From 034e3ea452933faa8e043349cf507c6f028539c2 Mon Sep 17 00:00:00 2001 From: Claude Date: Sat, 29 Aug 2026 18:50:20 +0000 Subject: [PATCH] board: arc entries for #1074 + #1075, batched Discharges the merged-PR obligation for both: PR_ARC_INVENTORY prepends and the LATEST_STATE table row. Batched deliberately -- the entries were owed by each PR's own commits (my miss), so they are retroactive either way, and one combined entry beats two separate retroactive cleanups. Per the termination clause this PR is hygiene-only and generates NO further obligations; the chain stops here. Records what a future session needs and cannot re-derive: - ISS-KANBAN-PLAN-EXIT-HAS-NO-NAMED-ROUTE is CLOSED (Plan was declared with semantics and reachable by no named primitive). - Two items OPEN ON MAIN, flagged not fixed: F-MEP-0b still reads "covariance or precision" though jc owns the covariance semantic (the real open question is the bounded monotone trust->step-transform mapping), and the Sigma->TrustTexture->GateDecision path in 0b/4 conflicts with EWA scoped as attention geometry. Both misdirect W1 before it runs. - The two prior-art corrections, both against my own earlier claims: Stamp::disjoint is canonical for S4 pooling under KNOWN overlap while the #854 ruling is canonical for independence NOT being established; and stance_panel is four non-destructive READINGS, not H3. - The next two plans and their gating, incl. the H3 wording correction: parallax earns the FIRST buyer for H3, it does not define H3. Supersession index regenerated. --- .claude/board/LATEST_STATE.md | 42 ++++++++++++++ .claude/board/PR_ARC_INVENTORY.md | 93 +++++++++++++++++++++++++++++++ 2 files changed, 135 insertions(+) diff --git a/.claude/board/LATEST_STATE.md b/.claude/board/LATEST_STATE.md index 733f0d849..75998cd44 100644 --- a/.claude/board/LATEST_STATE.md +++ b/.claude/board/LATEST_STATE.md @@ -14,6 +14,48 @@ untouched until that carve actually lands (it re-bases to 17 then). `TokenId` vs `WordId` control) → D-TVT-2 (the carve, A-or-B decided by D-TVT-1) → D-TVT-3 (lens write onto SPO-stream rows) → D-TVT-4 (BUY / NO-BUY). +## 2026-08-29 — #1074 + #1075 BOTH MERGED (plan, then the code split out of it) + +| PR | merge | content | +|---|---|---| +| **#1074** | `b871d4ea` | `mul-ewa-trust-propagation-v1` — PLAN/BOARD ONLY as merged (recut before merge) | +| **#1075** | `69016d99` | the epistemic triptych — `revision.rs`, `fusion.rs`, `KanbanColumn::{revise, advance_on_revision}` | + +**Contract inventory net delta (now on main):** `revision` and `fusion` +modules + two `KanbanColumn` primitives. Full type list in the #1075 arc entry. + +**`ISS-KANBAN-PLAN-EXIT-HAS-NO-NAMED-ROUTE` — CLOSED** by +`KanbanColumn::revise`. The `Plan` exit was declared with semantics and +reachable by no named primitive. + +**Open on main, deliberately flagged rather than silently fixed:** + +1. `F-MEP-0b` still reads "covariance *or* precision". `jc::ewa_sandwich_3d` is + explicit — *"world-space covariance matrices Σ ∈ ℝ³ˣ³ … pushed forward to + image-space"*, `M_k = sqrt(Σ_k)`. Precision would be a NEW quantity + borrowing the algebra, not reuse of the EWA operator. The open question is + the **bounded monotone trust→step-transform mapping**, not the kind. +2. The Σ→`TrustTexture`→`GateDecision` production path in §0b/§4 conflicts with + EWA scoped as attention/tension geometry. + +Both misdirect W1 before it runs; neither is a defect in the merged text's own +terms. **Follow-up PR, not a session note.