Repository navigation
D-CE64-CYCLE-0 + D-PUZZLE-ATTN-0: cross-cycle state, and entropy as the focus of attention on crosswords - #1397
Conversation
…eduler run_with(opts): heap scheduler (lazy stamps), FIFO worklist baseline, wall-clock window budget with continuation, Forget arms at window boundaries (belief, question bits, register, evidence, attention), #1390 Simpson+trial population, fold quadrants, polarity, end attractors, productive/redundant/rejected accounting, dense 1k..1M fold windows. Comparative tests now pool three seeds. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0177nSSyQaXqbNuGR1vMXXN6
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0177nSSyQaXqbNuGR1vMXXN6
…ake claim Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0177nSSyQaXqbNuGR1vMXXN6
|
Note Currently processing new changes in this PR. This may take a few minutes, please wait... ⚙️ Run configuration
📒 Files selected for processing (6)
✨ Finishing Touches📝 Generate docstrings
Warning Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption. Comment |
|
Add Carrot credits or activate Agent usage billing to use Autopilot |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ea16f1643f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| rep.belief_updates_no_epi += | ||
| u64::from(!epi_moved && (dh.is_some_and(|d| d.abs() > eps) || st.activation != 0)); |
There was a problem hiding this comment.
belief_updates_no_epi does not actually count belief updates: a nonzero st.activation from Fold::Confound increments it even though the non-hydration branch never changes b.belief, while two different posteriors can have equal entropy and a zero/quantized activation and therefore be missed. Consequently the reported “4648 belief moves” and the derived conclusion about how many updates W would reference are not supported; record and compare the posterior before and after each fold instead.
Useful? React with 👍 / 👎.
| Forget::Attention => { | ||
| b.habit = 0.0; | ||
| b.progress = PROGRESS_PRIOR; | ||
| b.coh = 0.0; | ||
| b.prev_act = 0; | ||
| b.surprise = 0.0; | ||
| } |
There was a problem hiding this comment.
The Forget::Attention experiment retains h_prev_end, even though that cross-window value is the input used to calculate the next learning-progress sample at lines 1397–1400. Thus this arm still carries learned attention history across every boundary, so its measured resolution cannot establish what happens when attention is actually forgotten; reset h_prev_end to a neutral/rebuilt value along with progress.
Useful? React with 👍 / 👎.
| let epoch = ((w * EPOCHS as u64) / s.windows.max(1)) as usize; | ||
| let last_epoch = epoch >= EPOCHS - 1 || s.windows < EPOCHS as u64; | ||
| let t0 = Instant::now(); |
There was a problem hiding this comment.
For the dense 327,680-basin runs, the 350 ms “per-window” clock starts only after event delivery and the full agg initialization pass, and window_ns is captured before the full end-of-window update/requeue pass. Both omitted sections are O(number of basins), so a window can exceed the advertised cap and the printed max ms/w and folds/s can materially understate its runtime; start timing before world.events and check/record through the post-processing pass.
AGENTS.md reference: AGENTS.md:L24-L27
Useful? React with 👍 / 👎.
AdaWorldAPI
changed the title
…ped) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0177nSSyQaXqbNuGR1vMXXN6
2 participants
This PR carries two probes. Neither changes production code, and both run under CI (
test = true).1. D-CE64-CYCLE-0: what a CE64 register must carry across cycles
The D-CE64-STAUNEN-0 executor (
lance-graph-planner/examples/staunen_scheduler_probe.rs) is extended, not duplicated:run_with(opts)Forgetarms at window boundariesThe comparative tests pool three seeds.
State forgotten at every window boundary:
W: belief moved on 4648 folds and Epi5 on 72;
revisenever changed F/C.Baselines (pooled, 3 seeds):
Correction to #1395: the "wake 4.2 vs 5.1" there was one seed. Pooled, the full scheduler is no faster to wake, and both simple baselines resolve more. What survives: it uses 11% fewer folds and wastes half as many on settled basins.
1M folds per window: FIFO runs about 220 ns per fold, so a full million fits in ~252 ms. The heap scheduler runs about 800 ns per fold and is cut at ~440k folds per 350 ms window. Within a window, later folds are less productive (0.95 → 0.48).
2. D-PUZZLE-ATTN-0 — moved to #1398
The crossword attention commits were pushed after this PR merged, so none of them are in it. They are now in #1398.
Falsifiers
Every disable run went red on its named falsifier.
Limitations
Board: the CYCLE-0 entry, the #1395 arc entry and LATEST_STATE, and status rows.
🤖 Generated with Claude Code
https://claude.ai/code/session_0177nSSyQaXqbNuGR1vMXXN6