deepnsm-v2: frequency-ordered lexical evidence (D-LXC-1) + coverage-band measurement (D-LXC-11)

.claude/board/AGENT_LOG.md

@@ -1,3 +1,41 @@
+## 2026-09-30 — correction to the two D-LXC-1 / D-LXC-11 entries below
+
+- Their "70,396 triples", "G6 = 25" and "25 tags moved (G6 exact)" record
+  frequency changing tags: unintended semantic interference, not a passing
+  gate. The #1304 repair removed the counted pick; the tag is `main`'s legacy
+  tagging again. KJV: 70,393 triples, 1,227 subjects, 1,941 predicates,
+  identical to `main`. G6 is now the invariant
+  `counts_change_evidence_never_the_readings_or_the_tag`.
+
+## 2026-09-30 — D-LXC-11 coverage bands (5+3 council, orchestrator implements)
+
+- The 5 (Sonnet): prior-art, iron-rule, code truth (general-purpose,
+  runtime-archaeologist charter), cascade-impact, creative-explorer.
+  0 VIOLATES; findings drove 17 v1→v2 changes (rank rule from ndarray
+  `rank_per_10000`, D-LXC-12 dropped for D-LXC-5, `CoverageBand` naming,
+  T3 split, reading-share key F8).
+- The 3 (Sonnet): overclaim-auditor, dilution-collapse-sentinel,
+  firewall-warden. 0 BLOCK, 3 P1, 9 P2; v2→v3 fixed one rank rule named
+  everywhere, receipts for every number, gate renames G8a-c, commit contents.
+- Code: `bible_wave.rs` gains `CoverageBand`, `BandCuts`, `calibrate`, G8c;
+  8 new tests (T1-T8). No tag changes.
+- Gates: 126 lib + 17 example tests, clippy `-D warnings`, fmt clean. KJV:
+  70,396 triples, G6 = 25, G8c = population 141, cuts (72, 97), bands
+  34/71/36 — equal to the Python receipt in the plan.
+
+## 2026-09-29 — D-LXC-1 implemented (orchestrator, no agents)
+
+- `crates/deepnsm-v2/examples/bible_wave.rs`: counted PoS pick replaces the
+  `word_forms.csv` first-wins; lemma table still first. `Cargo.toml`:
+  `[[example]] bible_wave test = true`. 9 new tests; 124 lib tests unchanged.
+- Disable runs: 8 guards, each turned its named test red. One first attempt at
+  the first-wins disable was not the old rule and stayed green; replaced.
+- clippy `-D warnings` and fmt are clean.
+- KJV before/after: 70,393 → 70,396 triples; 25 tags moved (G6 exact); 107 of
+  771,176 tokens changed; long-range shares unchanged. Counts-first
+  alternative: 71,088 triples, 141 tags.
+- `.claude/settings.json` gains `attribution` (D-LXC-7).
+
 ## 2026-09-29 — D-LXC-1 plan rewritten with Read (orchestrator, no agents, no code)
 
 - Operator-directed. The council run below read source with shell

.claude/board/entries/2026-09-29-deepnsm-v2-counted-pick-tag-deltas.md

@@ -1,32 +1,62 @@
-# DeepNSM-v2 counted PoS pick: tag-change counts (2026-09-29)
+# DeepNSM-v2 counted PoS pick: tag changes and the KJV run (2026-09-29)
 
-**Status:** MEASURED (offline, over committed CSVs plus the released KJV
-vocabulary). No code changed.
+**Status:** MEASURED. D-LXC-1 is implemented in `bible_wave` (this PR).
 
 The removed `deepnsm-v2/src/lexicon.rs` (`68955ecb^`) assumed "the first row
 [of a frequency-ranked table] is the dominant reading". For `word_forms.csv`
 that is false: the file is ordered by lemma rank, and 259 surfaces have a
-first row that is not their most frequent reading (e.g. `changes`: first row
-v 13,624 at `:707`, noun n 113,085 at `:853`).
+first row that is not their most frequent reading. For example, `changes` has
+its verb row first (v 13,624 at `:707`), but the noun row (`:853`) counts
+113,085.
 
