|
| 1 | +## 2026-06-21 — E-EQUIVALENCE-IS-THE-CRUX — template-equivalence is the load-bearing verifier of the whole loop; it MUST fail closed, and it rides on transparent Lance versioning (surrealdb #50) |
| 2 | + |
| 3 | +**Status:** FINDING (cross-session feedback, 2026-06-21). Reframing that |
| 4 | +re-grades the comparison gate from footnote to crux. The Cognitive Compilation |
| 5 | +loop's "take the training wheels off" decision — let a deterministic Elixir |
| 6 | +template REPLACE the LLM run — is exactly the output of `template-equivalence`: |
| 7 | +*does the template reproduce the recorded LLM run's results?* Therefore: |
| 8 | + |
| 9 | +1. **The verifier is load-bearing, not incidental.** If the comparison can pass |
| 10 | + when it shouldn't, the loop **self-certifies on a lie** — the LLM is removed |
| 11 | + from the hot path on a false equivalence. A codex P2 on this crate (a stable |
| 12 | + top item passing `RankOrder` despite dropped provenance) is really the crux, |
| 13 | + not a minor note. |
| 14 | +2. **Fail closed.** A pass must be an affirmative PROOF of reproduction, never |
| 15 | + "no difference detected". Hardened accordingly (commit `…`): claims compared |
| 16 | + as a SET both ways (no additions AND no omissions), source spans compared |
| 17 | + exactly, ranked-item SET preserved for `RankOrder` (only order may perturb, |
| 18 | + within tolerance), and every unevaluable dimension (incl. the deferred |
| 19 | + Semantic class) → `Failure`. 8 tests. |
| 20 | +3. **It rides on transparent versioning (surrealdb #50).** The loop records each |
| 21 | + orchestration step as a versioned Lance commit and verifies by **AS-OF replay |
| 22 | + + compare**. That requires the corrected version→snapshot mapping landed in |
| 23 | + surrealdb #50; under the old broken `checkout_version(versionstamp)` the |
| 24 | + verifier would replay the WRONG step and compare garbage. #50 is the substrate |
| 25 | + this loop records and replays on — load-bearing for the verifier, not |
| 26 | + incidental. |
| 27 | + |
| 28 | +Architecture note (cross-session): rig + rs-graph-llm (a Rust LangGraph) run the |
| 29 | +LLM-angle orchestration onto surrealdb-on-kv-lance; each step is a Lance commit; |
| 30 | +the run is transcribed into an Elixir template and verified by replay-and-compare |
| 31 | +against the recorded run. The reflex binary (`crates/cognitive-stack`) stays |
| 32 | +LLM-free; the LLM lives in the learning loop only. |
| 33 | + |
| 34 | +Cross-ref: `crates/template-equivalence/src/lib.rs` (fail-closed `compare`); |
| 35 | +PR #571; surrealdb #50 (transparent versioning); E-ELIXIR-TEMPLATE-IS-THE-GAP. |
| 36 | + |
| 37 | +## 2026-06-21 — E-ELIXIR-TEMPLATE-IS-THE-GAP — the compiled-cognition loop already has every organ EXCEPT the Elixir-shaped template; that template is the only new build |
| 38 | + |
| 39 | +**Status:** FINDING (operator scope correction, 2026-06-21). The "Cognitive |
| 40 | +Compilation" idea (LLM = teacher/compiler/critic, Lance-Graph = reflex runtime) |
| 41 | +maps onto organs the workspace already grew: **ractor** (control-plane ownership |
| 42 | +fence), **surrealdb-as-kv-lance** (provenance/timeline view), **Rubicon / |
| 43 | +kanbanview** (phase choreography), **thinking styles + JITson + i4-32D |
| 44 | +thinking-style vectors** (style dispatch + compiled kernels). The one missing |
| 45 | +organ is the **Elixir-shaped template** — the declarative `pipeline do step :x end` |
| 46 | +a proven LLM run compiles down to, whose steps bind to OGAR actions and run |
| 47 | +deterministically. Consequence: the additive build is small and surgical |
| 48 | +(`elixir-template` representation + parser + the `source_ranking_v1` first slice, |
| 49 | +plus the runtime/equivalence/compiler surfaces around it), NOT a from-scratch |
| 50 | +stack. Anything that proposes re-building Rubicon, a new orchestration crate, a |
| 51 | +surrealdb engine change, or a thinking-style system is drift — those exist. |
| 52 | +Cross-ref: plan `.claude/plans/cognitive-compilation-v1.md`; OGAR `ogar-from-elixir` |
| 53 | +(the eventual richer Elixir→OGAR front-end); §18 gates (no trace→no template, |
| 54 | +no replay→no promotion, no LLM in hot path once a template passes). |
| 55 | + |
1 | 56 | ## 2026-06-21 — E-S6-SOA-IS-ONE-FIXEDSIZEBINARY-NO-SECOND-COPY — operator override CORRECTS `E-S6-SCAN-SOA-NOT-ON-SHARED-VAL`: the SoA is stored ONCE as `FixedSizeBinary(512)` (single source, zero-copy) in lance-graph's own dataset; surrealdb is the Rubicon VIEW; NO second copy/column, NO time-series drop via tombstone+purge, lance 7.0.0 MANDATORY (not "unverified") |
2 | 57 |
|
3 | 58 | **Status:** FINDING (operator directive, 2026-06-21 — three explicit NOs that supersede the prior council's suggestions). CORRECTS the same-day `E-S6-SCAN-SOA-NOT-ON-SHARED-VAL` (whose "separate `soa_val` column / per-cell copy-fallback floor / deferred behind an unverified lance baseline" framing is RETRACTED). |
|
0 commit comments