D-PUZZLE-ATTN-0: entropy as the focus of attention on crosswords

.claude/board/LATEST_STATE.md

@@ -1,3 +1,8 @@
+## 2026-10-07 — lance-graph #1397 MERGED (`4a0a72c`); D-PUZZLE-ATTN-0 on branch
+
+- #1397: D-CE64-CYCLE-0, probe-only (no contract change).
+- D-PUZZLE-ATTN-0 (branch `claude/epistemicstate5-ce64-bits-3oe0dq`, unmerged): the crossword core moves to `cognitive-shader-driver/examples/shared/crossword_core.rs`; `crossword_attention_probe.rs` compares slot policies. No contract change.
+
 ## 2026-10-07 — lance-graph #1395 MERGED (`5658c10`); D-CE64-CYCLE-0 on branch
 
 - #1395: D-CE64-LOOP-0 and D-CE64-STAUNEN-0, probes in `lance-graph-planner/examples` (no contract change).

.claude/board/PR_ARC_INVENTORY.md

@@ -1,3 +1,10 @@
+## 2026-10-07 — lance-graph PR #1397 (merged `4a0a72c`, branch `claude/epistemicstate5-ce64-bits-3oe0dq`, head `ea16f16`, 3 commits) — D-CE64-CYCLE-0: what a CE64 register must carry across cycles
+
+- **Added:** `staunen_scheduler_probe.rs` gains `run_with(opts)`, a lazy-stamp heap scheduler, a FIFO worklist baseline, wall-clock windows with continuation, `Forget` arms, the #1390 Simpson + trial population, attractor and quadrant readings, dense 1k..1M windows; comparative tests pool 3 seeds.
+- **Locked (test-pinned):** the next cycle needs the evidence table, the register (Epi5) and the attention state; H and bits 40..42 are derived / echo; no wake advantage over round-robin when pooled (corrects #1395). 23 tests, 19 disable runs red.
+- **Deferred / open:** the heap scheduler costs ~3.5× FIFO per fold; Moore/EWA/Sudoku folds not in the loop; where cross-cycle state lives in the register.
+- **Confidence:** medium — pooled over 3 seeds, one world model.
+
 ## 2026-10-07 — lance-graph PR #1395 (merged `5658c10`, branch `claude/epistemicstate5-ce64-bits-3oe0dq`, head `086189a`, 5 commits) — D-CE64-LOOP-0 + D-CE64-STAUNEN-0: self-orchestrating fold loop, surprisal scheduler, bits 40..63 census
 
 - **Added:** `lance-graph-planner/examples/self_orchestration_probe.rs` (one register, five policies, 15 tests) and `staunen_scheduler_probe.rs` (many basins across windows, surprisal distinct from Shannon H, 18 tests), both `test = true`; the bits 40..63 reader/writer census (board entry).