** + +**Standing corrections carried from #1075 (both against earlier claims of +mine):** `Stamp::disjoint` is canonical for S4 pooling under KNOWN overlap, and +the **PR #854 ruling** is canonical for evidential independence NOT being +established (`event identity ≠ evidential-base membership ≠ source +dependence`). And `stance_panel` supplies four non-destructive READINGS, not +the H₃ synthesis — the Grail is unbuilt. + +**Next, per the operator's audit:** `perspective-parallax-v1` (bounded Shannon +control → `PerspectiveResidual` sibling differential → Three-Mountains rotation +→ Perspective/Meta band receipts; H₃ deferred — parallax earns the FIRST buyer, +it does not define H₃) and `mul-proprioception-v1` (MUL as a proprioceptive +revision loop mapping to existing `FieldModulation`; ambiguity widens cognition, +never Blocks; the Rubicon gate stays separate). + + ## 2026-08-29 — the epistemic triptych lands in the contract (BRANCH, not yet merged) > **✅ SCOPE RESOLVED (operator-instructed, 2026-08-29).** These commits were diff --git a/.claude/board/PR_ARC_INVENTORY.md b/.claude/board/PR_ARC_INVENTORY.md index 0c1b83a73..50abcf485 100644 --- a/.claude/board/PR_ARC_INVENTORY.md +++ b/.claude/board/PR_ARC_INVENTORY.md @@ -1,3 +1,96 @@ +## 2026-08-29 — lance-graph #1075 (MERGED 69016d99) — the epistemic triptych lands in the contract + +- **Added:** `contract/src/revision.rs` (`EvidenceMask`, `InterpretiveHorizon`, + `EncounterEvidence`, `BasisView`, `RevisionKind`×9, `EvidentialEffect`×3, + `RevisionDelta`, `RevisionPolicy`, `GadamerRevision`, `RevisionOutcome`, + `CounterfactualVerdict`×3, `RevisionVerdict`); `contract/src/fusion.rs` + (`FusionOutcome`×7, `SynthesizedClaim`, `FusionReceipt`, `fuse()`); + `KanbanColumn::{revise, advance_on_revision}`. Zero-dep throughout. +- **Why:** PR #1057 (MERGED) names revision *"the only write-back"* and *"the + court of appeal"* at its ACCEPT step (D-ECG-3/6), while `counterfactual.rs` + had shipped **twice** and `revision.rs` never landed — a merged plan's ACCEPT + step governed by a module that did not exist. Operator: *"the holy Grail + would be Gadamer Horizontverschmelzung of Thesis and antithesis synthesis + inference."* +- **The triptych, kept violently separate (operator):** `fusion` GENERATES a + candidate · `counterfactual` ATTACKS it · `revision` LICENSES it. `fuse()` + establishes synthesis **admissibility and provenance only** — + `SynthesizedClaim` carries no inferred relation `X`, deliberately, so the + module cannot both propose and license. +- **Mechanisms:** *an assumption IS an inherited root* (revisable iff inherited + and not independently grounded) makes `IrreducibleTension` checkable rather + than stylistic — **a fake intelligence always synthesises**. The anti-alchemy + law reads DISJOINT roots, never the union, so shared ancestry cannot mint + weight: *understanding may increase without evidence increasing*. +- **`ISS-KANBAN-PLAN-EXIT-HAS-NO-NAMED-ROUTE` CLOSED.** `Evaluation`'s + successors are `[Commit, Plan, Prune]`; `advance()` took the first non-`Prune` + (always `Commit`), `veto()` took `Prune`, and **nothing produced `Plan`** — + the documented revision exit had legal-edge status and no route. `revise()` + is the route; no edge added, no `next_phases` change. +- **Typed docket (operator correction):** `IncreaseEligible` means only *a new + independent root was introduced* — not that the counterfactual passed. It was + emittable from an `EncounterEvidence` **alone**, so routing it to `Commit` + bypassed Fusion→Counterfactual→Revision. `RevisionVerdict` + + `CounterfactualVerdict` close it; `NoIncrease → Plan` makes the closed-path + guarantee **motion** rather than prose. +- **Two prior-art corrections, both against my own earlier claims:** + `Stamp::disjoint` is **sound, not complete** — `Stamp` models source + MEMBERSHIP, is lossy (`1 << (id % 64)`), and `causal_audit.rs:346` leaves + `independent_strength` `None` for want of a dependence model (PR #854), so + `FusionReceipt::shared_roots` is CONJECTURE, not duplication. And + **`stance_panel` is NOT H₃** — four non-destructive READINGS of an existing + state, not the construction of a larger horizon. **The Grail survives.