-Replacing that first-wins layer in `bible_wave::load_pos` with a counted pick
-(highest summed `wordFreq` per folded parser state, tie order Noun > Verb >
-Adj > Det > Other) changes:
+D-LXC-1 replaces that first-wins layer with a counted pick: the highest summed
+`wordFreq` per folded parser state, ties in the order Noun > Verb > Adj > Det
+> Other. The lemma table stays first.
 
-| variant | over both tables | within `bible_vocab.txt` (12,543) |
-|---|---|---|
-| B: lemma table first, as pinned by `ec50f07b` (the plan's choice) | 105 | **25** |
-| A: counted pick first, lemma table as fallback | 288 (183 lemma-derived) | 141 |
+| | before (first-wins) | after (B, shipped) | A (counts first) |
+|---|---|---|---|
+| in-vocabulary tags moved | — | 25 | 141 |
+| KJV triples (31,102 verses) | 70,393 | 70,396 | 71,088 |
+| distinct subjects | 1,227 | 1,237 | 1,243 |
+| same-subject links beyond ±5 / ±8 | 60.3% / 52.7% | 60.3% / 52.7% | 60.3% / 52.6% |
+
+Under B, 107 of 771,176 in-vocabulary KJV tokens change tag. The offline
+predictions (25 and 141 words) match the in-code G6 count exactly. Whether the
+moved tags are more correct is not measured.
 
 Method: Python over `crates/deepnsm/word_frequency/{lemmas_5k,word_forms}.csv`
-and `bible_vocab.txt` from release `v0.1.0-cam96-data`. Surfaces are
-lowercased as `load_pos` does. Fold: `n|p→Noun, v→Verb, j→Adj, a|d→Det,
-else Other`. Whether either variant tags the KJV better is not measured; the
-KJV before/after is gate G5 of D-LXC-1.
+and `bible_vocab.txt` (release `v0.1.0-cam96-data`); `bible_wave` run on
+Gutenberg #10 with the same release artifacts, `main` `5282dfa3` against this
+branch. The KJV text is input only and is not committed.
+
+Reproduce: fetch the release assets per `crates/deepnsm-v2/data/README.md`,
+then `DEEPNSM_V2_DATA=<dir> cargo run --release --example bible_wave --
+<pg10.txt>`. The in-code G6 count recomputes the 25 and asserts it. The 141
+(variant A) was measured once in Python and is not re-run by any gate. The
+`bible_vocab.txt` used had SHA-256
+`8dc3a65dcd3af38a2f53308fb96ef5fb5b34336c14587f6971b75ac966c3e212` (recorded
+as provenance, not a gate).
 
 Pre-existing observations filed with the plan:
 - `archaic_pos` never fires for COCA-known words such as `art` (D-LXC-9).
-- No in-crate tagger in deepnsm-v2 produces `Pos::Rel`; external callers can
-  still pass it (D-LXC-10).
+- No in-crate tagger produces `Pos::Rel`; external callers can still pass it
+  (D-LXC-10).
+
+**Amended (same day).** The summed pick was replaced by the register:
+`LexicalEvidence` stores readings most frequent first with a cumulative
+percentile coverage, and the tagger reads position 0. The KJV run is identical
+to "after (B)" above (25 moved, 70,396 triples); the summing moved nothing.
 