.claude/board/STATUS_BOARD.md

@@ -48,7 +48,8 @@ Plan: `.claude/plans/2026-10-06-global-sudoku-replayable-orchestration-v1.md`. O
 | **D-EPI-CANON-0** | Canonical joint `EpistemicState5` reading of bits 59..63 in the contract | Shipped (#1381) | `contract::epistemic_state5`: Cartesian `Topology2 × Certification3`, `raw5 = t \| c << 2`, 24 meaningful / 8 reserved; per-factor facts; declarations + `project_state5`; `CausalEdge64::{epistemic_raw5, with_epistemic_raw5}`; non-declared-reading writers deprecated |
 | **D-EPI-POP-0** | Facts → 32-state population: the missing operator between one edge's coordinate (u5) and a semantic condition over all coordinates (u32) | Shipped (#1381) | `contract::epistemic_state5::{Population, facts_population, EpistemicState5::bit}`: `const fn`, no DSL / query object / planner. Tests pin exact sets (`CAUSES` = {20..23}, `IND_UNKNOWN \| RELATED` = {10,14,18,22}, `IND_KNOWN \| SUPPORTS` = {13,17,21}, `RELATED` = 8..23), operator == state enumeration over all 2048 requirements with no reserved code, conjunction == intersection; 2 disable runs red (any-of semantics; reserved read as fact-less) |
 | **D-EPI-LEGACY-DEPROJECT-0** | Deproject the remaining `ReasoningBand` / CE64-`TrustTexture` consumers: each goes to an exact coordinate, a `facts_population(..)`, MUL, or another subsystem (e.g. counterfactual context) — never a 1:1 replacement enum | Shipped (#1381) | `entries/2026-10-07-epi-legacy-deprojection.md`: 6 ReasoningBand caller sites (1 certification question, 1 non-interference witness → both exact coordinates; 2 context/obsolete → compat; 2 refusal witnesses), 0 production CE64 `TrustTexture` consumers, 4 `TrustTexture` types with no bridge; `EdgeRole::band` removed; hinge + V3 lift migrated; 4 code loops → `facts_population` |
-| **D-CE64-CYCLE-0** | Cross-cycle executor experiments on the surprisal scheduler: forget arms, FIFO baseline, 350 ms windows, 1k..1M folds, attractors | In PR | the next cycle needs evidence table + register (Epi5) + attention state; H and bits 40..42 are derived/echo; pooled 3 seeds: no wake advantage over round-robin, 11% fewer folds; FIFO fits 1M folds in a 350 ms window at ~220 ns/fold, the heap scheduler ~440k at ~800 ns (`entries/2026-10-07-ce64-cycle-0-cross-cycle-executor.md`) |
+| **D-PUZZLE-ATTN-0** | Entropy as the focus of attention on crosswords (`cognitive-shader-driver/examples/crossword_attention_probe.rs`, shared `crossword_core.rs`) | In PR | popcount (uniform entropy) needs 14.5× fewer nodes than fixed order to prove uniqueness; frequency-weighted Shannon costs 5.9× more than popcount; dom/wdeg reaches first fills sooner (41k vs 92k on open puzzles); contradiction rate flat within a search (`entries/2026-10-07-puzzle-attn-0-entropy-attention-crosswords.md`) |
+| **D-CE64-CYCLE-0** | Cross-cycle executor experiments on the surprisal scheduler: forget arms, FIFO baseline, 350 ms windows, 1k..1M folds, attractors | Shipped (#1397) | the next cycle needs evidence table + register (Epi5) + attention state; H and bits 40..42 are derived/echo; pooled 3 seeds: no wake advantage over round-robin, 11% fewer folds; FIFO fits 1M folds in a 350 ms window at ~220 ns/fold, the heap scheduler ~440k at ~800 ns (`entries/2026-10-07-ce64-cycle-0-cross-cycle-executor.md`) |
 | **D-CE64-STAUNEN-0** | Surprisal scheduler over many CE64 basins across windows (`lance-graph-planner/examples/staunen_scheduler_probe.rs`) | Shipped (#1395); wake claim corrected by D-CE64-CYCLE-0 (pooled: no wake advantage) | naive raw surprise is captured by noise (resolved 0.649); habituation + learning progress fix it (0.952, wake 4.2 windows vs round-robin 5.1); round-robin still resolves slightly more (0.964); bits 50..52 left as plasticity (`learn` gates on them); 18 tests, 13 disable runs red (`entries/2026-10-07-ce64-staunen-0-surprisal-scheduler.md`) |
 | **D-CE64-LOOP-0** | Bounded self-orchestrating fold loop over one CE64 register (`lance-graph-planner/examples/self_orchestration_probe.rs`) + bits 40..63 census | Shipped (#1395) | combined policy reaches `IndirectKnown × Causes` with H ≤ 0.5 bit in 3 folds (round-robin 10, random median 14); ~830 ns/fold release; 15 tests in CI, 11 disable runs red; H and activation stay out of the register (`entries/2026-10-07-ce64-loop-0-self-orchestration.md`) |
 | **D-PEARL-PROD-0** | Pearl's ladder in production: `lance_graph_planner::pearl::{reason, revise, hydrate}` over `lance_graph_contract::certification` | Shipped (#1391) | the P7a folds become `contract::certification::{CertificationModel, ModelBuilder}`; the dispatch, write-back and topology gate become `planner::pearl`, taking the caller's declared reading. The 14 falsifier tests run in the planner suite (CI); 12 disable runs red against the production module. No runtime caller yet: nothing in production produces sealed populations or chains for it (`entries/2026-10-07-pearl-prod-0-production-wiring.md`) |

