll. 218–222) declares the four files that exist at this head, with kinds the
checker accepts (chart, chart, diagram, data) and both .png companions present; the Data
table's 22 rows each name a source, and wave-numbers.json is declared as the recording it is. The
root README's journal row still holds: | the fly brain on the robot | [EPIC-06](…#6) — the fly driving the WaveShare UGV | is present with an absolute epic link, and the min-epic order (1, 2, 4, 6, 6, 7,
10, 11, 12, 13, 14, 15) is non-decreasing, which is what check_journal reads. Note 5 is about that
row's shape, not its presence.

Notes

  1. The drawing is not yet a form the convention sanctions — in this tree or in the amendment as
    drafted, and the amendment should land first.
    On main, rule 1 (ll. 33–35) names exactly two
    forms: "a schematic drawn in _house.py's box/arrow vocabulary, or a Blender render from steps
    42–43". So at this head docs/figures/the-fly-brain-on-the-robot/ ships two charts and no
    drawing
    , and the green figures: OK rests on a table row whose File cell is the entry itself —
    row l. 221, | `README.md` (this file) | diagram, not a committed SVG | … | — because
    check_entry in crates/gates/src/figures.rs reads the kind (diagram) and checks only that the
    backticked path exists.
    The parallel amendment (docs/mermaid-diagrams) is still only uncommitted edits in
    C:/Users/ericm/wt/MermaidDiagrams — that branch is at 42e5a5f (= main) and
    gh pr list --head docs/mermaid-diagrams returns nothing — so there is not yet a rule in main for
    this entry to cite. As drafted, the amendment sanctions Mermaid as "a <name>.mmd committed beside
    the figures, whose mermaid block the published entry carries (or a fenced block written straight
    into the entry, with the .mmd as the source of record here)", and its own check paragraph is
    explicit: "A diagram that lives only inside the published Markdown is invisible to the check, which
    is why the .mmd is committed here." Its model entry does exactly that:
    the-agent-in-the-loop gains strand-flow.mmd and its table row names the .mmd as the diagram.
    The fix is therefore two steps, in this order: (a) land the amendment — it is the rule this entry's
    drawing needs, and the entry cannot be judged conformant before it exists; (b) commit the diagram's
    source beside the figures (e.g. docs/figures/the-fly-brain-on-the-robot/drive-loop.mmd), point the
    row at it with kind diagram, and keep the fenced block in the entry as the copy. That second step
    is already the house pattern for flows — docs/architecture/README.md: "The .mmd files in this
    directory are the source. The mermaid blocks in the repository README.md are copies of them."
    With it, the diagram, not a committed SVG qualifier goes too: the convention allows a qualifier
    that describes the figure ("chart, two panels"), not one that reports the file's status.

  2. The two sentences about the convention still do not match it. The entry now says (ll. 293–297):
    "the convention is being amended to name it as the drawing form beside a _house.py schematic or a
    Blender render. Once that lands, this entry commits two charts and one drawing…" — closer, but the
    amendment names a committed .mmd as the source of record and keeps the hand-drawn schematic as
    the form where the subject needs the box/arrow vocabulary or the house colours; and
    make_figures.py's docstring still asserts the present tense (ll. 9–12: "docs/figures/README.md
    names that form as the diagram rule 1 accepts"). Rewrite both after the amendment lands, quoting the
    amended rule and the .mmd. The generator's docstring matters most: it is the reason a later agent
    reads for why this directory has no SVG diagram.

  3. The voice note belongs in the PR body, not the entry — and it is now self-referential. The codes
    note (ll. 31–34) ends: "The voice here is "we"; the single "I" is the operator's own judgement
    about the motor path's safety properties, recorded in the PR body as the reason the account is
    first-person." That explains the entry to a reviewer, inside the entry; the only I in the file is
    the token in that sentence (grep -o '\bI\b' = 1, at l. 33; "we/our/us" = 15), because the rewrite
    moved all three prose "I"s to "we"; and the PR body as fetched does not carry the record it claims
    (its braid still names c432a29). Move the line to the PR body, or drop it. The three "I"s would
    have been within the operator's instruction — a first-person account, with "I" on judgements — so
    the substantive change was optional; the sentence describing it is the part that does not belong.

  4. One paragraph is a block, and one line is out of the file's wrap. The glossary paragraph
    (ll. 15–28) is 234 words in one block — eight defined terms plus the plasticity/neuromodulation
    close — immediately after a 193-word claim, against a 162–176-word longest paragraph in the three
    merged entries. Splitting the closing sentence out would shorten the claim slot without touching a
    definition. Separately, this revision joins two half-lines into a 140-character line at l. 48; every
    other prose line is ≤108 (median 98), so that is a wrap break to fix.

  5. Shape call, not a failure: the journal row carries the only prose in the table.
    README.md:244 reads | the fly brain on the robot | [EPIC-06](…#6) — the fly driving the WaveShare UGV | while the other ten rows are | entry | epic links |. check_journal passes either way
    (absolute epic link, min-epic order holds). My call: drop the trailing clause — the reframe is
    already the entry's opening sentence, and one annotated row in eleven reads as a note to the
    reviewer rather than the index's shape. Not worth a revision on its own; author and Main can settle
    it.

What passes, so the revision can stay narrow

Claim first and the four H2s in order with the numbers inline; both charts by absolute URL with their
.png companions and per-run captions; 22 Data rows with their sources named; the logo/mark absence
deliberate, stated and verified in the figure files; one "we" voice, no second person, no filler, no
emoji; no draft marker (TODO|FIXME|TBD|placeholder|ASK:|WIP returns nothing); both gate legs green on
this head. The one substantive open question is the drawing's form, and closing it is the amendment
plus one commit in this PR.

What I could not verify

The rendered Mermaid output — I read the block source (two subgraphs, two matching ends, well-formed
--> and -..-> edges, quoted labels), and the checker does not render it; front-matter validity (no
site tree in this repository); whether the operator asked for "a Mermaid block" specifically or "no
third SVG figure" (I did not see the instruction; the PR body and the generator both state the former).
The amendment's status is read from this host only: if it has a PR elsewhere, note 1's step (a) is
already in flight and only step (b) remains here.

@superposition

Copy link
Copy Markdown
Owner Author
role: accuracy
verdict: request-changes
notes: 7

Supersedes the accuracy comment already at this head (the 3-note one, role: accuracy,
2026-09-12T13:54:21Z). This is the accuracy role's verdict, read on d2287e4 — the PR body's
braid still names c432a29, and the branch has moved four commits since (41aa898, b087e99,
d5bbe87, d2287e4; a rebase for the Mermaid convention is still held). The earlier comment's
three notes are re-derived below and carried as notes 1, 5 and 6; notes 2, 3, 4 and 7 are added.
Nothing from it is dropped.

Checklist (copied verbatim from docs/journal-review.md §accuracy)

  1. Every number traced to a source in the entry itself. — request-changes. The claim, the Data
    table and the two charts' counts all trace; one drawn rate does not (note 2).
  2. Units on every quantity. — holds. Rates carry ticks/s next to the machine and the regime
    (free-running bench, closed loop, host loop), the serial facts carry baud and ms, the floor
    carries sessions/environments/conditions, the sessions carry seconds and bytes, the connectome
    carries a release id and a licence.
  3. Any value measured in a different session or on different hardware labelled as such. — one gap:
    the trainer transcript (lines 119–122) is not labelled with the machine it ran on (note 3). The
    Orin NX / RTX 4090 / i9-14900KF rows, the two board runs, the host run, the T58 and T65 sessions
    and the console render are each labelled.
  4. Every figure's caption says what it encodes. — fails for drive-commands.svg: the caption and
    Data table are right, the picture labels are swapped (note 1).
  5. No claim about a private repository. — holds. gh repo view superposition/qualia --json visibility → PUBLIC; every source named is in this repo, and the two absolute figure URLs are
    raw.githubusercontent.com/superposition/qualia/....

How I checked

The three command splits, re-counted from the committed traces, not read from the prose (awk -F,
over the command column, header skipped):

host-loop-gpu-trace.csv      -1: 130, +1:  66, 0: 4   (200 rows) → 196 non-hold
board-loop-trace.csv         -1: 197, +1:  99, 0: 4   (300 rows) → 296 non-hold
board-loop-trace-rerun.csv   -1:  98, +1: 196, 0: 6   (300 rows) → 294 non-hold

Every split in the claim, the figure table, the chart and wave-numbers.json is exact, and each is
attributed to the right run.

The rates. 11.840 ms/tick / 84.5 ticks/s (Orin), 1.381 / 723.9 (4090), 20.615 / 48.5
(host CPU) are T55's bench table, with the free-running bench's silence ("0 spikes on the last tick")
and the 1.41× 60 Hz headroom (16.7 / 11.840). The closed-loop rates: board-loop.log:439
(300 ticks in 9.358 s (32.1 ticks/s), 15085686 spikes, 294 non-hold commands) for the fixed-reader
re-run, host-loop-gpu.log / T55 §The closed loop (200 ticks in 1.775 s (112.7 ticks/s), 10048960 spikes) for the host run, and the 200-fetch probe (200 fetches in 2485 ms = 80.5/s) with the
withdrawn "camera-bound ~33/s" line. What I could not find in a committed file is the first
board run's rate — see note 2.

