AdaWorldAPI · Oct 14, 2026 · Oct 14, 2026 · Oct 14, 2026
diff --git a/.claude/board/entries/2026-10-14-ledger-revision-after-1553.md b/.claude/board/entries/2026-10-14-ledger-revision-after-1553.md
@@ -0,0 +1,17 @@
+# 2026-10-14 — Recovery ledger revised against #1495–#1553
+
+**Status:** WORKING-MODEL (research document; no code changed).
+
+`.claude/research/2026-10-recovery/LEDGER-REVISION-2026-10-14.md` revises
+`CONSOLIDATED-LEDGER.md` after the 49 PRs merged since #1494. New closed ends C15–C18, C20, C21
+(O25 already pinned on 2026-10-08; D2's refusal core shipped as
+`nal::Revision` in #1538; the Staunen-table finding is answered by its own
+append-only addendum). New open items N1–N14,
+merged dots K1–K5.
+
+The sharpest: `syllogize` writes bits 50..52 as `0b111` and 53..63 as 0
+(N1), so the upper-half reading question (HL-O1/O2/O3, O4, O10, O14, the
+pyramid's promotion) becomes one decision, K1: declare bits 40..63 per class
+through the ClassView (plateau U1).
+
+OPEN: K1, the O25 policy (refuse or choice), O7, O23, O26 — operator decisions.
diff --git a/.claude/board/entries/README.md b/.claude/board/entries/README.md
@@ -25,11 +25,12 @@ 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.
 
-330 entries, 2026-08-06 .. 2026-10-14.
+331 entries, 2026-08-06 .. 2026-10-14.
 
 | date | entry id | finding | file |
 |---|---|---|---|
 | 2026-10-14 | `revision-layer-invariants` |  | [2026-10-14-revision-layer-invariants.md](2026-10-14-revision-layer-invariants.md) |
+| 2026-10-14 | `ledger-revision-after-1553` |  | [2026-10-14-ledger-revision-after-1553.md](2026-10-14-ledger-revision-after-1553.md) |
 | 2026-10-14 | `D-NALP-4` |  | [2026-10-14-isa-truth-exact-codes.md](2026-10-14-isa-truth-exact-codes.md) |
 | 2026-10-14 | `D-HOLO-H6b` | normal forms on the Context × Operation tile (measured) | [2026-10-14-d-holo-h6b.md](2026-10-14-d-holo-h6b.md) |
 | 2026-10-14 | `D-HOLO` |  | [2026-10-14-d-holo-falsifiers.md](2026-10-14-d-holo-falsifiers.md) |

diff --git a/.claude/research/2026-10-recovery/LEDGER-REVISION-2026-10-14.md b/.claude/research/2026-10-recovery/LEDGER-REVISION-2026-10-14.md
@@ -0,0 +1,220 @@
+# Ledger revision — after #1495 … #1553 (2026-10-14)
+
+Revises `CONSOLIDATED-LEDGER.md` (2026-10-11) in place of a new list. Every
+O-id keeps its number; new items get N-ids, merged dots get K-ids. The
+hypothesis-layer plateau (`ce64-reasoning-substrate-plateau.md` §4) has its
+own O1..O8, which collide with this ledger's ids; this file calls them
+**HL-O1..HL-O8**.
+
+Method: lance-graph, OGAR and ndarray fast-forwarded to `main` (49 lance-graph
+PRs merged after #1494, OGAR #357, ndarray +16; nothing rewritten). Read in full: PR arc
+#1505–#1552, `revision-layer-invariants`, `isa-truth-exact-codes`,
+`ce64-hypothesis-layers-plateau`, plateau §4/§5/§8, `nal-polyfill-v1` §D-NALP-3,
+`inverted-temporal-ablation` entry and plan §0–§1. Haiku harvest (search and
+maps) located candidates; every claim below that relies on a hit was read at
+the cited line.