.claude/board/entries/2026-10-07-puzzle-attn-0-entropy-attention-crosswords.md

@@ -0,0 +1,59 @@
+# 2026-10-07 — D-PUZZLE-ATTN-0: on crosswords, uniform entropy focuses search; frequency-weighted Shannon misdirects it
+
+**Status:** TEST-PINNED (`crates/cognitive-shader-driver/examples/crossword_attention_probe.rs`, `test = true`, 9 tests run by the shader-driver `cargo test` step). MEASURED (`cargo run --release -p cognitive-shader-driver --example crossword_attention_probe`; 20 created puzzles and 20 empty NYT grids, sides 5 and 7; 200,000 nodes per run).
+
+## DECISION
+
+- **SCOPE:** crossword solving is a fixed point (#1387), so deduction needs no attention. This probe changes only the search's choice of the next slot once deduction stalls. Everything else is held fixed:
+  - the puzzles;
+  - the propagation, moved unchanged into `examples/shared/crossword_core.rs` (#1387's 20 tests pass on it);
+  - the value order, most frequent word first.
+- **Policies:**
+  - fixed order;
+  - seeded random;
+  - widest first (negative control);
+  - **popcount**: most-constrained first, i.e. entropy under a uniform prior, `log2(count)`;
+  - **Shannon**: `H = -Σ p log2 p` with `p ∝` COCA-All frequency;
+  - **dom/wdeg** (Boussemart et al. 2004): count divided by accumulated contradictions on the slot that ran dry, so earlier failures steer later choices.
+- **Contradiction:** a slot with no candidate has no entropy under either reading and is never settled (pinned).
+
+## Measured
+
+| policy | prove: solved | prove: nodes | open: solved | open: nodes / to first | fill: solved |
+|---|---|---|---|---|---|
+| fixed | 16/20 | 852,535 | 7/20 | 2,858,517 / 2,488,660 | 5/20 |
+| random | 15/20 | 1,049,801 | 5/20 | 3,151,741 / 2,572,628 | 0/20 |
+| widest | 16/20 | 916,437 | 5/20 | 3,186,550 / 2,653,695 | 10/20 |
+| **popcount** | **20/20** | **58,729** | **20/20** | 570,450 / 91,930 | 10/20 |
+| Shannon | 19/20 | 348,578 | 16/20 | 1,156,793 / 550,562 | 10/20 |
+| **dom/wdeg** | **20/20** | 53,892 | **20/20** | 539,520 / **40,938** | 10/20 |
+
+(prove = count fills to 2 on created puzzles; open = half the givens, count to 50; fill = an empty grid.)
+
+- **Entropy as focus of attention works, read as uniform entropy.** Popcount needs 14.5× fewer nodes than fixed order to prove uniqueness, and 5× fewer on the open puzzles, where fixed order finishes only 7 of 20.
+- **Frequency-weighted Shannon is worse than popcount**: 5.9× the nodes on prove; on open, 2× the nodes and 4 puzzles unfinished. It picks a different slot on 36–50% of decisions.
+  - The frequency prior measures how *predictable* a slot is; a search pays for how many candidates it must *try*.
+  - So for attention in search, `log2(count)` is the right reading, not a proxy for something better.
+  - The `ms` column overstates popcount: its runs also compute Shannon on every decision to count disagreements.
+- **Earlier contradictions do direct later choices.** dom/wdeg uses about the same total nodes as popcount but reaches the first fill sooner: 33,017 vs 41,029 on prove, 40,938 vs 91,930 on open.
+- **The contradiction rate does not improve within a search.** Per tenth of the search it stays flat at 0.73–0.85 for both popcount and dom/wdeg. The gain shows up as an earlier solution, not a lower failure rate.
+- **Empty grids:** 10 of 20 stay unfilled within 200,000 nodes under every good policy. This is likely the vocabulary rather than attention; not measured.
+
+## Gates
+
+7 disable runs, each red on its named falsifier:
+- popcount picks the first slot
+- Shannon reads popcount
+- an empty mask has H = 0
+- a flat prior
+- no failure weights
+- no propagation
+- a wall-clock seed
+
+One falsifier was vacuous on first write: the dom/wdeg "differs from popcount" check compared whole stats, which differ anyway through popcount-only counters. It now compares the search path, and its disable goes red.
+
+## OPEN
+
+- Shannon may belong in *value* ordering (which word to try first) rather than slot choice. Untested: the value order here is fixed by frequency.
+- Coverage is narrow: one frequency prior (COCA academic), one vocabulary, sides 5 and 7 only. Side 9 takes minutes per created puzzle.
+- No CE64 register in this loop. The slot state is the mask (the evidence); entropy is derived from it.