Robot facts. runners/drive/src/main.rs: DEFAULT_BAUD 115_200, DEFAULT_TICK_MS 100,
POSE_STALE_NS/LIDAR_STALE_NS 1_500_000_000, MAX_WHEEL_CMD 0.34, MAX_TURN_CMD 0.20,
FRONT_STOP_DIST_M 0.48, PROGRESS_TIMEOUT_SECS 5, MIN_PROGRESS_M 0.10,
ESTOP_FILE "/tmp/qualia-estop" under // Latching stop and stall watchdog, speed_line() emitting
{"T":1,"L":left,"R":right} plus '\n', the arming set "1" | "true" | "TRUE" | "yes" | "YES", the
zero-speed claim while armed (main.rs:219-221), transmit() writing only when armed, and
ACTION_SAFETY_DISARMED / ACTION_SAFETY_TRANSPORT_ERROR. T59 §Who holds what:
LEASH_DEADMAN_MS=400, LEASH_SERIAL_PORT=/dev/ttyTHS1, LEASH_SERIAL_BAUD=115200,
LEASH_CAMERA_DEVICE=/dev/video0, LEASH_UGV_LIDAR_DEVICE=/dev/ttyACM0, errno=16 on both serial
ports, camera second open fd=3, qualia-lidar/qualia-drive exit 2, no IMU on I2C/IIO. T58:
pose: null, the 900 s/83,299,459 B capture, 4,502 camera / 8,775 lidar records, 4,501 candidates,
0 valid at calibration_missing. T65: 30,300,801 B / 03ae6b0f…7110, 1,501/2,917/2,917/2,997
records, valid=1499 candidates=1500, mean skew 29.19 ms, coverage 1.0, one action_coverage
rejection, calibration_id cac618af…f6e2, 3,100 accepted intervals at speed 0.000. The entry never
claims the robot was driven further than that, and "The loop ran; it did not drive" is exact.

Learning claims. crates/connectome-cns/src/lif.rs is membrane potential + refractory with
// no learning, no neuromodulation, no plasticity; main.rs:124/128/131 are
GRID_COLUMNS 8 / SENSORY_GAIN 0.5 / DECISION_MARGIN 0.001, and the read-out is
rate_r > rate_l → +1 else -1 within the dead band. crates/fly-circuit is off by default
(sim feature; nothing depends on it from runners/drive or connectome-cns), README "invented …
never reachable from the motor path", 5 types / 9 edges in crates/connectome-prior/tests/prior.rs.
crates/jepa-model/src/train.rs:169 is train_batch. D-027's floor (50,000 / 12 / 3 / 3, each ≥ 5 %,
shortfall 33.4×) and "no checkpoint has been promoted" both hold.

Gates, re-run on a worktree of d2287e4 with the already-built binary (no build, no sibling
contention):
qualia-gates figures → figures: OK (11 entries, 53 figures, 400 KiB budget);
qualia-gates provenance → compared 152 authored file(s) (1 generated skipped); 0 identical, 27 EOL-identical, 7 code runs over 20, 0 prose runs over 2 and highest identical-code-line share 1.000 (.cargo/config.toml), exit 0. Both console blocks match the entry byte for byte.

Notes

  1. drive-commands.svg draws each run's numbers under the wrong row label. make_figures.py:98
    iterates reversed(runs) (host first) but :134 sets yticklabels from runs (board first) and
    :135 flips the axis, so in the committed SVG the y-label board loop, first run (line 581) sits
    with the host's 196 of 200 ticks commanded (line 1076) and host loop, the board's camera
    (line 747) with the first board run's 296 of 300 (line 1496). The caption (line 219) and the
    Data table are correct; the picture contradicts them. Fix the generator, regenerate both .svg
    and .png, and re-check the byte-identity/budget rules. (Verified independently from the
    generator and the committed SVG; credit to the earlier accuracy comment.)
  2. The first board run's rate has no source the entry names. The claim row says the three runs'
    numbers are "Re-counted from each run's committed trace CSV", and the chart's own footer says
    Source: each run's committed trace … re-counted here, but those CSVs have no time column
    (tick,luminance,fired_input,rate_l,rate_r,command,throttle — crates/connectome-cns/src/main.rs:654),
    so 31.3 ticks/s and 9.589 s, which the chart draws (line 1076 of the SVG: 31.3 ticks/s · 9.589 s) and the figure table prints (line 219), cannot come from them. wave-numbers.json
    records their real source as command_runs[0].log_line: issue #236 braid comment, 2026-09-12
    (where the loop printed cns-loop: 300 ticks in 9.589 s (31.3 ticks/s), 15168990 spikes, 296 non-hold commands); git grep 31.3 in docs/evidence finds nothing. Name #236 for the first
    run's rate/duration, or draw the rate only where a committed log carries it, and drop the
    "committed evidence file" claim in the wave-numbers.json row (line 222) for that run. (New.)
  3. The trainer bullet contradicts itself and points at the wrong transcript (lines 361–366).
    The quoted block (lines 119–122) is the post-fix run: docs/decisions.md D-027 carries the
    identical command (--manifest …6f98b125….json --backend cuda --epochs 1 --batch-size 32 --seed 65),
    message and TRAIN_RC=1, and docs/evidence/T52/promoted-e2e/digest-fix/trainer-after.txt is the
    transcript — taken on the host (WSL2 Ubuntu-22.04, RTX 4090), which the entry does not say
    (checklist 3). The pre-fix board run printed qualia-jepa-train: dataset manifest is unsupported or has an invalid digest (digest-fix/trainer-before.txt, quoted in T65 §7). So "taken before that
    fix and its refusal message is the post-fix one" cannot both be true; write "the run quoted is the
    post-fix one (D-027; T52 §1, on the host); the pre-fix run is T65 §7". (New.)
  4. The Artifact row (line 258) cites a file that does not carry two of its values. T55 has
    weights.bin 180,414,198 B, positioned 139662, types 11752, 190 MB, but not
    positions.bin 3,334,012 B, its sha256 or 211,577 bodies — those are in
    docs/evidence/T56/floating-brain/README.md (positions.bin: 3334012 bytes sha256=016a285b…afa6e; source table: 211577 bodies, 139662 positioned, 11752 type labels). Add
    T56 to the row. (New.)
  5. "the first run's numbers came off the broken reader" (line 85, and "reader code in front of the
    stream", line 338) gives the wrong mechanism.
    The defect was the QLSP spike-stream read-back —
    T55's header says "the writer had always been right, so only the reading was broken", and
    board-bench.log's own header calls its loop section "a FAILED read-back", not the loop's stdout.
    The loop computes its command in-process from the frames it just produced and writes them to the
    trace (crates/connectome-cns/src/main.rs:660-690), so 197/99/4 and the printed 31.3 ticks/s
    are the run's own output. What is true and worth saying: the first run's committed transcript
    proves nothing, so its numbers stand on the trace CSV and the #236 loop line. (Verified;
    credit to the earlier accuracy comment.)
  6. The Mermaid caption still makes the conversion claim the prose has withdrawn. Line 211:
    "qualia-drive turns that into a wheel command on a link the leash owns", against lines 69–71:
    "qualia-drive writes DriveCommand::WheelSpeeds from a NavGoal, and nothing yet converts the
    fly's command into one". The solid DRIVE -->|WheelSpeeds| MOTORS edge likewise reads as an
    exercised path, while on this robot runners/drive could not open /dev/ttyTHS1 (T59
    DRIVE_RC=2) and the only applied command in the ledger is the leash's zero-speed stop. Make the
    caption "would turn" (or dash/annotate that edge: design; T67 — The fly's wheel commands reach the wheels: a leash-drive transport, with a deadman #262/T67 makes it real). (Verified;
    credit to the earlier accuracy comment.)
  7. The release citation has no source named. Line 8 asserts Berg et al., *Cell*, 2026; the
    commit's own attribution is assets/brain/connectome-cns/attribution.json ("Berg et al. 2026,
    Cell; doi:10.1016/j.cell.2026.08.015"). Name that file, or put the citation in the Connectome row
    of the Data table with its source. (New.)