 Plan: `.claude/plans/deepnsm-v2-lexical-evidence-consumer-v1.md`.
+
+**Correction (2026-09-30, operator ruling on #1304).** Frequency is evidence,
+not a lexical decision. The 25 moved tags and the 70,396 triples above record
+frequency changing the tag: unintended semantic interference, not a
+correction. The counted pick and `dominant_pos` are removed. The tag is
+`main`'s tagging again, named `load_pos_legacy_first_wins`; its dependence on
+source-row order is inherited debt pending D-LXC-2/D-LXC-3, not an authorized
+resolver. KJV after the repair: 70,393 triples, 1,227 subjects, 1,941
+predicates — identical to `main`. G6 is now a paired invariant pinned by
+`counts_change_evidence_never_the_readings_or_the_tag`: count changes may move
+the order and the coverage, never the reading set or the tag. The table above
+is kept as the historical measurement.

.claude/board/entries/2026-09-30-deepnsm-v2-coverage-bands.md

@@ -0,0 +1,40 @@
+# DeepNSM-v2 coverage bands at population-calibrated thresholds (2026-09-30)
+
+**Status:** MEASURED. D-LXC-11 is implemented in `bible_wave` (report only;
+no tag changes).
+
+Each word's dominant reading share (`coverage(id)[0]`, from D-LXC-1's
+frequency-ordered register) is banded Contested / Leaning / Decisive at the
+quartiles of the band population, calibrated once at load. The population is
+the vocabulary words that are not lemma-table keys, have known coverage, and
+have readings that fold to at least two parser states.
+
+On the release KJV vocabulary (`v0.1.0-cam96-data`):
+
+| quantity | value |
+|---|---|
+| band population | 141 (of 635 cross-state words; 494 are lemma-table keys) |
+| cuts (lo, hi), `rank_per_10000` indices 35 / 105 | 72, 97 |
+| share range | 50 … 99 |
+| contested / leaning / decisive | 34 / 71 / 36 |
+| absolute alternative "< 60 contested" | 20 |
+
+Contested words include flies, leaves, lies, promises, needs, means.
+
+The Rust gate G8c and the Python receipt in the plan give the same numbers.
+The KJV run is otherwise unchanged (70,396 triples, G6 = 25).
+
+Open for the operator: relative bands (this) vs an absolute floor.
+
+Ready-to-paste STATUS_BOARD row (the file is deny-listed for the agent):
+
+`| D-LXC-11 | coverage bands (population quartiles) in bible_wave | In PR | deepnsm-v2-coverage-bands-v1 |`
+
+Plan: `.claude/plans/deepnsm-v2-coverage-bands-v1.md`.
+
+**Correction (2026-09-30).** The bands are a measurement, not a decision: the
+labels are report vocabulary in `bible_wave`, nothing reads a band to select
+or drop a reading, and no downstream consumer exists. The numbers (141,
+(72, 97), 34 / 71 / 36) stand. "70,396 triples, G6 = 25" above is stale: after
+the #1304 repair the KJV run is 70,393 triples, identical to `main`, and G6 is
+the non-interference invariant (see the 2026-09-29 entry's correction).

.claude/board/entries/README.md

@@ -25,10 +25,11 @@ 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.
 
-176 entries, 2026-08-06 .. 2026-09-29.
+177 entries, 2026-08-06 .. 2026-09-30.
 
 | date | entry id | finding | file |
 |---|---|---|---|
+| 2026-09-30 | `deepnsm-v2-coverage-bands` |  | [2026-09-30-deepnsm-v2-coverage-bands.md](2026-09-30-deepnsm-v2-coverage-bands.md) |
 | 2026-09-29 | `deepnsm-v2-counted-pick-tag-deltas` |  | [2026-09-29-deepnsm-v2-counted-pick-tag-deltas.md](2026-09-29-deepnsm-v2-counted-pick-tag-deltas.md) |
 | 2026-09-26 | `deepnsm-v2-lexical-evidence-survives-routing` |  | [2026-09-26-deepnsm-v2-lexical-evidence-survives-routing.md](2026-09-26-deepnsm-v2-lexical-evidence-survives-routing.md) |
 | 2026-09-25 | `window-scheduling-and-two-level-ternlog` |  | [2026-09-25-window-scheduling-and-two-level-ternlog.md](2026-09-25-window-scheduling-and-two-level-ternlog.md) |