+
+## 1. What merged since #1494, grouped by the thread it touches
+
+| PRs | what | touches |
+|---|---|---|
+| #1495, #1496 | NAL-1..9 × 34-recipe audit with OpenNARS/ONA oracles; CE64 hypothesis-layer plateau and plan (D-HL-0..10, D-NALP-0..11) | O35 (schedule depth measured), K1, K2 |
+| #1525, #1538 | `lance-graph-nal`: NAL queries as NodeGuid apertures; **revision as one grouped fold keyed by evidential base** | **K2/D2 core landed**, O8 (C3 arm), O25 |
+| #1548 | exact `u8` truth codes for deduction / induction / abduction; Gadamer gate invariants G1–G7; Rubikon R1–R4 specs | O22, O23, O26, O33, **N1** |
+| #1546 | `lance-graph-relalg`: relational IR normalized to a fixpoint before it runs | O32 (closure home), D-RA-4 |
+| #1550, #1551, #1553 | D-HOLO: address-indexed algebra falsifiers (H6 interval closure holds) | O28 |
+| #1549 | D-ORACLE-0 oracle fidelity matrix | N13 |
+| #1497–#1500, #1543, #1545 | research: PLN×Shannon, Hyperon/MeTTa, NARS self-evidence, Rubicon volition, flow/cog, semantic Morton | O33, K4 |
+| #1509, #1523, #1504–#1520, #1524, #1532 | SAG; class-method layer built then moved to backup (D-CMR reset) | D-SAG-3 builds on the pyramid (§5 K5) |
+| #1505, #1526, #1529–#1531 | native target-cpu default, debug 0, production v4 tiers | build only |
+| #1534, #1537, #1540, #1552 | r2il workspace member; core depends on the fold path; supervisor backoff; Haiku cards | infrastructure |
+
+## 2. Closed ends
+
+Kept from 2026-10-11: C1–C14 unchanged. New:
+
+| id | item | how it closed | evidence |
+|---|---|---|---|
+| C15 | O25 "pin first" | already pinned before the ledger was written | `planner/tests/chain_confidence.rs:88-100` (`self_revision_raises_confidence_on_the_replay_path`, FAILS IF `replay_step` gains a stamp guard), commit `1bcb15b6`, 2026-10-08 |
+| C16 | D2 core: a revision that refuses dependent evidence | `nal::Revision` classes each piece as independent / duplicate (counted once) / overlapping / no stamp; refuses the last two | `nal-polyfill-v1.md` §D-NALP-3; `lance-graph-nal/src/lib.rs:292-404`; 7 disable runs red |
+| C17 | exact NARS truth codes (part of the quantiser debt in coresearch G2 / plateau U8; not this ledger's O7) | deduction / induction / abduction exact over all 2³² inputs | #1548 entry, `truth_exact_probe` |
+| C18 | Gadamer invariants as contract | G1–G7 TEST-PINNED, G6 makes `is_acceptable()` the commit gate | `revision-layer-invariants.md` §1 |
+| C20 | CodeRabbit "Staunen in the codebook table" | no change: board entries are append-only; the same-day addendum records the rename to Astonishment | `2026-10-11-maturation3-ladder-probe.md:76-80` |
+| C21 | O9 resident-cycle operand design | unchanged, stays settled | — |
+
+## 3. Excluded
+
+X1–X7 unchanged. New:
+
+| id | item | why |
+|---|---|---|
+| X8 | NAL as a core module (D-FOLD-CORE-c) | dropped in #1537; `lance-graph-nal` is the authoritative NAL crate |
+| X9 | a second substrate for the exact truth law | #1548 DECISION: it lives in `causal_edge::isa::truth` |
+| X10 | ONA as a runtime | pinned oracle only (#1536) |
+
+## 4. Open ends — revised, deduplicated
+
+Status: OPEN / DECISION / MERGED-INTO. Fruit: L hours, H one PR with a probe,
+Y heavy. **Agent** = the cheapest tier that may do the grindwork (Haiku =
+harvest-scout / quote-harvester / census-counter / substrate-mapper /
+spec-editor / guarded-executor under orchestrator supervision; S = Sonnet
+worker; O = Opus main thread).