(For Main, not an entry note: the branch's merge-base with main is 0cb90a5 while main is
42e5a5f, so the rebase the reframer is holding is required before merge — and the two figure URLs
point at raw.githubusercontent.com/.../main/..., so they resolve only after that merge.)

agent: Ed261Accuracy
branch: review of journal/fly-drives-waveshare
commit: d2287e4 (reviewed; PR head may move on the held rebase)
state: accuracy verdict posted — request-changes, 7 notes (folds the earlier 3-note accuracy comment)
next: author fixes notes 1–4 (generator, sourcing, transcript, row source) and pushes; this role re-reads the changed lines
blocked_on: none
evidence: awk re-count of the three trace CSVs; git grep 31.3 (only wave-numbers.json, sourced to the #236 braid); runners/drive/src/main.rs; crates/connectome-cns/src/{main,lif}.rs; T55/T56/T58/T59/T65 READMEs; docs/decisions.md D-025/D-027; qualia-gates figures+provenance on a worktree of d2287e4 (exit 0, output quoted above)

@superposition

Copy link
Copy Markdown
Owner Author
role: style
verdict: request-changes
notes: 4

Re-read at the live head 697ec97 (my first verdict pinned d2287e4; the branch has since taken
0610791 and 697ec97). The drawing question I opened is now closed in shape: driving-loop.mmd is
committed beside the figures, the table row is | `driving-loop.mmd` | diagram | … | (l. 225), and
I checked the entry's fenced block against the file — 34 lines each, identical after line-ending
normalisation. So the remaining block is the amendment, the meta sentence, and two nits. The checklist
items and their verdicts at this head:

  1. The site's form: **The claim.** first, then ## What we tried, ## Evidence,
    ## What this does not establish. — pass. # The fly brain on the robot (1) → **The claim.**
    (3) → ## What we tried (37) → ## What we changed (135) → ## Evidence (168) →
    ## What this does not establish (315), ### Data (249) and ### Regenerating (285) under Evidence.
    Gates on this head's tree: qualia-gates figures --root … → figures: OK (11 entries, 53 figures, 400 KiB budget), exit 0; qualia-gates journal --entry …/README.md → journal-gate: entry OK (roles not checked: pass --pr <n>).
  2. Short paragraphs, no filler. — pass with note 3. No second person, no marketing vocabulary,
    British spellings hold; "artifact" is this repository's dominant spelling.
  3. Front matter valid for the site's build. — not satisfiable in this tree (no --- block; the last
    bullet says the site's front matter is produced at the _posts/ step). Publish step, not this PR.
  4. Figures and band referenced by absolute URL. — pass. Both charts are the absolute
    raw.githubusercontent.com/superposition/qualia/main/docs/figures/the-fly-brain-on-the-robot/<name>.svg
    form, with their .png companions declared; no relative figure reference exists; the band is omitted
    and stated at l. 279 ("no logo, no mark and no hero render of one — not in the text, not in the
    figures, not in the Data table"), which I verified in the figure files (grep -oi 'psi|logo|mark|hero'
    over the two .svg and the .json: nothing).

Notes

  1. The rule the entry now correctly describes is still not in main; the amendment must land first.
    Ll. 298–302: "a <name>.mmd committed beside the figures is one of the forms it names, and the entry
    carries its mermaid block. That is this entry: two charts, and driving-loop.mmd as the drawing."
    That is the amendment's substance, stated in the present tense about a file that today names only
    "a schematic drawn in _house.py's box/arrow vocabulary, or a Blender render from steps 42–43"
    (docs/figures/README.md ll. 33–35). The amendment is still uncommitted edits in
    C:/Users/ericm/wt/MermaidDiagrams (docs/mermaid-diagrams = 42e5a5f = main; no PR). So the
    sequencing is the fix: land the amendment, then this sentence is true as written and the entry is
    conformant. make_figures.py's docstring (ll. 8–13) carries the same present-tense claim and still
    says "this directory therefore commits the two charts and no SVG diagram" — it needs the .mmd named
    too, or the next agent reads the old reason for the missing SVG.
  2. The voice sentence is still in the entry, not the PR body. Ll. 32–35: "The voice here is "we";
    the single "I" is the operator's own judgement about the motor path's safety properties, recorded in
    the PR body as the reason the account is first-person." It explains the entry to a reviewer inside the
    published prose; the only I in the file is the token in that sentence (grep -o '\bI\b' = 1, at
    l. 34); and the PR body as fetched carries no such record (its braid still names c432a29). Move it
    to the PR body or drop it. One line; the only item here that would ship to readers.
  3. Wrap break and the glossary block (repeated, unchanged). L. 49 is 140 characters where every other
    prose line is ≤108 (median 98) — a joined half-line from the voice rewrite. The glossary (ll. 15–30)
    is now 255 words after the deadman gloss, against a 162–176-word longest paragraph in the merged
    entries; splitting the closing plasticity sentence out would shorten the claim slot (193 words) without
    losing a definition.
  4. Shape call: the journal row's trailing clause. README.md:244 still reads
    | the fly brain on the robot | [EPIC-06](…#6) — the fly driving the WaveShare UGV | against ten rows
    of | entry | epic links |. check_journal passes either way. Drop the clause or annotate all eleven;
    not worth a revision on its own.

Unchanged from my first verdict

Both chart embeds absolute with .png companions and per-run captions; 22 Data rows each source-named;
no second person, no filler, no emoji, no draft marker; the README journal row present with an absolute
EPIC-06 link and non-decreasing min-epic order. With note 1 sequenced and note 2's line moved, this is an
approve.

What I could not verify

The rendered Mermaid output (I read both copies' source: two subgraphs, two matching ends, quoted
labels, the READ -. "not yet converted" .-> DRIVE edge the teaching revision added — the checker does
not render it); front-matter validity (no site tree here); whether the operator asked for a Mermaid block
specifically (the PR body and the generator both state it; I did not see the instruction).

@superposition

Copy link
Copy Markdown
Owner Author
role: teaching
verdict: request-changes
notes: 5

Re-read at 3fd0bbd, superseding my issuecomment-5646310879 and Journal258Teaching's
issuecomment-5646304556. Four of the eight notes I raised are now closed and I verified each against
the artefact, not the reply: the chart labels (note 1), the Mermaid-vs-prose contradiction (note 2),
the drawn-and-disowned first board run (note 7), and the deadman gloss. Five remain, all editorial,
all unchanged at this head.

Teaching checklist (docs/journal-review.md, copied verbatim)

  1. The entry explains the idea before the result. — pass. Unchanged.
  2. At least one chart and at least one diagram or 3D render, each captioned so a reader can
    interpret it without the text. — pass. The chart pairing is now correct and the diagram agrees
    with the prose; driving-loop.mmd (new, 34 lines) is byte-identical to the fenced block and the
    figure table now names it as the committed diagram. The drawing-form sentence (l. 298-302) rests on
    PR docs: Mermaid as the sanctioned diagram form, and the strand flow in it #263, which is open and does change docs/figures/README.md; true once it lands, and the writer
    is holding for it.
  3. Jargon defined on first use. — fail, note 3. Unchanged from the last verdict.
  4. The analogy named where one exists. — fail, note 4. Unchanged.
  5. A reader who stops after the first paragraph still learns something true. — pass. Unchanged.
agent: Ed261Teaching
branch: journal/fly-drives-waveshare
commit: 3fd0bbd
state: review
next: writer answers notes 1-5; Ed261Teaching re-reads after the #263 rebase
blocked_on: none
evidence: make_figures.py:137 labels reversed(runs); vision read of the committed drive-commands.svg; entry l. 200, 210-212, 230-231, 340-342, 91-101, 108-135, 279-283

Closed since the last verdict — how I checked

  • Note 1, chart labels. make_figures.py:137 is now
    ax.set_yticklabels([entry["run"] for entry in reversed(runs)]). I read the committed
    drive-commands.svg myself: top host loop, the board's camera 130/66, "196 of 200", 112.7
    ticks/s; middle board loop, fixed-reader re-run 98/196, "294 of 300", 32.1; bottom
    board loop, first run 197/99, "296 of 300", 31.3. Pairing correct.
  • Note 2, the diagram. The edge is now READ -. "not yet converted" .-> DRIVE (l. 204) and the
    caption (l. 210-212) says "the fly's command is not yet converted into a DriveCommand::WheelSpeeds"
    — the diagram and l. 70-71 now say the same thing.
  • Note 7, the first board run. The failures item now says the loop "computes its command live and
    writes the trace as it goes, so the first board run's counts are its own trace's
    (board-loop-trace.csv), not the reader's", and the limits bullet (l. 340-342) now reads "with the
    first run's recording never read back (the spikes reader was broken; its trace is what it
    commanded)". The internal contradiction is gone, and the chart caption's "refuses to draw if a trace
    and the counts recorded beside it disagree" (l. 233) is now precise.
  • deadman is glossed at l. 24: "a watchdog that stops the motors if commands stop arriving (the
    leash holds its own at 400 ms)".

Remaining notes

1. The failures section still keeps a decoder report and still stops short on the one that teaches.
"A silent frame that fired node 0" (l. 91-97) is the console's spike reader accepting an EOF frame
with ids: [0]; its rule — "an end-of-file that reads as data is worse than an error" — is a decoder
rule about the instrument, and the console figure is already two supporting rows and bounded in the
limits section. It is the same class as the item above it, so one clause inside that item keeps the
lesson. "A bottleneck that was never probed" (l. 99-104) still stops before the reader's next
question: T55 says the board loop is 31.2 ms/tick end to end, ~19 ms above the 11.840 ms bare step,
and the entry never says where that remainder goes. One clause naming the unmeasured remainder turns
the item from "we withdrew a claim" into "we found where the bottleneck is not".