** +- **Falsifiers:** five, every one disable-verified and surgical (each disable + fails exactly its own test). `revision_verdict_alone_cannot_reach_commit` + PROVES the shortcut rather than asserting it. +- **Docs:** module docs; `LATEST_STATE` inventory + five-loci finding; this entry. +- **Confidence:** high on the code (1247 contract tests, 8/8 CI, clippy + `-D warnings`); the `shared_roots` contract is explicitly CONJECTURE until an + evidence-event identity exists. + +## 2026-08-29 — lance-graph #1074 (MERGED b871d4ea) — mul-ewa-trust-propagation-v1, measure-before-carve + +- **Added:** `.claude/plans/mul-ewa-trust-propagation-v1.md` (PLAN/BOARD ONLY + as merged) + INTEGRATION_PLANS prepend + STATUS_BOARD `D-MEP-0..4` + five + `ISSUES` entries. Supersession index regenerated in-commit. +- **Why:** MUL answers *"how much do I trust HERE"* (scalar `TrustQualia`); + nothing answers *"what does that trust become N hops away."* jc's EWA + sandwich is the CANDIDATE operator — never certified for the semantic claim. +- **The two hinges (§0b, operator):** (1) the MUL revamp — Dunning-Kruger + overconfidence vs trusted epistemic knowledge vs counterfactual, as + thesis/antithesis/synthesis, asking how grounded the known and **indirect + intermediate** unknowns are; (2) the EWA sandwich borrowed from 3DGS — + filling a known unknown by *Oberflächenspannung* vs HHTL parent inheritance + vs thinking-style reasoning. **This plan measures only EWA against a scalar + baseline**, so a BUY licenses "EWA beats naive decay", NOT "EWA is the right + fill". +- **The danger, named:** an epistemic frontier handed to math goes circular or + becomes accidental entropy-based intelligence. `F-MEP-0a` (§0a) is the + capability gate — redistribute/rank **yes**; eliminate worlds **no**; mint + evidence **never** — above the sharpened invariant: *no operation whose + inputs are entirely graph-internal may convert a hypothesis into epistemic + evidence merely by reducing its uncertainty.* +- **Three on-paper defects caught before any run:** the suspicion score was + **inverted** (`M_k = √trust·I` SHRINKS covariance as trust falls, so a + distrusted 5-hop chain scored `Calibrated`); the readout is **hop-count + dominated** (`∏t_k ≈ t̄ⁿ`); circularity was unnamed. `F-MEP-0b`/`0c`/`0d` + gate all three, `0d` on synthetic closed graphs before W1's cohort. +- **Statistical pre-registration:** cohort, one normalized readout + (`trace(Σₙ)/trace(Σ₀)`), per-half floors, leakage-safe `blake3` split, the + chain-level observation unit, a 1000-redeal one-sided null, and a three-part + BUY (`AUC(b) > 0.5` **and** ΔAUC ≥ 0.05 **and** 2σ). Five review rounds; the + `AUC > 0.5` condition exists because ΔAUC alone could buy a *bigger error*. +- **Scope correction, in-arc:** contract code was landed on this branch against + its own "no contract change" status line; split to #1075 on operator + instruction and this PR recut to plan-only (0 `crates/` files) before merge. +- **Open on main, flagged not fixed:** `F-MEP-0b` still reads "covariance *or* + precision" though jc owns the covariance semantic; and the + Σ→`TrustTexture`→`GateDecision` path in §0b/§4 is in tension with + EWA-as-attention-geometry. Both misdirect W1 and want a follow-up. +- **Docs:** the plan; this entry. +- **Confidence:** high that §1's facts are verified at HEAD file:line; the + transplant itself is deliberately UNBOUGHT until W1. + ## 2026-08-28 — lance-graph #1072 (MERGED 3a3e15b6) — token-value-tenant-v1 PROPOSED: a byte-exact span address inside the 40,767-triple stream - **Added:** `.claude/plans/token-value-tenant-v1.md` (PLAN/BOARD ONLY, no