+
+### Thread 1 — audit and hygiene
+
+| id | item | status now | fruit | agent |
+|---|---|---|---|---|
+| O1 | board supersession of tenants 18–20 claims, #1427 label | OPEN, unchanged | L | Haiku spec-editor on an O-written spec |
+| O2a | Moore tenant hygiene (`project_state5`, `tenant_update`, Energy4 name) | OPEN, unchanged (0 doc hits since) | L | S |
+| O3 | `RelativeOffset` in `RegisterLaw` | OPEN, still at `register128.rs:228`, `hotplug.rs:274` | H | O decides, S edits |
+| O7 | model-identifier trailer on commits | DECISION, unchanged: OGAR, tesseract-rs and stockfish-rs forbid model identifiers in committed artifacts; lance-graph's CLAUDE.md does not, and `LATEST_STATE.md:1748` leaves the workspace-wide policy to the operator. This session omits the trailer, which is a practice, not a rule | — | O |
+| O4 | CE64 replay compatibility | DECISION, now part of **K1** (one version gate for every bits-40..63 reading change) | Y | O |
+| O5→O6 | `lab.rs` into jc, then its five defects | OPEN, unchanged | H | S + guarded-executor |
+| N3 | PR arc has no entries for #1494–#1504 (11 merged PRs) | OPEN | L | Haiku quote-harvester (PR bodies) → O writes |
+| N4 | STATUS_BOARD has no rows for D-RESIDUE-FP-0, D-MATURATION3-0 | OPEN | L | Haiku spec-editor |
+| N10 | `causal-edge` fails `clippy -D warnings` on six pre-existing lints; CI runs only `cargo test` there | OPEN | L | S, guarded-executor gate |
+
+### Thread 2 — learning seam and evidence pooling
+
+| id | item | status now | fruit | agent |
+|---|---|---|---|---|
+| O8 | F.2: HashMap learner vs quack `GroupAddr::Pair` + SumI32 | OPEN, **narrowed**: arm C3 (echo with and without dedup) is now answerable by `nal::Revision`'s duplicate class instead of a new dedup | L | S |
+| O25 | `replay_step` revises every hop with no evidential base | pin DONE (C15); **fix OPEN**, with two shipped patterns to reuse: refuse (NAL) or choice (`BeliefArena::revise_at`) | H | O picks refuse vs choice, S implements, the pinned test flips deliberately |
+| N6 | exact revision / synthesis: both-at-255 packs `(0, 0)` (pinned known defect) | OPEN (#1548) | H | O |
+| N8 | `nal::Choice` returns the maximum confidence, not the belief (no argmax fold) | OPEN (#1525) | H | S |
+| N7 | R2IL `RANGE` takes `u8` immediates: no range past row 255 | OPEN, needs an `ogar-r2il` change | Y (cross-repo) | O |
+
+### Thread 3 — resident Moore × CE64 cycle, and the upper half of the word
+
+| id | item | status now | fruit | agent |
+|---|---|---|---|---|
+| O10 | op read from bits 46..49 vs op from the recipe | MERGED-INTO **K1** (= HL-O1) | — | — |
+| O11 | a `SlabReading` variant for Moore lanes | OPEN: still four variants (`hotplug.rs:231-258`) | H | O designs, S edits |
+| O12 | the resident-cycle seam test | OPEN, can start test-local | H | S |
+| O13 | phase / activation rule | DECISION, unchanged | — | O |
+| O14 | freeze-only plasticity | MERGED-INTO **K1** (= HL-O2 + the pyramid's migration + N1) | — | — |
+| O17, O18, O19 | Experiment B, diagonal ownership, Experiment E | OPEN, after O12 | Y/L/H | — |
+| N1 | `syllogize` writes bits 50..52 as `0b111` and 53..63 as 0; `execute` writes W and 59..63 as 0. A derived conclusion inherits none of its premises' hydration, witness or maturation; under the pyramid reading every syllogism would read as Astonishment | OPEN (#1548 entry, `isa::contracts::{FORWARD, SYLLOGIZE}`) | part of K1 | O |
+
+### Thread 4 — event detection
+
+| id | item | status now | fruit | agent |
+|---|---|---|---|---|
+| O15 | Experiment A (detection quality) | DECISION, unchanged | H | — |
+| O20 | Cascade Stroke 1 drops unresolved candidates | OPEN in this ledger; **verify** against ndarray `main` (a Stroke-1 change was worked on in an earlier session) | L to verify | Haiku harvest-scout over ndarray `src/hpc/cascade*` |
+| O21 | `meta_prefilter` returns row indices, not a mask | OPEN (still 2 code hits) | L | S |
+
+### Thread 5 — reasoning: hydration, ablation, Gadamer, intention
+
+| id | item | status now | fruit | agent |
+|---|---|---|---|---|
+| O22 | nothing computes `CounterfactualVerdict` | OPEN, verified: only `revision.rs` and `kanban.rs` tests construct `Necessary`/`Dispensable`. Now gates G6 *and* D-SAG-CF-1 W6 | L (one-claim producer) | S |