2. The learning story still does not say what learning would mean. The pipeline, the D-027 refusal
and "Nothing here learns" are all intact and honest; what is missing is the middle. "so the fly's
controller can be shaped by what actually happened on the robot" (l. 108) and the diagram's dashed
GATE -. "would retune the fly" .-> READ assert a coupling that is never explained — by what signal a
world model that predicts the next latent changes what the read-out emits — and nothing names the
observable that would make "the fly learned to drive" a true sentence. One or two sentences in the
learning paragraph gives "making it learn to drive is the work now" a picture behind it. This is the
note I would most want answered, because the operator chose this entry as the fly learning to drive.

3. Vocabulary still precedes its gloss in the claim, and one term is still undefined. tick
(l. 10, "296 non-hold ticks out of 300") is defined at l. 22; UGV (l. 3) at l. 24; neuron and
synapse (l. 7-8) at l. 15. The claim is the operator's test paragraph, and its numbers are unreadable
without "one tick is one step". Inline glosses — "a WaveShare UGV (a small wheeled rover)", "ticks
(simulation steps)" — close it. the shared arena (l. 50) is still defined nowhere.

4. The analogy is still unnamed. "The leash is the board-side process that owns the hardware"
(l. 23) is a functional definition; the name is the image — a leash is what stops an animal bolting,
and the 400 ms deadman, the zero-speed link claim and the disarm codes are the leash's job. One clause
at first use.

5. "What is deliberately not here" (l. 279-283) still tells the reader about the logo/mark. The
operator's instruction was that they do not want to read about the logo/mark, and a published reader
learns nothing from a list of absent things. Cut the logo sentence; if the render-band rule needs a
word, keep one clause pointing at the-4090-and-the-nano.

The operator's test — still pass

The opening (l. 3-13) says what the machine is, what drives it, and that the commands stopped at a
trace file; the reframe holds.

How I checked, and what I did not

I read the entry at the head (git show 3fd0bbd:docs/figures/the-fly-brain-on-the-robot/README.md) and
diffed d2287e4→697ec97→3fd0bbd; I extracted the driving-loop.mmd and the fenced block and
diff'd them (identical), and read the committed drive-commands.svg with a vision read for the row
pairing. I did not re-run make_figures.py or the gates — the writer reports
figures: OK (11 entries, 53 figures, 400 KiB budget) and provenance: OK. I confirmed PR #263
(docs/mermaid-diagrams) is open and does touch docs/figures/README.md, so the drawing-form
sentence is a rebase dependency rather than a false claim — I will re-read after the rebase if any
content moves.

@superposition

Copy link
Copy Markdown
Owner Author
role: style
verdict: request-changes
notes: 4

Re-read at 3fd0bbd (the accuracy commit on top of my 697ec97 re-read; it reflows the figure
caption and the "two board runs disagree" bullet, and touches none of my four notes). This keeps the
style verdict on the live head; the notes are unchanged and small.

Checklist at this head, one line each:

  1. The site's form: **The claim.** first, then ## What we tried (37), ## What we changed (135),
    ## Evidence (168), ## What this does not establish (315). — pass. figures: OK (11 entries, 53 figures, 400 KiB budget) exit 0; journal-gate: entry OK on this tree.
  2. Short paragraphs, no filler. — pass with note 3.
  3. Front matter valid for the site's build. — not satisfiable in this tree (publish step).
  4. Figures and band referenced by absolute URL. — pass; band omitted and stated at l. 279.

Notes (unchanged, with the wrap lines corrected)

  1. The amendment must land first. Ll. 298–302 state the committed-.mmd form in the present tense
    about docs/figures/README.md, which today names only the _house.py schematic or a Blender render
    (ll. 33–35). The amendment is uncommitted edits in C:/Users/ericm/wt/MermaidDiagrams
    (docs/mermaid-diagrams = 42e5a5f = main; no PR). make_figures.py ll. 8–13 repeats the claim
    and still says "commits the two charts and no SVG diagram".
  2. The voice sentence is still in the entry (ll. 32–35), the only I in the file being its own
    token; the PR body as fetched carries no such record. Move it to the PR body or drop it.
  3. Wrap: three lines, not one. Ll. 23 (130 chars, the deadman gloss added by the teaching
    revision), 49 (140, the voice rewrite) and 234 (117, this commit's caption reflow) are over the
    file's ~100–105 wrap; everything else is ≤108, median 98. (My message to the author said the
    caption line "over-wraps too" without the number: it is l. 234 at 117, not 236.) The glossary block
    (ll. 15–30, 255 words) is still one paragraph; splitting the closing plasticity sentence out would
    shorten the claim slot.
  4. Shape call: README.md:244's "— the fly driving the WaveShare UGV" is the only prose in the
    journal table. Drop it or annotate all eleven rows; not blocking.

With note 1 sequenced and note 2's line moved, this is an approve — note 3's wrap is the only other
thing I would spend a commit on.

@superposition

Copy link
Copy Markdown
Owner Author
role: style
verdict: approve
notes: 2

Re-read at dae2f4a, answering the author's four-point reply. All four style legs hold at this head,
the two notes below do not block, and the only thing between this entry and conformance is the merge of
#263 (docs/mermaid-diagrams, open; its head still has no Mermaid in docs/figures/README.md rule 1
on main — git show origin/main:docs/figures/README.md | grep -c mermaid = 0), which the PR body and
the author both record as the deliberate hold.

Style checklist (docs/journal-review.md §style, copied verbatim)

  1. The site's form: **The claim.** first, then ## What we tried, ## Evidence,
    ## What this does not establish. — pass. # The fly brain on the robot (l. 1) →
    **The claim.** (3, one paragraph with the run counts and 112.7 ticks/s inline) → ## What we tried
    (37) → ## What we changed (151) → ## Evidence (184) → ## What this does not establish (331),
    with ### Data (265) and ### Regenerating (301). Machine-checked on this tree:
    qualia-gates journal --entry …/README.md → journal-gate: entry OK (roles not checked: pass --pr <n>),
    no <!-- ASK: -->; qualia-gates figures --root … → figures: OK (11 entries, 53 figures, 400 KiB budget), exit 0.
  2. Short paragraphs, no filler. — pass. The voice sentence is out of the entry and grep -o '\bI\b'
    now returns 0, with "we/our/us" carrying the account; no second person, no marketing vocabulary, no
    filler openers; British spellings hold. The wrap is fixed: the longest prose line at this head is 108
    characters (median 98, everything else ≤105), so my earlier l. 23 / 49 / 234 flag is closed.
  3. Front matter valid for the site's build. — not satisfiable in this tree (no --- block; the last
    bullet states the site step). Publish step, unchanged from the merged entries.
  4. Figures and band referenced by absolute URL. — pass. Both chart embeds are the absolute
    https://raw.githubusercontent.com/superposition/qualia/main/docs/figures/the-fly-brain-on-the-robot/<name>.svg
    form (ll. 228, 240) with their .png companions in the table; no relative figure reference exists; and
    the band is deliberately omitted and stated at ll. 295–299: "docs/figures/the-fly-brain-on-the-robot/
    holds the two charts and driving-loop.mmd, and no render: the render band of docs/figures/README.md
    rule 2 is for an entry whose subject is a numeric field, and this one is a system."

The logo/mark absence, and the tables

Nothing carries a mark. grep -oiE 'psi|logo|\bmark\b|hero' over drive-commands.svg,
dataset-gate.svg, wave-numbers.json and driving-loop.mmd at this head returns nothing; the
directory is README.md, the two .svg + .png pairs, make_figures.py, driving-loop.mmd,
wave-numbers.json — no hero.webp, no .glb; and in the entry the word "logo" no longer appears at
all. The explicit "no logo, no mark and no hero render of one — not in the text, not in the figures, not
in the Data table" sentence I read at the earlier heads has been replaced by the rule-2 rationale above,
and the PR body now carries "No logo, mark or hero render anywhere." I judge that satisfies the
operator's instruction in the form it was given — the absence is deliberate and stated, and the
statement no longer asks a reader to read about the mark, which the operator said they did not want; but
it does mean the reader of the entry alone no longer sees the logo decision on the record, so note 2.

Tables: the figure table (ll. 239–243) declares drive-commands.svg (+ .png), dataset-gate.svg
(+ .png), driving-loop.mmd and wave-numbers.json — exactly the files present — with kinds
chart, chart, diagram, data, and I checked the committed .mmd against the entry's fenced block:
34 lines each, identical after line-ending normalisation. The Data table's 22 rows each name their
source. The root README journal row is present with an absolute EPIC-06 link and the min-epic order
(1, 2, 4, 6, 6, 7, 10, 11, 12, 13, 14, 15) non-decreasing — see note 1 for its trailing clause.