.claude/board/entries/README.md

@@ -25,12 +25,13 @@ index row, (3) no duplicate entry id. Checks 1 and 2 are deliberately
 opposite directions; the stranding this convention prevents shows up in
 exactly one of them, never both.
 
-260 entries, 2026-08-06 .. 2026-10-07.
+262 entries, 2026-08-06 .. 2026-10-07.
 
 | date | entry id | finding | file |
 |---|---|---|---|
 | 2026-10-07 | `tinker-janus-fold-harvest` | TinkerPop bulk = K (GroupReduce), ONE_BULK = S, GValue pinning = bundle invalidation; JanusGraph slice = OrderedLaneWitness→Range (30–38×); no new V4 op | [2026-10-07-tinker-janus-fold-harvest.md](2026-10-07-tinker-janus-fold-harvest.md) |
 | 2026-10-07 | `selector-seq-aware-coverage` |  | [2026-10-07-selector-seq-aware-coverage.md](2026-10-07-selector-seq-aware-coverage.md) |
+| 2026-10-07 | `D-PUZZLE-ATTN-0` |  | [2026-10-07-puzzle-attn-0-entropy-attention-crosswords.md](2026-10-07-puzzle-attn-0-entropy-attention-crosswords.md) |
 | 2026-10-07 | `D-PUZZLE-0` |  | [2026-10-07-puzzle-0-sudoku-population-fold.md](2026-10-07-puzzle-0-sudoku-population-fold.md) |
 | 2026-10-07 | `D-PUZZLE-0-XWORD` |  | [2026-10-07-puzzle-0-crossword-same-algebra.md](2026-10-07-puzzle-0-crossword-same-algebra.md) |
 | 2026-10-07 | `D-PUZZLE-0-XWORD-REAL` |  | [2026-10-07-puzzle-0-crossword-real-words.md](2026-10-07-puzzle-0-crossword-real-words.md) |

crates/cognitive-shader-driver/Cargo.toml

@@ -171,6 +171,12 @@ test = true
 name = "crossword_real_words_probe"
 test = true
 
+# D-PUZZLE-ATTN-0: entropy as the focus of attention on crosswords (search
+# slot policies over the shared crossword core); tests run under `cargo test`.
+[[example]]
+name = "crossword_attention_probe"
+test = true
+
 [[example]]
 name = "crossword_mask_propagation_probe"
 test = true