+| O23 | TruthOnly → verdict | DECISION, **now with a recommendation from two independent sources**: D-SAG-CF-1 F8 requires TruthOnly to stay distinguishable from Inert and Necessary → a fourth reading, not a remap | — | O |
+| O26 | Intentional → (Do, Causes) | OPEN; the "pin" at `causal_bridge.rs:450` asserts the defect **as correct**, so a fix must flip a passing test. Target now has a home: Rubicon (an intention is a commitment phase; R3 "a goal is never revised by counting") | L (re-pin as documented defect) / DECISION (target) | Haiku spec-editor for the re-pin text; O for the target |
+| O27 | `first_divergence` unused in production; no cone | OPEN | H | S |
+| O28 | 64 cut scenarios as planes need a population-cache descriptor | OPEN, **new evidence**: D-HOLO H6 shows symbolic interval closure is the cheapest arm in all 24 cells; a cut set that is an interval needs no plane at all | H | O |
+| O29, O30, O31 | horizon identity revision, horizon persistence, eligibility masks | OPEN, unchanged | Y/H/H | — |
+| O32 | deductive closure over hydrated relations | OPEN, **candidate home**: `relalg::normalize` fixpoint + D-RA-4 bounded recursion | Y | O |
+| O33, O34 | teleology evidence classes; `learning::scm` unconnected | OPEN; O33 joins K4 | Y | O |
+| O35 | recipe schedule depth 1 | MEASURED in #1495 (`knowledge/nal-recipe-audit/schedule_depth.md`); no action proposed | — | — |
+| N9 | Rubikon R1–R4 tests (goal revision refused in `CognitiveWork`; agreement as `HorizonFusion`) | OPEN (#1548 entry §3) | L for R1 | S |
+| N13 | deepnsm-v2 `answered_questions` returns a Pearl `CausalMask` for syntactic S/P/O completion | OPEN, marked BLOCKING in `d-oracle-fidelity-matrix-v1.md:68` | L | S |
+
+### Thread 6 — the pyramid, residue and frontier (this session's)
+
+| id | item | status now | fruit | agent |
+|---|---|---|---|---|
+| N11 | promote Maturation3 beyond test-local | DECISION = **K1**; D-SAG-3 asks the same question (`sag-wisdom-frontier-ona-hyperon-v1.md:180`) | Y | O |
+| N12 | production feedback of resolved relations into `hydrate` | OPEN, needs K2's root id | H after K2 | O |
+| N14 | Haiku cards registered only for sessions started after #1552; a broad 24-pattern brief returned counts without hits | process note | — | brief one question per card run |
+
+## 5. The dots, re-drawn
+
+**K1 — one decision about the upper half of CE64 replaces five.** HL-O1 /
+O10 (46..49: opcode or activation), HL-O2 / O14 / N11 (50..52: plasticity,
+Staunen or maturation), HL-O3 (53..58: W), N1 (derived edges zero their
+witness and epistemic bits and set plasticity to `0b111`) and O4 (replay of
+words written under an older reading) are one question: *who declares what
+bits 40..63 mean for a class?* The plateau's U1 answers it the same way V3
+facets and D-GSO-7a already work: the ClassView declares a reading; the ISA
+reading becomes `isa-v2`; a pyramid reading is a second declaration; an
+undeclared read is refused; replay goes through a version gate
+(`I-LEGACY-API-FEATURE-GATED`). The test-local `M3Reading::Experimental` in
+`maturation_ladder.rs` is that mechanism in miniature: it refuses every
+existing `EdgeProvenance`. Why now: N1 makes the cost of not deciding
+concrete — the moment any class declares the pyramid reading, every
+`syllogize` output is born as Astonishment.
+
+**K2 — the provenance identity exists four times; the work is to make it one.**
+`nars::belief::Stamp(u64)` (disjoint → revise, overlap → choice), `nal`
+`Evidence { stamp: &[u64] }` (refuse overlap and no-stamp), contract
+`revision`'s `independent_roots` / `inherited_roots` masks (Gadamer), and the
+maturation probe's `roots: u64` + derivation lineage all answer "which
+independent sources does this rest on". What is still missing is the same
+in all of them: `replay_step` uses none (O25), a CE64 derived edge carries
+none (N1), and nothing maps a root to the split-tunnel log record (the
+original D2). D-HL-3 is the probe that decides the record shape (U6: the
+stamp lives in the delta record, not in the CE64 word). Fixing O25 with the
+NAL policy is the smallest step that makes two of the four agree.