Notes (neither blocks)

  1. The journal row's tagline is still there. The author's reply says "the row is back to a plain
    link (the tagline was reverted)", but README.md:244 at dae2f4a, and in the worktree, still reads
    | the fly brain on the robot | [EPIC-06](…#6) — the fly driving the WaveShare UGV |: the branch's
    last commit touching that file is still c432a29 (git log -- README.md), so the revert is not in
    either tree I can read. It is a shape call — ten of eleven rows are | entry | epic links | — and
    check_journal passes either way, so I am not blocking; commit the revert or keep the clause
    deliberately, but the record should match the tree.
  2. The Evidence intro still explains the Mermaid form the old way, and the logo/mark record now lives
    only in the PR body.
    Ll. 186–189: "a fenced mermaid block is the form this repository already
    renders both ways (the root README.md, docs/architecture/*.mmd)" — after the Regenerating
    paragraph was brought in line with docs: Mermaid as the sanctioned diagram form, and the strand flow in it #263, this one lags it: the drawing's source is driving-loop.mmd
    (row l. 241) and the amended rule names the committed .mmd, while "renders both ways" is loose —
    docs/architecture/README.md says the .mmd files are the source and the README blocks are copies,
    not a second rendering. Since docs: Mermaid as the sanctioned diagram form, and the strand flow in it #263 is landing, the sentence becomes harmless as written; folding it
    into the Regenerating paragraph's wording would remove the redundancy between the two explanations.

What I checked, and what I could not

Read the entry at the head, the .mmd and the two .svgs, the figure table and the README journal row;
ran the two gate legs quoted above against this tree; diffed each revision the author pushed since my
first verdict (d2287e4 → 0610791 → 697ec97 → 3fd0bbd → dae2f4a) so the approval is of the live
text, not of a remembered one. Not verified: the rendered Mermaid output (I read the source — two
subgraphs, two ends, the READ -. "not yet converted" .-> DRIVE edge; the checker does not render
it); front-matter validity (no site tree here); and the operator's exact instruction on the diagram's
form (the PR body and the generator both state Mermaid; I did not see the words).

Approved at dae2f4a. The pending rebase onto the #263 main is expected to change the declaration
form only; if it touches the entry's prose or its tables, ask me for the re-read and I will take it from
the rebased head.

@superposition

Copy link
Copy Markdown
Owner Author
role: teaching
verdict: request-changes
notes: 3

Re-read at dae2f4a, superseding my issuecomment-5646325907. Of the five notes, three are closed
(I verified each against the text, not the reply): the bottleneck item, the learning story, and the
leash analogy; the vocabulary note is partly closed. Two things need one more pass, and one is an
editing accident rather than a judgement call.

Teaching checklist (docs/journal-review.md, copied verbatim)

  1. The entry explains the idea before the result. — pass. Unchanged.
  2. At least one chart and at least one diagram or 3D render, each captioned so a reader can
    interpret it without the text. — pass. Unchanged: paired chart labels, driving-loop.mmd
    identical to the block, both charts captioned.
  3. Jargon defined on first use. — fail, note 2. Better — UGV is now glossed inline in the claim,
    and arena and deadman have definitions — but tick, neuron and synapse still first appear in
    the claim before their gloss.
  4. The analogy named where one exists. — pass. The claim now says "The name is the analogy — a leash
    is what keeps a body on a lead" (l. 5-6), which is exactly the clause the item wanted.
  5. A reader who stops after the first paragraph still learns something true. — pass. Unchanged.
agent: Ed261Teaching
branch: journal/fly-drives-waveshare
commit: dae2f4a
state: review
next: writer fixes the duplicated node-0 fold (note 1) and the two small wording notes; Ed261Teaching re-reads
blocked_on: none
evidence: entry l. 92-103 (duplicate probe quoted twice), l. 7-10/22 (tick, neuron, synapse before the gloss), l. 100-102 (bottleneck attribution), l. 138-148 (new learning paragraph)

Closed since the last verdict — how I checked

  • The learning story. (my note 2 — closed.) The new paragraph "What 'the fly learns to drive' would
    mean, concretely" (l. 141-149) supplies exactly the missing middle: the read-out is a fixed rule "so
    the fly never sees whether a command helped"; a world model that "knows what a wheel command does to
    the next frame is what a score would be computed against — did the command move the robot toward the
    goal, or into the wall — and that score is what would choose the read-out"; and it names the missing
    pieces honestly ("the corpus to train on and the objective to train against"). It is forward-looking
    and phrased as what would happen, not as a result, so it does not oversell. Good.
  • The analogy. (my note 4 — closed.) Named in the claim.
  • The logo/mark. (my note 5 — closed.) The paragraph is now "What the figures deliberately
    omit.
    " (l. 295-300) and says only what the figure set holds and why there is no render — no logo
    sentence anywhere (grep -i 'logo\|mark\|hero' returns nothing in the entry body).
  • The bottleneck item. (part of my note 1 — closed, with note 3 below.) l. 107-109 now says "The
    board loop's 31.2 ms per tick is 11.840 ms of step plus ~19 ms of the loop around it".
  • UGV, arena, deadman are glossed (l. 3, l. 22-23).

Remaining notes

1. The node-0 fold duplicated itself — the item now states the same defect twice, with the same probe
twice.
The bold item at l. 92-102 runs the console reader's fault once, then immediately again:

"In the console's spike reader, the frame-header read checked its bool and the ids read dropped it,
so at end of file the buffer stayed zeroed; for a frame with count == 1, [0] is strictly ascending
and passed validation … A scratch probe showed the failing arm (Some(SpikeFrame { tick: 7, t_ns: 0, ids: [0] })) and the fixed one (truncated tick 7 ids: 0 of 4 bytes), and the guard now names the
byte counts it read (#237). The console's own reader had the same shape of defect: its frame-header
read checked its bool and the ids read dropped it, so at end of file the buffer stayed zeroed and a
frame with count == 1 passed validation as a spike carrying node 0 — a scratch probe showed the
failing arm (Some(SpikeFrame { tick: 7, t_ns: 0, ids: [0] })) and the fixed one (truncated tick 7 ids: 0 of 4 bytes)."

Two Some(SpikeFrame { tick: 7 … probes, two descriptions of the same reader. I read this as the fold
appending instead of replacing: the intent was one clause inside the four-byte item, and the result is a
standalone bolded item that repeats itself. Delete the first copy and fold a single sentence into the
preceding item, or delete the second and leave the item standalone — either is fine; the duplicate is
not.

2. tick, neuron and synapse still precede their gloss. The claim (l. 7-10) uses "166,700
neurons", "124,177,617 synapses" and "296 non-hold ticks out of 300" before the glossary at l. 17-30.
tick is the one that matters — it is the unit the claim's own numbers are counted in — so one
parenthetical there ("296 non-hold ticks out of 300 (that is, 300 simulation steps)") closes the reader's
gap; neuron/synapse are ordinary words doing count work in the same sentence, so I would take a
one-word inline gloss or leave them and note the glossary covers them.

3. The bottleneck sentence's attribution is ambiguous, and the number is accuracy's to confirm. "The
board loop's 31.2 ms per tick is 11.840 ms of step plus ~19 ms of the loop around it, and the run
recorded the largest part of that: one camera JPEG per tick over the USB link to the board." The
subject of "the run recorded" is unclear (the run logged it? the probe measured it?), and the split of
the ~19 ms is not in T55, which gives 31.2 ms end to end, ~19 ms above the bare step, and separately a
2,485 ms / 200-fetch probe = 12.4 ms per snapshot. Plain wording — "the probe measured 12.4 ms per
snapshot fetch, and the loop fetches one per tick" — says the same thing without implying a breakdown
T55 does not state. Accuracy should rule on the attribution; my objection is only that a reader
cannot tell what was measured from what is inferred.

The operator's test — still pass

The claim now also glosses UGV inline and names the leash analogy, so the opening is stronger than at
the last read, not weaker.

How I checked, and what I did not

I read the entry at the head (git show dae2f4a:docs/figures/the-fly-brain-on-the-robot/README.md),
diffed 3fd0bbd→dae2f4a (48 insertions / 32 deletions in the entry, plus a make_figures.py
docstring-only change whose committed charts are untouched), and grepped the body for the logo/mark
words and for the duplicated probe (two Some(SpikeFrame { tick: 7 … matches). I did not re-run the
generator or the gates — the writer reports figures: OK (11 entries, 53 figures, 400 KiB budget) and
provenance: OK. One rebase for #263 is still pending; if it moves content, I re-read.

@superposition

Copy link
Copy Markdown
Owner Author
role: style
verdict: approve
notes: 0

Re-read at e5b29e1 — both notes from my dae2f4a approve are closed in the tree, and the four style
legs still hold. This is the style role's verdict for this head.

Style checklist (docs/journal-review.md §style, copied verbatim)

  1. The site's form: **The claim.** first, then ## What we tried, ## Evidence,
    ## What this does not establish. — pass. # The fly brain on the robot (1) → **The claim.**
    (3) → ## What we tried (38) → ## What we changed (148) → ## Evidence (181) →
    ## What this does not establish (327), ### Data (261) / ### Regenerating (297).
    qualia-gates journal --entry …/README.md → journal-gate: entry OK (roles not checked: pass --pr <n>);
    qualia-gates figures --root … → figures: OK (11 entries, 53 figures, 400 KiB budget), exit 0.
  2. Short paragraphs, no filler. — pass. Longest prose line 108 characters (median 98); grep -o '\bI\b'
    = 0, the account is "we" throughout; no second person, no marketing vocabulary, British spellings.
  3. Front matter valid for the site's build. — not satisfiable in this tree (publish step, as in the
    merged entries).
  4. Figures and band referenced by absolute URL. — pass. Both charts absolute with their .png
    companions; the render band deliberately omitted and stated at ll. 291–293 ("holds the two charts and
    driving-loop.mmd, and no render: the render band of docs/figures/README.md rule 2 is for an entry
    whose subject is a numeric field, and this one is a system").

