docs: v2.x specs for replay/derivation-worker/view-flip chain - #267
Conversation
Three cascade addenda to docs/design/write-log-as-truth.md, one per post-#205 chain issue: - docs/v2_replay.md (#262): shape-equality at ingest time, not bit-equality. Posterior drift is feedback doing its job, not drift; legacy-origin backfill bucketed separately. Drift exit threshold and example sample sizes settled. - docs/v2_derivation_worker.md (#264): single-transaction ingest path, idempotency check on (derived_belief_ids non-empty AND beliefs exist), recover-by-replay between batches, edges derived in same pass as beliefs. - docs/v2_view_flip.md (#265): WRITE_LOG_AUTHORITATIVE flag with two-release soak before default-on, aelf rebuild preserves alpha/beta and replays feedback by default, --rule-set and --as-of ship as stubs in v2.x, drift alarm wires replay_full_equality into aelf doctor. Each memo follows the cascade-addendum pattern: what's being decided, recommendation, decision asks, downstream impact, out-of-scope, provenance. Recommendations only; decisions are the maintainer's.
|
Review: discretion grep flags one phrase — |
|
Review: discretion leak — three docs added in this PR carry tier vocabulary that's reserved for internal coordination, not public artifacts.
Suggested fix: drop the section entirely, or rename to something neutral like Once sanitized, push the fixup and re-request review. Dropping |
robotrocketscience
added
review-Setr
|
Discretion review — three uses of
Tier vocabulary doesn't belong in public artifacts. Suggest reframing as "Why this is design-heavy" / "Risk surface" or similar — the substance of the sections is fine, just the label needs to come out. Dropping |
robotrocketscience
added
review-Kulili
|
Review: blocked on one fix. Discretion leak — Otherwise the cohort looks coherent: shape-equality contract in v2_replay matches worker invariants in v2_derivation_worker matches rebuild semantics in v2_view_flip; Push the one-word fix and history-scan should re-run clean. Dropping review label. |
robotrocketscience
added
review-Kulili
|
Review: blocked — needs rebase before this is reviewable. Branch last commit 2026-04-28T23:55, predates #253, #266, #259, #273, #276, #271, #277, #279 landing on
Action requested: rebase onto current Dropping |
robotrocketscience
added
review-Kulili
|
Blocker on merge: all three new memos ship tier-reasoning vocab in public-readable docs.
Per coordination protocol, "queen" / "queen-tier" / "rook" / "rook-tier" are banned in any third-party-readable artifact. The substance of the section (high-judgment scope, settle calls before implementing) is fine — just rename to e.g. Sanitize in place, force-push, ping. Dropping |
|
[claim:review:Kulili:2026-04-29T01:22:43Z] |
|
Re-review after rebase: rebase is clean (251 ins / 0 del, three new docs only) and CI is green, but the discretion leak flagged in prior reviews has not been addressed. Still blocked. Confirmed instances in current diff:
Per coordination protocol, Sanitize in place, force-push, re-request review. Dropping |
|
[release:review:Kulili:2026-04-29T01:23:30Z] |
Renames '## Why queen' to '## Why this is judgment-scope' and rephrases bodies in v2_replay.md, v2_derivation_worker.md, v2_view_flip.md to drop 'queen'/'queen-tier' references. Substance unchanged.
|
[claim:review:Kulili:2026-04-29T01:25:54Z] |
|
[release:review:Kulili:2026-04-29T01:26:12Z] |
## Summary Cascade addendum to `docs/bayesian_ranking.md`. Drafts the v2.0 spec memo for the residual scope of [#151](#151) — the work that remains after #146 shipped the core posterior-weighted ranking at v1.3. - New file: `docs/v2_posterior_ranking_residual.md` - Pattern matches `substrate_decision.md`, `v2_replay.md`, `v2_view_flip.md`, `v2_derivation_worker.md`. ## Recommendation at a glance | Slice | What | When | |---|---|---| | 1 | MRR uplift + ECE eval harness under `benchmarks/posterior_ranking/` | ships first; no #150 dependency | | 2 | Heat-kernel composition into the score equation | gated on #150 | | 3 | Per-corpus weight sweep (`aelf bench --sweep`) | last; needs slice 1's evaluator | Real-feedback retest is held out of v2.0 and filed as a follow-up — pending corpus size threshold (≥1000 events / ≥100 beliefs). ## Decision asks (in the memo) 1. Three-slice sequencing. 2. MRR uplift threshold `+0.05`. 3. ECE threshold `0.10` for synthetic. 4. Weight sweep range `[0.0, 1.0]` step `0.1`. 5. Real-feedback retest deferred outside v2.0. ## Cross-links - Closes (spec only) #151. Implementation tracked separately under the same issue. - Adjacent v2.x cascade specs: #267 (replay/derivation/view-flip). - v1.3 partial: `docs/bayesian_ranking.md` (already on main). ## Notes Docs-only. No code, no tests. Recommendation only — ratification is a separate sweep, same as the other cascade addenda.
1 participant
Summary
Three cascade addenda to
docs/design/write-log-as-truth.md, one per post-#205 chain issue. Drafts the formal spec memos called for by theneeds-speclabel on each.docs/v2_replay.md— issue [v2.x] Implement replay_full_equality — flip-readiness probe #262 (replay_full_equalityflip-readiness probe)docs/v2_derivation_worker.md— issue [v2.x] Derivation worker — beliefs become materialized state #264 (derivation worker)docs/v2_view_flip.md— issue [v2.x] View-flip — beliefs/edges become materialized; ingest_log is canonical #265 (view-flip — `ingest_log` becomes canonical)Pattern matches the substrate-cascade addenda merged via #257 / #259: each memo states what's being decided, recommendation, decision asks, downstream impact, out-of-scope, provenance. Recommendations only; ratification is a separate sweep.
Recommendations at a glance
Cross-links
Test plan