+
+**K3 — the ablation verdict is one producer away.** G6 makes
+`is_acceptable()` the commit gate; nothing computes its input (O22). Pearl
+reactions, `first_divergence` (O27) and the pyramid's ablation impact are
+three readouts of the same cut. One function from (cut, two traces) to
+`CounterfactualVerdict` closes O22, gives D-SAG-CF-1 W6 its admission
+policy, and needs O23 decided first only for the TruthOnly case.
+
+**K4 — intention is not cause, and now it has a place to go.** O26
+(`Intentional → Causes`), O33 (no teleology evidence class), D6 (purpose is
+never derived), the Rubikon spec R3 (a goal is never revised by counting) and
+`rubicon-volition-v1.md:87` (revising a belief can leave a commitment active)
+point the same way: an intention maps to a commitment in the kanban phase
+machine, with its own evidence class, never to an `EdgeType::Do / Causes`.
+
+**K5 — residue, frontier and calibration are one loop.** D-RESIDUE-FP-0
+proves "computation alone cannot lower the residue". D-SAG-3 asks whether a
+stage-0 anchor exposes a frontier of consequences, and D-SAG-3.1 asks for
+incremental truth maintenance checked against a full-recompute oracle after
+every delta — the fixed-point test generalized to deletions. D-HL-4 (Narsese
+run in parallel) is the external oracle for the same loop. All three already
+keep the fences the pyramid set (no production 50..52, no 59..63 as desire,
+no loop as independent support).
+
+## 6. Fruit, in order
+
+**Low (hours, no operator):**
+1. N3 + N4 board hygiene — Haiku quote-harvester pulls #1494–#1504 PR
+   bodies, the orchestrator writes the arc entries, spec-editor adds the two
+   STATUS rows.
+2. O22-min — one-claim producer: LoadBearing → Necessary, Inert →
+   Dispensable, TruthOnly → NotRun; first end-to-end `is_acceptable()`.
+3. O26 re-pin — rewrite the `causal_bridge.rs:450` assertion as a documented
+   defect (FAILS IF the mapping is fixed) without changing behaviour.
+4. N9 R1 — a goal revision attempted in `CognitiveWork` is refused.
+5. N13, N10, O21 — each a small Sonnet edit behind a guarded-executor gate.
+6. O20 verify — Haiku harvest over ndarray cascade.
+
+**High (one PR with a probe, or one answer):**
+7. O25 fix (K2 step 1) — `replay_step` refuses overlapping evidence the way
+   `nal::Revision` does; the pinned test flips with a recorded reason.
+8. D-HL-3 — per-delta evidential base on D-RPF-P5's duplicate fixture.
+9. O27 + K3 — divergence cone feeding the verdict producer.
+10. N6, N8 — exact revision at 255, argmax fold for `Choice`.
+11. O11 → O12 — Moore `SlabReading`, then the resident-cycle test.
+
+**Heavy (design, cross-repo or open science):**
+12. K1 — the per-class reading of bits 40..63 (absorbs O4, O10, O14, N1, N11).
+13. N7 — widen R2IL `RANGE` past 255 in OGAR.
+14. O32 on relalg, O29, O33/O34 (K4), D-HOLO H11/H12.
+
+## 7. Decisions only the operator can make
+
+| id | question | suggestion |
+|---|---|---|
+| K1 | Declare bits 40..63 per class through the ClassView (U1)? | yes; it turns five conflicts into declarations and keeps `learn` correct under `isa-v2` |
+| O25 policy | On overlapping evidence in replay: refuse (NAL) or choice (`BeliefArena`)? | refuse in replay (a replay that silently picks hides the overlap) |
+| O23 | TruthOnly: fourth reading or remap? | fourth reading (D-SAG-CF-1 F8 requires it distinguishable) |
+| O26 | What does "A intends B" map to? | a Rubikon commitment with its own evidence class, never Causes |
+| O13 | phase / activation rule | unchanged; needs your rule |
+| O15 | go for Experiment A? | unchanged |
+| O7 | Is "no model identifier in committed artifacts" workspace-wide (then lance-graph's CLAUDE.md should say so), or per repo? | make it workspace-wide and write it into lance-graph's CLAUDE.md, so commits and harness trailers stop disagreeing |