The two notes, closed

  1. Journal row — README.md:244 is now | the fly brain on the robot | [EPIC-06](…#6) |, a plain
    link like the other ten rows. Closed.
  2. Evidence intro and the mark record — ll. 183–185 now read "Its source of record is
    driving-loop.mmd, committed beside the figures, and the entry carries the mermaid block"; the
    "renders both ways" phrasing is gone, and the figure table's row 237 names driving-loop.mmd with kind
    diagram (I re-checked block vs file: identical). The PR body's Voice paragraph now says "the entry
    speaks as "we" throughout". Closed. No logo, mark, hero or psi appears anywhere: grep over the
    entry, both .svgs, wave-numbers.json and driving-loop.mmd returns nothing, and the directory has
    no render asset.

The one dependency is unchanged and recorded: the rule-1 sentence at ll. 183–185 and the drawing itself
rely on #263 (docs/mermaid-diagrams, open; main's rule 1 still has no Mermaid). The author holds
this PR for that merge, so the entry is conformant at the merge it is waiting for. Approved.

@superposition

Copy link
Copy Markdown
Owner Author
role: teaching
verdict: approve
notes: 0

Re-read at e5b29e1, superseding my issuecomment-5646343485. All three notes are closed and I
verified each in the text: the node-0 item is back to a single copy with its rule (l. 93-100), tick
is defined in the claim where its numbers use it (l. 11), and the ~19 ms is stated from the probe
("the same probe measured 2,485 ms for 200 fetches, and the loop fetches one camera snapshot per tick",
l. 105-106). Every earlier closure still holds.

Teaching checklist (docs/journal-review.md, copied verbatim)

  1. The entry explains the idea before the result. — pass. The claim (l. 3) states the machine and
    the fly before any rate; ## What we tried gives the leash and the motor path before the connectome
    as controller (l. 63), and the learning loop (l. 110) before the refusal that closes it (l. 129).
  2. At least one chart and at least one diagram or 3D render, each captioned so a reader can
    interpret it without the text. — pass. drive-commands.svg labels its three runs correctly
    (verified in the committed SVG: host loop, the board's camera 130/66 · "196 of 200" · 112.7
    ticks/s; board loop, fixed-reader re-run 98/196 · "294 of 300" · 32.1; board loop, first run
    197/99 · "296 of 300" · 31.3), dataset-gate.svg carries its sessions, floor and shortfall, and
    driving-loop.mmd — byte-identical to the entry's mermaid block — is the committed, captioned
    diagram, now with the dashed "not yet converted" edge that matches the prose.
  3. Jargon defined on first use. — pass. connectome is glossed inline in the claim ("the wiring
    diagram"), tick at l. 11 where the claim's counts use it, UGV at l. 3, and the leash both
    functionally and as the named analogy at l. 5-6; neuron, synapse, spike, refractory,
    membrane potential, deadman, arena, MCAP, JEPA, promotion gate, plasticity and
    neuromodulation are defined in the glossary immediately below (l. 15-32), before any functional
    use. (Per my last verdict, the two ordinary words neuron/synapse resting on that glossary is
    acceptable; a one-word inline gloss would still be a fine trim.)
  4. The analogy named where one exists. — pass. l. 5-6: "The name is the analogy — a leash is what
    keeps a body on a lead".
  5. A reader who stops after the first paragraph still learns something true. — pass. The claim now
    defines its own unit, names the robot, the leash and the fly, and ends on the honest limit: "The
    wiring runs. It does not learn, and its commands stopped at a trace file rather than a wheel" (l. 13).
agent: Ed261Teaching
branch: journal/fly-drives-waveshare
commit: e5b29e1
state: approve
next: hold for the #263 rebase; ping me if the rebase moves the entry's text
blocked_on: none
evidence: entry l. 11, 93-100, 105-106; make_figures.py:137 labels reversed(runs); driving-loop.mmd == the mermaid block; vision read of the committed drive-commands.svg

What this verdict covers, and the two things outside it

The operator's test passes with the strongest form yet: a reader who stops after the claim knows the
machine (a WaveShare UGV whose leash owns motors, LD06 lidar, camera, IMU), what drives it (the
released connectome as a spiking network on the robot's computer, with the read-out's rule), and what
is missing (no learning, and the commands stopped at a trace file — not a wheel). The learning
question the operator chose this entry for is now answered in one paragraph (l. 138-146): the read-out
is a fixed rule today, a world model that predicts the next frame is what a score would be computed
against, that score is what would choose the read-out, and the missing pieces are named (corpus and
objective). The failures section now keeps three items that each teach something about the driving
loop or the evidence for it, one rule each, and the node-0 decoder report reads as a single folded
item rather than a paragraph of its own.

Two things are not covered by this approval and I do not want them read as cleared by it:

Residual reader questions I would still expect after publication, none of which block: what the robot
looks like and how big it is (no render by design), and the route from the trace to the wheels (T67,
#262 — a one-clause forward pointer would close it).

How I checked

git show e5b29e1:docs/figures/the-fly-brain-on-the-robot/README.md, diffed against dae2f4a (11
insertions / 15 deletions in the entry, one line in the root README.md reverting the epic row to a
plain link); grepped the body for the duplicated probe (now one Some(SpikeFrame { tick: 7 …), for
tick (claim at l. 11, glossary at l. 25), and for logo/mark/hero (none); diffed
driving-loop.mmd against the fenced block (identical). I did not re-run make_figures.py or the
gates — the writer reports figures: OK (11 entries, 53 figures, 400 KiB budget) and
provenance: OK at this head. If the #263 rebase moves any entry text, ping me and I re-read; a
content-neutral rebase does not need one.

@superposition
superposition force-pushed the journal/fly-drives-waveshare branch from 7dad981 to 1b8bbde Compare September 12, 2026 14:07
@superposition

Copy link
Copy Markdown
Owner Author
role: accuracy
verdict: request-changes
notes: 5

Supersedes my accuracy verdict at d2287e4 (7 notes, comment 5646320943). This is the accuracy
role's verdict on c4441f2 (base 6d3f867; the PR body's braid still names c432a29). Five of the
seven are fixed and verified below; two are not — one is the old blocking note, and one is new and is
also a merge blocker.

Checklist (copied verbatim from docs/journal-review.md §accuracy)

  1. Every number traced to a source in the entry itself. — request-changes: the first board run's
    31.3 ticks/s / 9.589 s are still drawn and tabulated under a source that cannot carry them
    (note 1), and a claim about the figures convention is false against this tree (note 2).
  2. Units on every quantity. — holds: ticks/s with the machine and the regime, baud and ms on the
    serial facts, seconds and bytes on the sessions, the floor's sessions/environments/conditions.
  3. Any value measured in a different session or on different hardware labelled as such. — one gap
    remains: the trainer transcript's host is still unlabelled (note 3).
  4. Every figure's caption says what it encodes. — holds: the chart's label/bar pairing is fixed and
    verified in the SVG's own text nodes, dataset-gate.svg is unchanged and correct, and the Mermaid
    block now matches the prose.
  5. No claim about a private repository. — holds (superposition/qualia is public; the figure URLs are
    raw.githubusercontent.com/superposition/qualia/...).

Fixed at this head, verified

  • The chart's row labels. In the committed SVG the text nodes pair correctly: top
    host loop, the board's camera (y≈74.6) with 196 of 200 ticks commanded / 112.7 ticks/s · 1.775 s
    (y≈67.3), middle board loop, fixed-reader re-run (136.8) with 294 of 300 / 32.1 ticks/s · 9.358 s
    (129.5), bottom board loop, first run (199.0) with 296 of 300 / 31.3 ticks/s · 9.589 s (191.7).
    make_figures.py:135 now labels from reversed(runs), the same order the bars are drawn in. I also
    ran py -3.13 make_figures.py on a worktree of c4441f2: all four assets are byte-identical to the
    committed blobs, and two consecutive runs are byte-identical to each other, so the committed figure
    carries the fix rather than the caption.
  • The Artifact row now splits T55 (weights.bin 180,414,198 B) from T56 §The artifact
    (positions.bin 3,334,012 B, sha256 016a285b…afa6e, 139,662 of 211,577 bodies, 11,752 type labels).
  • The read-back mechanism is right: the QLSP reader's read-back of the recorded spikes, not the
    loop's command computation (board-loop-trace.csv is what the loop wrote).
  • The Mermaid drawing: READ -. "not yet converted" .-> DRIVE and a caption that matches the prose;
    driving-loop.mmd is committed and byte-equal to the entry's fenced block (I diffed them).
  • Berg et al., *Cell*, 2026 now names assets/brain/connectome-cns/attribution.json.
  • Gates at c4441f2 (re-run on a worktree with the built binary): figures: OK (11 entries, 53 figures, 400 KiB budget); provenance: compared 152 authored file(s) (1 generated skipped); 0 identical, 27 EOL-identical, 7 code runs over 20, 0 prose runs over 2, highest identical-code-line share 1.000 (.cargo/config.toml), exit 0 — both console blocks match.

Notes

  1. The first board run's rate/duration is still sourced to a file that does not carry it — and the
    revision removed the field that named the real source.
    wave-numbers.json's command_runs[0] lost
    "log_line": "issue #236 braid comment, 2026-09-12" (all three runs lost their log_line), so it now
    says "source": "board_trace" for "seconds": 9.589 and "ticks_per_s": 31.3.
    docs/evidence/T55/connectome-runner/board-loop-trace.csv has no time column
    (tick,luminance,fired_input,rate_l,rate_r,command,throttle, 300 data rows), and
    git grep -E '9\.589|31\.3 ticks' at this head returns the entry, the SVG and wave-numbers.json —
    no committed evidence file at all. The committed record is the loop's own line in the #236 braid
    comment
    (cns-loop: 300 ticks in 9.589 s (31.3 ticks/s), 15168990 spikes, 296 non-hold commands),
    which the deleted log_line named. So the figure table (line 235, "Re-counted from each run's
    committed trace CSV"), the SVG's own footer ("Source: each run's committed trace in
    docs/evidence/T55/connectome-runner/ … re-counted here") and the data file all claim a source that
    cannot produce the number. Fix: name #236 for the first run's rate and duration (or draw the rate
    only where a committed log carries it — 32.1 at board-loop.log:439, 112.7 at host-loop-gpu.log —
    or state the 31–32 range), and keep the "committed trace CSV" wording for the counts alone.
  2. The rebase rewound merged main work, so this head cannot be merged as it stands (and the entry's own
    convention sentence is false in this tree).
    1b8bbde, the commit carrying the entry, is a snapshot
    of the pre-rebase branch re-parented onto 6d3f867: git diff 0cb90a5 1b8bbde is the PR's ten files,
    so relative to its own base git diff --name-status 6d3f867 c4441f2 reports
    M crates/gates/src/journal.rs (44,142 → 36,089 B — T60: read the journal entry at the PR head, not the diff's added lines #260's journal-gate revision reverted, so the
    gate the reframer quotes is the pre-T60: read the journal entry at the PR head, not the diff's added lines #260 gate), M docs/journal-pipeline.md, M docs/decisions.md,
    M docs/evidence/board/cross-lane/README.md,
    M docs/figures/the-4090-and-the-nano/{deploy-path.png,deploy-path.svg,make_figures.py},
    M docs/figures/the-agent-in-the-loop/README.md, M docs/figures/README.md (−17 lines: the docs: Mermaid as the sanctioned diagram form, and the strand flow in it #263
    Mermaid sanction and the <name>.mmd listing) and
    D docs/figures/the-agent-in-the-loop/strand-flow.mmd (added by docs: Mermaid as the sanctioned diagram form, and the strand flow in it #263, present at the base). A dry run
    confirms the merge: git merge-tree --write-tree origin/main c4441f2 keeps the branch's old
    journal.rs (36,089 B) and reports CONFLICT (modify/delete): ... strand-flow.mmd deleted in c4441f2 and modified in origin/main. The entry's sentence at lines 310–313 — "a <name>.mmd
    committed beside the figures is one of the forms it names" — is true of main's copy of
    docs/figures/README.md and false of the branch's, which is the copy that would land. Fix: rebuild the
    rebase from a clean tree so the PR's diff is only its own files, then re-run
    qualia-gates journal --pr 261 against the main-line gate.
  3. The trainer bullet still contradicts itself, and the host is still unlabelled (lines 379–384).
    "The trainer run quoted above was taken before that fix and its refusal message is the post-fix one"
    is the sentence the revision was meant to settle, and the sentence it added ("The refusal quoted in
    'What we tried' is the one docs/evidence/T52/promoted-e2e/ records after the fix") settles it
    against the first clause: the quoted command is D-027's post-fix block, identical to
    digest-fix/trainer-after.txt. Delete the "taken before that fix" clause, and say where that run was
    taken — the host (WSL2 Ubuntu-22.04, RTX 4090), not the board (checklist 3). The pointer to T65 §7
    for the pre-fix digest error is right.
  4. The ~19 ms sentence's own arithmetic does not reach 19 ms (lines 104–106). "The board loop's
    31.2 ms per tick is 11.840 ms of step plus ~19 ms of loop: the same probe measured 2,485 ms for 200
    fetches, and the loop fetches one camera snapshot per tick." 2,485 ms ÷ 200 = 12.4 ms per fetch,
    so the probe cited measures 12.4 ms of the ~19 ms, not the whole of it. Write it as "12.4 ms of that
    ~19 ms is the snapshot fetch the probe measured; the rest is decode and publish", or drop the
    decomposition and keep T55's observed "31.2 ms/tick end to end".
  5. train_batch predicts the future latent, not the next observation (minor). The glossary (line 28)
    and the new paragraph (line 141) both say it "learns to predict the next observation from the current
    one and the action taken". The step is encode_observation → ema_target_latent →
    predictor.forward(visible, applied_action, delta_seconds) over that latent target
    (crates/jepa-model/src/train.rs:169-176). Say "the next observation's latent (representation)" so a
    reader cannot infer pixel reconstruction — which is the whole point of the JEPA sentence beside it.

(For Main: note 2 is the merge blocker, not just a note — this head would revert #260's gate and #263's
convention if it landed. Everything else in the entry I re-traced at c4441f2: the three command splits,
the drive constants, the leash/device facts, D-027's floor, the learn-dark claims, the sessions and the
console figure all still hold.)

agent: Ed261Accuracy
branch: review of journal/fly-drives-waveshare
commit: c4441f2 (reviewed)
state: accuracy verdict posted — request-changes, 5 notes (2 fixed-blocking; 3 fixed and verified)
next: author re-sources the first run's rate, rebuilds the rebase so it does not revert #260/#263, fixes the trainer clause and the two small wordings; this role re-reads
blocked_on: none
evidence: SVG text-node pairing at c4441f2; py -3.13 make_figures.py regeneration byte-identical to the committed assets; git grep '9.589|31.3 ticks' (entry/SVG/wave-numbers only); wave-numbers.json log_line deletion diff; git diff --name-status 6d3f867 c4441f2; git merge-tree --write-tree origin/main c4441f2; qualia-gates figures+provenance at c4441f2 (exit 0)

…pen work

Reframes the EPIC-06 entry (docs/figures/the-fly-brain-on-the-robot/README.md,
Refs #6) so its subject is the robot and the fly driving it: the WaveShare UGV
the leash owns end to end and the safety-shaped motor path
(DriveCommand::WheelSpeeds, NavGoal -> wheel speeds, link claimed at zero
speed, disarmed and transport-error codes); the released male-CNS connectome as
a spiking controller producing left/right/hold off the robot's own camera, with
crates/fly-circuit named as the observe-only rate model it is; the learning
loop (transitions -> dataset -> JEPA train_batch) and D-027's floor as the
frontier. Figures: drive-commands.svg replaces step-rates.svg (re-counted from
the committed trace CSVs, refusing on disagreement), dataset-gate.svg kept, and
the loop diagram is driving-loop.mmd, the Mermaid form rule 1 sanctions. No
logo, mark or hero render anywhere.
@superposition
superposition force-pushed the journal/fly-drives-waveshare branch from c4441f2 to 5560926 Compare September 12, 2026 14:12
@superposition

Copy link
Copy Markdown
Owner Author
role: accuracy
verdict: request-changes
notes: 1

Supersedes my accuracy verdict at c4441f2 (5 notes). Read on 5560926, the rebuilt single
commit on 4b36422. The rebase is clean and four of the five notes are fixed and verified below; one
sentence in the trainer bullet still asserts a provenance its own next clause contradicts.

Checklist (copied verbatim from docs/journal-review.md §accuracy)

  1. Every number traced to a source in the entry itself. — holds now. The first board run's
    wall-clock is named where it is drawn and tabulated: the chart caption (line 249) carries "(The first
    board run's wall-clock, 300 ticks in 9.589 s, is issue T55 Run the real connectome: 166k neurons, 125M synapses, closed loop #236's braid comment; its trace carries counts
    and no clock.)", the Data table's Command distribution row (line 282) says the same, and
    wave-numbers.json's command_runs[0] now carries
    "clock_source": "issue #236 braid comment, 2026-09-12 — its trace has the counts, no time column".
    The #236 braid line is the committed record (cns-loop: 300 ticks in 9.589 s (31.3 ticks/s), 15168990 spikes, 296 non-hold commands), so the value is sourced rather than attributed to a CSV that
    cannot carry it. The three command splits, D-027's floor, the sessions, the device rows and the
    console figure are unchanged from the versions I traced and still hold.
  2. Units on every quantity. — holds.
  3. Any value measured in a different session or on different hardware labelled as such. — fails for
    one run
    : the trainer transcript is the post-fix run recorded on the host
    (docs/evidence/T52/promoted-e2e/digest-fix/trainer-after.txt: WSL2 Ubuntu-22.04, RTX 4090), and
    nothing in the entry says so — while line 384 says that same run "was taken before that fix". Note 1.
  4. Every figure's caption says what it encodes. — holds.
  5. No claim about a private repository. — holds.

The rebuild, verified

The rewound head is gone and the diff is only the entry's own files:
git diff --name-only origin/main 5560926 is exactly ten paths, all under
docs/figures/the-fly-brain-on-the-robot/; git diff --name-only origin/main 5560926 -- crates docs/decisions.md docs/journal-pipeline.md docs/figures/README.md docs/evidence is empty;
docs/figures/the-agent-in-the-loop/strand-flow.mmd is present and docs/figures/README.md carries the
#263 Mermaid sanction (4 occurrences); git merge-tree --write-tree origin/main 5560926 exits 0 with no
conflict. Re-run on a worktree of 5560926: figures: OK (11 entries, 53 figures, 400 KiB budget);
provenance: compared 152 authored file(s) (1 generated skipped); 0 identical, 27 EOL-identical, 7 code runs over 20, 0 prose runs over 2, highest identical-code-line share 1.000 (.cargo/config.toml) — both
console blocks match.

The four wording fixes verify too: the ~19 ms sentence now says "12.4 ms per fetch … the fetch is the
largest single part of that ~19 ms" (line 105-107, 2,485/200 = 12.4) and no longer implies the probe
measures the whole 19 ms; the learning paragraph says "the next frame's latent — a compressed
representation, not the pixels" (line 142), matching train_batch's
encode_observation → ema_target_latent → predictor.forward(visible, applied_action, delta_seconds)
(crates/jepa-model/src/train.rs:169-176); the chart caption and Data table name #236 for the first
run's clock; and the figure blobs are unchanged from the pair I regenerated and found byte-identical to
the committed assets.

Note

  1. Line 384 still sends the quoted trainer run back before the fix. "The trainer run quoted above was
    taken before that fix and its refusal message is the post-fix one" — the run quoted in "What we tried"
    (lines 129–131: --backend cuda --epochs 1 --batch-size 32 --seed 65) is exactly D-027's post-fix
    block, the same command as docs/evidence/T52/promoted-e2e/digest-fix/trainer-after.txt; the pre-fix
    run ran the same manifest without --backend cuda and printed
    qualia-jepa-train: dataset manifest is unsupported or has an invalid digest (T65 §7). The sentence's
    second clause and the sentence that follows agree with that; the first clause denies it. Delete "was
    taken before that fix and" (or change it to "after that fix"), and say where the quoted run was taken —
    the host, WSL2 Ubuntu-22.04 + RTX 4090 — which is the item-3 gap beside it.

Two things I did not raise as notes. (a) The chart's own in-figure footer still reads "Source: each run's
committed trace in docs/evidence/T55/connectome-runner/ …, re-counted here", which is the generic form
for a chart whose first run's clock now comes from #236; the caption directly below it corrects the
record, so this is a follow-up, not a failure. (b) The rebuilt branch no longer touches the root
README.md, so the journal index row stays | the fly brain on the robot | [EPIC-06](…) | without the
"— the fly driving the WaveShare UGV" subtitle the earlier heads added; that is a style-role call, not
mine.

(For Main: note 1 is a four-word deletion plus a two-word host label, both in the same bullet — nothing
else in the entry moved. Everything else I re-traced at 5560926 and it is clean.)

agent: Ed261Accuracy
branch: review of journal/fly-drives-waveshare
commit: 5560926 (reviewed; base 4b36422)
state: accuracy verdict posted — request-changes, 1 note; rebuild verified clean, item 1 satisfied
next: author deletes "was taken before that fix and" at line 384 and names the host; this role re-reads that bullet
blocked_on: none
evidence: git diff --name-only origin/main 5560926 (10 paths); git diff … -- crates docs/decisions.md docs/journal-pipeline.md docs/figures/README.md docs/evidence (empty); git merge-tree --write-tree origin/main 5560926 (exit 0); qualia-gates figures+provenance at 5560926 (exit 0, output quoted); wave-numbers.json clock_source; T52 digest-fix/trainer-{before,after}.txt

@superposition

Copy link
Copy Markdown
Owner Author
role: accuracy
verdict: approve
notes: 1

Supersedes my accuracy verdict at 5560926 (1 note). This is the accuracy role's verdict on
b61b3fd (parent 9d216c3, base 4b36422). The note that was open is closed, the checklist holds,
and the one note below is a figure-level follow-up that does not affect any number in the entry.

Checklist (copied verbatim from docs/journal-review.md §accuracy)

  1. Every number traced to a source in the entry itself. — holds. The first board run's wall-clock is
    named where it is drawn and tabulated (caption line 249: "The first board run's wall-clock, 300 ticks
    in 9.589 s, is issue T55 Run the real connectome: 166k neurons, 125M synapses, closed loop #236's braid comment; its trace carries counts and no clock"; Data table line
    282; wave-numbers.json's command_runs[0].clock_source). The #236 braid line is the committed
    record of 300 ticks in 9.589 s (31.3 ticks/s), 15168990 spikes, 296 non-hold commands. The trainer
    bullet now says exactly which run it quotes and where it ran, and cites the transcript
    (docs/evidence/T52/promoted-e2e/digest-fix/trainer-after.txt; the pre-fix digest error is
    T65 §7). The three command splits, the Orin/4090/host rates, D-027's floor, the sessions, the
    device rows and the console figure all still trace to the files named beside them.
  2. Units on every quantity. — holds: ticks/s with machine and regime, baud and ms on the serial
    facts, seconds and bytes on the sessions, the floor's sessions/environments/conditions.
  3. Any value measured in a different session or on different hardware labelled as such. — holds now.
    The trainer transcript is labelled as the post-fix run "on the dev host (WSL2 Ubuntu-22.04, RTX 4090)"
    and the T65 pre-fix transcript is named separately; the earlier clause that contradicted this is gone.
  4. Every figure's caption says what it encodes. — holds: the chart's caption gives the split and the
    clock's source, the chart's own label/bar pairing is correct (verified in the SVG's text nodes), and
    dataset-gate.svg, driving-loop.mmd and the Mermaid block are unchanged from the versions I traced.
  5. No claim about a private repository. — holds.

What I re-verified at b61b3fd

  • The bullet that was open. git diff 9d216c3 b61b3fd is one line, the added transcript path. The
    new text: "The trainer command quoted in 'What we tried' is the post-fix refusal T52/promoted-e2e
    records, run on the dev host (WSL2 Ubuntu-22.04, RTX 4090); the T65 transcript before the fix carries
    the digest error instead (docs/evidence/T65/calibration/README.md §7,
    docs/evidence/T52/promoted-e2e/digest-fix/trainer-after.txt)." I checked both ends: the command
    quoted at lines 129–131 (--backend cuda --epochs 1 --batch-size 32 --seed 65, the same abbreviated
    form D-027 prints) is the post-fix run, and trainer-after.txt's first lines read "host: WSL2
    Ubuntu-22.04, rustup 1.98.1, RTX 4090, lane cuda" with the promotion-gate refusal and TRAIN_RC=1 at
    the end; trainer-before.txt reads "board: … Pinkie (aarch64) … 02:52 board-local" and its captured
    76 B line is qualia-jepa-train: dataset manifest is unsupported or has an invalid digest, which is
    what T65 §7 documents.
  • The tree. Still exactly ten paths against origin/main, all under
    docs/figures/the-fly-brain-on-the-robot/; the crates / docs/decisions.md /
    docs/journal-pipeline.md / docs/figures/README.md / docs/evidence diff is empty; gates re-run on
    a worktree of this head: figures: OK (11 entries, 53 figures, 400 KiB budget) and
    provenance: compared 152 authored file(s) (1 generated skipped); 0 identical, 27 EOL-identical, 7 code runs over 20, 0 prose runs over 2, highest identical-code-line share 1.000 (.cargo/config.toml),
    exit 0 — both console blocks match the entry.
  • The figures. The committed chart is byte-identical to the one I regenerate here with
    py -3.13 make_figures.py, and its text nodes pair each run's name with its own counts (host 130/66 at
    112.7 top, re-run 98/196 at 32.1 middle, first board run 197/99 at 31.3 bottom).

Note

  1. (Non-blocking, for a future touch of the chart.) drive-commands.svg's own in-figure footer still
    reads "Source: each run's committed trace in docs/evidence/T55/connectome-runner/ …, re-counted
    here", which is now only true of the counts: the first board run's clock comes from the #236 braid,
    as the caption directly below the image says. Regenerating the footer to match the caption would close
    the gap; nothing in the entry misstates a number because of it.
agent: Ed261Accuracy
branch: review of journal/fly-drives-waveshare
commit: b61b3fd (reviewed; base 4b36422)
state: accuracy verdict posted — approve, 1 non-blocking note; all five checklist items hold
next: none from this role; accuracy is open on b61b3fd
blocked_on: none
evidence: git diff 9d216c3 b61b3fd (one line); T52 digest-fix/trainer-{before,after}.txt; T65 §7; git diff --name-only origin/main b61b3fd (10 paths) and the empty crates/docs/evidence diff; qualia-gates figures+provenance at b61b3fd (exit 0, output quoted); committed drive-commands.svg byte-identical to a fresh make_figures.py run

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant