oleg.m/waggle-os
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
b20b138fe485817e8d69d0c60e376e3c53af149a
waggle-os/docs/decisions/2026-04-28-gepa-faza1-launch.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
Details
moving
2026-09-02 10:10:29 +02:00

30 KiB
Raw Blame History

decision_id, date, authority, session, mission, status, chain, manifest, substrate, cost_cap, expected_wall_clock, verdict
decision_id date authority session mission status chain manifest substrate cost_cap expected_wall_clock verdict
2026-04-28-gepa-faza1-launch 2026-04-28 PM (Marko Markovic) — RATIFIED via Amendment 1 CC-2 (filename retains cc4 historical naming; CC-2 is operational executor) GEPA Tier 2 Prompt-Shapes Evolution Faza 1 (proof-of-concept pilot) LOCKED upon authoring
briefs/2026-04-28-cc4-gepa-tier2-evolution-faza1-brief.md (PM brief, 266 lines)
briefs/2026-04-28-cc4-faza1-preflight-report.md (CC-2 pre-flight, 239 lines)
briefs/2026-04-28-cc4-faza1-amendment-1.md (PM Amendment 1, 252 lines)
briefs/2026-04-28-cc4-faza1-amendment-2.md (PM Amendment 2 — Phase 4.5 retrieval-engagement signal)
PM Amendment 3 (oral ratification embedded in CC-2 session 2026-04-28T00:45:00Z) — cost cap raise + Pre-Gen-1 re-projection rule
PM Amendment 4 (oral ratification embedded in CC-2 session 2026-04-28T01:00:00Z) — Option B retry of 3 failed cells via JSON-mode + binding texture-audit caveat
PM Amendment 5 (oral ratification at Checkpoint A 2026-04-28T11:00:00Z) — judge metric parallel-report (raw agreement primary) + F-saturated-baseline rule [PARTIALLY REVOKED by Amendment 6: F-saturated rule PAUSED pending real NULL data]
PM Amendment 6 (oral ratification post Checkpoint A bug discovery 2026-04-28T15:30:00Z) — NULL-baseline shape-override bug fix + re-run authorization + reversal of original Checkpoint A per-shape findings (artifactual)
decisions/2026-04-28-phase-4-3-rescore-delta-report.md (Phase 4.3 verdict — GEPA motivation; Amendment 6 REVOKED prior PM-clarification-note proposal)
decisions/2026-04-28-phase-4-5-tools-audit-results.md (Phase 4.5 — Amendment 2 trigger)
benchmarks/preregistration/manifest-v7-gepa-faza1.yaml (in waggle-os-faza1-wt; 5 SHA pins in §B for initial LOCK + Amendments 2/3/4/5) feature/c3-v3-wrapper @ c9bda3d (Phase 4.7) via isolated git worktree $115 hard / $90 internal halt / $109.08 expected (raised from $100/$80/$100.50 by Amendment 3 — inherited estimate correction, not scope creep) 3-5 days CC time, 2-3 days wall-clock (no rate-limit blockers) LOCKED (no further halts beyond 4 mandatory checkpoints unless sub-rule trigger fires)

Faza 1 Launch Decision — LOCK

This decision LOCKS Faza 1 scope, methodology, cost ceiling, and acceptance criteria. All Faza 1 work executes against this LOCK + manifest v7 audit chain. Deviation from §A inherited rules, §B audit chain, §C substrate anchor, §D cost discipline, §E checkpoint protocol, or §F acceptance criteria triggers immediate halt-and-PM per brief §10 deviation policy (inherited from manifest v6).

The contents of this decision are self-contained binding contract — Faza 1 audit references that would have cited an external feedback_config_inheritance_audit.md instead cite "Faza 1 launch decision §A" per Amendment 1 Ask E alternative resolution.

§A — Inherited Pre-flight Rules (binding for entire Faza 1)

The 8 sub-rules below are PM's brief §6.1-§6.8 verbatim. Each sub-rule has a binding-checklist line for CC-2 operational compliance. Sub-rule violation triggers halt-and-PM per brief §5 halt protocol.

A.1 — Config inheritance audit (brief §6.1 verbatim)

Eksplicitno specify Qwen + Opus + GPT + MiniMax model strings + reasoning_effort + max_tokens u manifest v7. Ne nasleduj iz manifest v6 implicitly. Naročito: Qwen reasoning mode (thinking vs non-thinking) MUST match per-shape configuration (qwen-thinking shape → reasoning enabled, qwen-non-thinking → disabled).

Binding compliance: Manifest v7 §judges + §subject blocks declare model strings + max_tokens + thinking flags explicitly with inherited_from: pilot_2026_04_26_runner_sha256_8a6251e2_line_626. Per-shape thinking flag enforcement is via existing prompt-shapes/{qwen-thinking,qwen-non-thinking}.ts baselines (verified at substrate anchor SHA per §B).

A.2 — Mixed-methodology baseline (brief §6.2 verbatim)

NULL-baseline mora prijavljivati trio-strict + self-judge razdvojeno (ne shared aggregate). Phase 4.3 koristi trio-strict; pisanje "GEPA delta" mora citirati trio-strict numbers, ne self-judge.

Binding compliance: All Faza 1 reports — Pre-A through Checkpoint C — report trio-strict + self-judge as separate columns. Acceptance §F applies to trio-strict only. Self-judge is supplementary diagnostic.

A.3 — Scope verification (brief §6.3 verbatim)

Pre run, CC-4 verifikuje da H3 cell ima ≥40 instanci u source corpus (potrebno za 8 NULL + 24 GEPA + 5 held-out = 37 instances + buffer). Ako H3 ima <40 instanci, halt-and-PM (signal da scope estimate pogrešan).

Binding compliance: Per Amendment 1 Ask A Option C, H3 source corpus = 50 net-new NorthLane CFO synthesis instances (corpus_design.total_instances=50 in manifest v7 §corpus_design). This satisfies ≥40 with comfortable margin (8 NULL + 24 Gen 1 + 5 held-out + 13 buffer). Pre-A halt-and-PM verifies corpus existence + spot-audit before NULL-baseline kick.

A.4 — Cell semantics prompt strictness preservation (brief §6.4 verbatim)

Audit step pre commit Faza 1 results: za svaki GEPA candidate prompt, diff vs baseline. Diff mora biti samo unutar prompt-shape template body (between defined boundaries u shape file). Diff koji touch-uje cell.system_prompt ili cell.scoring_rubric = automatic INVALID, candidate dropped, mutation oracle re-prompted.

Binding compliance: Manifest v7 §gepa.mutation_validator specifies allowed/invalid diff targets. Boundary anchor = MULTI_STEP_ACTION_CONTRACT constant in packages/agent/src/prompt-shapes/types.ts (byte-level SHA pinned in §B). Validator runs as automated check before any GEPA candidate enters evaluation queue. INVALID candidate triggers oracle re-prompt; 2 consecutive INVALID per shape triggers halt (brief §5).

A.5 — σ-aware acceptance range (brief §6.5 verbatim)

N=8 per cell daje cca CI ± 17pp at 95% (binomial), što je široko. +5pp acceptance threshold je ne-statistički-rigorozan na N=8 — uzima se kao fitness signal indicator, ne kao publishable claim. To je razlog zašto Faza 1 = proof-of-concept, ne paper-ready evidence. Faza 2 scale-up je tek tu za publishable σ-bounded delta.

Binding compliance: All Faza 1 reports include σ-aware acceptance disclaimer verbatim. Manifest v7 §faza_1_acceptance.condition_1_updated.signal_disclaimer encodes this. Public-facing language post-Faza-1 is "fitness signal indicator", never "statistically significant".

A.6 — Mixed-methodology variant (brief §6.6 verbatim)

Trio-strict je primary; self-judge je supplementary diagnostic only. Faza 1 acceptance rule (§4) bazira se na trio-strict, ne self-judge.

Binding compliance: §F acceptance criteria conditions all reference trio-strict. Self-judge appears only in supplementary diagnostic columns, never in PASS/FAIL determination.

A.7 — Cost super-linear input growth (brief §6.7 verbatim)

GEPA candidates have variable token length (mutations may grow prompts). Cost calculation must use worst-case 1.5× baseline token count per candidate (encodes mutation overhead). If actual mid-run cost exceeds projection by >30%, halt.

Binding compliance: Manifest v7 §cost_governance.super_linear_buffer encodes 1.5× projection multiplier + 30% mid-run halt threshold + every-20-evaluations audit frequency. Pre-A through Checkpoint C reports include cost-projection-vs-actual delta tracking.

A.8 — Source data structure (brief §6.8 verbatim)

Verify H3 source data is agentic knowledge work format (not factoid LoCoMo). Phase 4.3 categorization confirms H3 = agentic. CC-4 spot-check 3 random H3 instances pre run, confirm task structure matches pilot 2026-04-26 corpus.

Binding compliance: Pre-flight report §3.8 verified pilot artifact format = agentic synthesis (NorthLane CFO 6-document knowledge work, not LoCoMo factoid Q&A). Amendment 1 Ask A Option C corpus extends this format via task families F1-F5 (manifest v7 §corpus_design). Pre-A spot-audit (5 random instances) verifies new corpus matches the pattern. Note: Pre-A audit is 5 instances rather than 3 per brief §6.8 minimum — buffer for 50-instance scale.

A.9 — Phase 4.5 retrieval-engagement signal (Amendment 2 binding addition)

Pilot empirical signal: Qwen retrieves 1.33×/task vs Opus 2.33×/task on byte-identical MULTI_STEP_ACTION_CONTRACT surface (your 70a1701d... hash); H4 score gap mechanistically traces to under-engagement, NOT tool format. GEPA fitness for Qwen-targeted shapes weights retrieval-engagement bonus per Amendment 2 §3; mutation oracle for Qwen-shapes emphasizes anti-premature-finalization scaffolding per Amendment 2 §4.

Binding compliance:

  1. Per-shape fitness function fork (manifest v7 §metric_operationalization.per_shape_fitness_formula): Qwen-targeted shapes (qwen-thinking, qwen-non-thinking) compute fitness = trio_strict_pass_rate + retrieval_engagement_bonus − cost_penalty; non-Qwen shapes (claude, gpt, generic-simple) compute fitness = trio_strict_pass_rate − cost_penalty (no retrieval engagement weighting — these shapes don't have the gap).

  2. Retrieval engagement bonus bands (manifest v7 §metric_operationalization.retrieval_engagement_bonus.bands):

    • +0.05 (5pp) if mean retrieval_calls per task ≥ 2.0 (Opus parity proxy)
    • 0.00 if mean retrieval_calls per task in [1.5, 2.0)
    • −0.05 (5pp) if mean retrieval_calls per task < 1.5 (Qwen baseline behavior penalty)

  3. Mutation oracle fork (manifest v7 §mutation_oracle_design): two prompt template paths — mutation-prompt-template-qwen.md (anti-premature-finalization scaffolding) and mutation-prompt-template-non-qwen.md (standard guidance). Forking is by shape class string match (qwen-thinking, qwen-non-thinking → Qwen branch; claude, gpt, generic-simple → non-Qwen branch).

  4. Acceptance criteria update — see §F condition 1 (third update) + §F.5 (NEW FAIL).

  5. Telemetry source: retrieval_calls counter is existing agent harness telemetry (per pilot 2026-04-26 trace data) — no new API calls; fitness function reads existing telemetry.

  6. Test coverage requirements (manifest v7 §amendment_2_integration.scaffold_test_coverage_NEW_requirements): 5 retrieval_engagement boundary tests (1.49/1.50/1.99/2.00/2.50) + 5 shape-routing tests (claude/gpt/generic-simple excluded; qwen-thinking/qwen-non-thinking included) + §4.5 FAIL test + §4.5 PASS-path test. ≥80% coverage on per-shape fitness function module.

  7. Phase 5 forward record (NOT Faza 1): CC-2 must NOT optimize Faza 1 selection for Phase 5 GEPA-evolved variant criteria (engagement parity ≥ Opus + score parity narrowed by ≥0.30 H4 trio_mean delta). Faza 1 selection per §F only.

A.10 — Pre-phase-boundary cost re-projection (Amendment 3 binding rule)

After NULL-baseline run completes (Checkpoint A), BEFORE Gen 1 kick, CC-GEPA must:

  1. Compute actual cost-per-evaluation from NULL-baseline telemetry (5 shapes × 8 instances × actual subject + judge cost)
  2. Project Gen 1 cost = 5 shapes × 3 candidates × 8 instances × actual per-eval cost
  3. If projected Gen 1 > $78 (30% over $60 manifest projection), halt-and-PM with options:
    • (a) raise Gen 1 cap proportionally (Amendment 3-style correction for inherited estimate)
    • (b) reduce Gen 1 scope (3 candidates × 6 instances OR 2 candidates × 8 instances)
    • (c) pause Faza 1 + PM decides path

Binding compliance: Manifest v7 §cost_governance.pre_phase_boundary_reprojection encodes the rule. Phase-boundary re-projection catches projection errors that continuous monitoring misses — re-baselines against fresh telemetry rather than original projection. This rule is codified in response to the corpus-generation cost surprise (Amendment 3) where inherited $0.10/instance estimate was 170% off vs actual $0.27/instance Opus 4.7 cost.

The rule applies to Pre-Gen-1 boundary as the canonical instance. PM may extend to other phase boundaries via future amendment.

A.14 — NULL-baseline shape-override bug fix (Amendment 6 binding)

Bug discovered post-Checkpoint-A: original run-null-baseline.ts:runOneEval(shape, ...) received PromptShape but did NOT forward it to runRetrievalAgentLoop as promptShapeOverride. All 40 evals used the model-alias-default shape (qwen-thinking for Qwen subject). The "per-shape pass rates" were 5×8 replicates of qwen-thinking, NOT shape-vs-shape comparison.

Fix: added promptShapeOverride: shape.name to runRetrievalAgentLoop call. Regression test at benchmarks/gepa/tests/faza-1/null-baseline-shape-override.test.ts (4 tests, all passing) verifies source-text invariant.

Re-run: NULL-baseline rerun with the fix; 5 sunk artifacts preserved as *-artifactual-bug-superseded.{ext} for audit trail; new artifacts written to canonical paths.

Reversals:

  • Amendment 5 §F-saturated-baseline-rule PAUSED until real per-shape NULL data confirms qwen-thinking ≥ 0.88 (saturated threshold per N=8 binomial CI). Re-instate if condition met; revoke entirely otherwise.
  • PM-proposed Phase 4.3 clarification note REVOKED — original "qwen-thinking outperforms claude" finding was artifactual; Phase 4.3 verdict (72.2% T2 reasoning failure) remains binding as authored.
  • Amendment 5 §judge_metric_design STAYS — judge ensemble metrics computed on actual response content; methodology valid regardless of artifactual shape labels.

Cost impact: $4.95 sunk + $5 re-run new = ~$10 total NULL-baseline. Cumulative Faza 1 spend post re-run: ~$25.13. Headroom under $115 cap: ~$90.

Unaffected: mutation oracle 10 candidates ($1.43, valid); corpus 50/50; manifest v7 Amendments 1-5 conceptually correct; Pre-Gen-1 cost projection $14.86 still valid (cost is shape-independent in practice).

A.12 — F-saturated-baseline rule (Amendment 5 — PAUSED per Amendment 6)

For shapes at saturated NULL-baseline (trio_strict_pass_rate (op ii) = 1.0 = 100% all evals pass), §F condition 1 ≥+5pp delta is structurally inapplicable (cannot improve beyond 100%). Reformulated acceptance:

  • (a) No-regression: Best GEPA candidate maintains trio_strict_pass_rate = 100% (i.e., all 8 Gen 1 instances pass)
  • (b) Mechanistic improvement (Qwen-targeted): mean retrieval_calls per task ≥ 1.5 (escape Amendment 2 penalty zone)
  • (b') Mechanistic improvement (non-Qwen): mean retrieval_calls per task ≥ NULL-baseline retrieval mean (no regression on engagement)

For non-saturated shapes (NULL pass rate < 100%): original §F condition 1 ≥+5pp criterion applies unchanged.

Binding compliance: Manifest v7 §F_saturated_baseline_rule encodes per-shape rule selection. Initial classification at Checkpoint A:

  • Saturated (1 shape): qwen-thinking (8/8 = 100% NULL → saturated rule applies)
  • Non-saturated (4 shapes): claude (50%), qwen-non-thinking (75%), gpt (88%), generic-simple (88%) → original ≥+5pp rule

If a non-saturated shape reaches 100% on Gen 1, the saturated rule retroactively applies (documented at Checkpoint C).

A.13 — Judge metric parallel-report binding rule (Amendment 5)

For synthesis Likert evaluations (Faza 1 + downstream where pass-rate base rate may exceed 80%), the judge ensemble health metric is raw agreement rate (primary) + Cohen's κ (audit reference).

  1. Compute pairwise raw agreement at trio_strict_threshold (default 4.0): agree_pct = count(pair agrees pass-vs-fail) / n
  2. Compute pairwise Cohen's κ at the same threshold for audit reference
  3. Drift verdict (primary): min(raw agreement across pairs) ≥ 65% → PASS
  4. Drift signal (secondary): flag if ≥ 2 pairs simultaneously go below 50% raw agreement (genuine ensemble drift)
  5. PM verdict primary per checkpoint with full context; no automatic verdict from κ alone
  6. Canonical κ=0.7878 retained as audit reference with Cohen-1960 high-base-rate paradox annotation; explicitly noted as measured on LoCoMo factoid binary (~50% base rate), NOT directly comparable to synthesis Likert (~88% base rate)

Binding compliance: Manifest v7 §judge_metric_design encodes the rule. NULL-baseline at Checkpoint A reported:

  • Raw agreement (PRIMARY): Opus↔GPT 75%, Opus↔MiniMax 80%, GPT↔MiniMax 70% → min 70% ≥ 65% PASS
  • κ literal (audit): Opus↔GPT +0.342, Opus↔MiniMax −0.111, GPT↔MiniMax +0.211 → low due to Cohen paradox, NOT genuine drift

Per-judge mean distributions (N=40 evals): Opus mean 4.317 stdev 0.360; GPT mean 3.917 stdev 0.311; MiniMax mean 4.521 stdev 0.458 — internally consistent.

A.11 — JSON-mode retry texture-audit binding rule (Amendment 4)

Whenever JSON-mode response_format is used to retry corpus instances (or any prompt-controlled generation):

  1. Spot-audit retry instances against same quality criteria as original spot-audit (same validateInstance rules per manifest v7 §corpus_design.per_instance_quality_floor)
  2. Side-by-side narrative texture comparison: pick 5 random instances from originals using a different seed than the original spot-audit (e.g., spot-audit uses seed=42, texture audit uses seed=99); read first 2 documents from each retry + each sampled original
  3. Score texture match qualitatively (paragraph length, sentence length, bullet density, table density, pronoun register, persona-stage consistency, framing) AND quantitatively (per-metric mean delta vs original-sample mean)
  4. If texture drift detected (visibly shorter/longer paragraphs, different framing, different register): PIVOT TO previous-corpus path (e.g., accept partial corpus); document drift as caveat in checkpoint addendum + manifest amendment
  5. If texture matches: accept retried-corpus version, kick downstream phase

Binding compliance: Manifest v7 §amendment_4_integration.texture_audit_binding_rule encodes the rule. Rationale per Amendment 4: JSON-mode response_format changes generation control flow (constrained decoding); subtle narrative texture shift possible that non-side-by-side spot-audit doesn't catch. Insurance value > 5-10 min audit cost.

This rule was first invoked in the Pre-A halt-and-PM addendum (benchmarks/results/gepa-faza1/corpus/h3-spot-audit-pre-a-addendum.md) for the 3-cell JSON-mode retry; verdict was NO_DRIFT_DETECTED, accept 50/50.

PM may extend scope to future JSON-mode retries within Faza 1, Faza 2 expansion, or Phase 5 GEPA-evolved variant.

§B — Manifest v7 audit chain (SHA pins at LOCK time)

Item SHA-256 Path
Manifest v7 (Amendment 6 supplemented — CURRENT BINDING) 0b55d8e353299594254e1a4a76f26f53014d726315dc6a0e5d6dc1a3a44a368a benchmarks/preregistration/manifest-v7-gepa-faza1.yaml (in worktree) — supplemented at 2026-04-28T15:30:00Z; NULL-baseline shape-override bug fix + Amendment 5 §F-saturated PAUSED + Phase 4.3 clarification REVOKED
Manifest v7 (Amendment 5 — superseded by Amendment 6) 062dfc4935aaa89f0b25595c5dc3ce4af06c95c4c261075a1f0226d8af3f3dee same path — historical SHA at 2026-04-28T11:00:00Z; judge metric parallel-report (STAYS) + F-saturated-baseline rule (PAUSED)
Manifest v7 (Amendment 4 — superseded twice) 1f7a6d6fa01403f6c8d6855893adbfa5e82898a81b7583cfa55628e5eba60196 same path — historical SHA at 2026-04-28T01:00:00Z; corpus retry methodology + texture-audit binding rule
Manifest v7 (Amendment 3 — superseded twice) e43d13793535077c92a0e2c24f948ebb9d6e04000293690fdf38c4ba957aa972 same path — historical SHA at 2026-04-28T00:45:00Z; cost cap raise + Pre-Gen-1 re-projection rule
Manifest v7 (Amendment 2 — superseded twice) 583712dde139ffc87fb1ab21643f68d52c56469ded9e8090a624980b05969beb same path — historical SHA at 2026-04-28T00:30:00Z
Manifest v7 (initial LOCK — superseded thrice) 1d592a6113c918b7a07fc9aba748c8bdd12a6ce1c6943943c0492678299fa700 same path — historical SHA at 2026-04-28T00:00:00Z initial lock
H3 corpus JSONL (50 instances, BINDING) file: 9fa2bef83eb604f361419bf0ead70cf1560484a44ea01c5ebdc170a2c25c4ea3 / canonical fields: 9336ae2467e0728f20dd64a8972e3095b795f248676d679039bd1dd79a11bfef benchmarks/results/gepa-faza1/corpus/h3-northlane-cfo-50-instances.jsonl
H3 corpus pre-retry (47 instances, historical) cc9b9ae210cbd20f48f98675a45551366eebb9aa15fca93fd2eda6b366a2b912 same path — superseded by 50-instance version post Option B retry
Manifest v6 (parent inheritance) 5d5c1023421cd1a79f4913bb4c0a59415e21f50797255bff7dfec8e16b68e3ed benchmarks/preregistration/manifest-v6-preregistration.yaml
κ anchor file 657d4490bab28d35cf8a9c3ccea8a6b79e92835d700155184e51f3900836684c benchmarks/calibration/v6-kappa-recal/_summary-v6-kappa.json
κ memo 24b18112f7648ea3aa235281af19970ff4712925124301a0e60a8fd05bf5bb33 benchmarks/calibration/v6-kappa-recal/v6-kappa-memo.md
κ analysis 457357db1ad7f5941c045c3ef6724b653d2050ba8a4b61bf3f02a751adae5d47 benchmarks/calibration/v6-kappa-recal/kappa-v6-analysis.md
Pilot runner (judge config archeology source) 8a6251e2fc4e3c44ba2f23bfe7a452c316cd58f2d30a5ae45928238d72e01104 scripts/run-pilot-2026-04-26.ts
Cell-semantic boundary anchor (whole file) 1a9fa329e4b66ed9f0abe8bc22cbbf0124e0c879e1e78ec806d557cab25bc94d packages/agent/src/prompt-shapes/types.ts
MULTI_STEP_ACTION_CONTRACT (linchpin string) 70a1701dfa126f8dc1df9c116f0a8469da005821ecadc59d9b8f348568e755ba byte-level SHA of constant body (252 bytes)
Baseline shape: claude.ts cbaf0c37b067b025a1fe97f2feeec11fae4070a8b3fcfaad1da8775dda451cc0 packages/agent/src/prompt-shapes/claude.ts
Baseline shape: qwen-thinking.ts 848a4e4917baa5c7bbcc3bb35fb8cb4b4ac8f0ab537243f14cbef3a99197aacb packages/agent/src/prompt-shapes/qwen-thinking.ts
Baseline shape: qwen-non-thinking.ts 35be379be9a8caafc2c419e32da5f63f92fc83f6f6d70d9df76029c1e8584572 packages/agent/src/prompt-shapes/qwen-non-thinking.ts
Baseline shape: gpt.ts 5dc6d750d52a68feb9d37ad8384b2bcd59d70962066122ff086b0e5888413576 packages/agent/src/prompt-shapes/gpt.ts
Baseline shape: generic-simple.ts 81189817f560e26a69394248d8bd9089cae72c7d40825323e2b7407e36026172 packages/agent/src/prompt-shapes/generic-simple.ts
κ canonical value 0.7877758913412564 constant — drift band [0.7378, 0.8378] (±0.05)

The 5 baseline shape SHAs serve as the delta-zero reference for the GEPA mutation validator. Each Gen 0 NULL-baseline candidate must match its baseline SHA exactly (zero diff). Each Gen 1 mutation candidate must produce a non-zero diff in shape body but zero diff in types.ts/selector.ts/index.ts/metadata-except-evidence_link.

§C — Substrate anchor + isolated worktree (Discovery 4.5)

  • Branch: feature/c3-v3-wrapper
  • Anchor commit: c9bda3d6dd4c0a4f715e09f3757a96d01ff01cd7 (Phase 4.7 — compression-engaged-end-to-end assertion test post-fold-in)
  • Anchor verified ancestor of HEAD: PASS (verified 2026-04-28 via git merge-base --is-ancestor c9bda3d HEAD)
  • Isolation method: git worktree add D:/Projects/waggle-os-faza1-wt c9bda3d — detached HEAD, race-condition-guarded against CC-1 parallel Phase 4.4/4.5 work
  • All Faza 1 reads/writes go through the worktree. Main repo D:/Projects/waggle-os receives only the final integration commits at Checkpoint C (cherry-pick or merge; CC-2 designs integration sequence).

Note on remote: git fetch origin feature/c3-v3-wrapper returned fatal: couldn't find remote ref — repo has no origin remote configured for this branch. Substrate freshness verified locally only via git rev-parse + ancestry check. This does NOT affect Faza 1 (work is local; integration-back-to-branch is local; no fetch dependency).

§D — Cost discipline (per brief §5 + Amendment 1 §4 + Amendment 3 cap raise)

Phase Subtotal (Amendment 3) Running cumulative Pre-Amendment 3
Corpus generation (50 × Opus 4.7 @ $0.27/inst) $13.58 $13.58 $5.00
NULL-baseline (5 shapes × 8 instances) $20.00 $33.58 $25.00
GEPA Gen 1 (5 × 3 × 8) $60.00 $93.58 $85.00
Held-out validation (5 × 1 × 5) $12.50 $106.08 $97.50
Mutation oracle (5 × 2 × 2) $3.00 $109.08 $100.50
  • Hard cap: $115.00 (raised from $100 by Amendment 3)
  • Internal halt: $90.00 (raised from $80 by Amendment 3)
  • Corpus generation halt: $15.00 (raised from $7 by Amendment 3 — 40% buffer over $13.58 expected)
  • Super-linear sub-rule (§A.7): if mid-run actual exceeds projection by >30%, halt
  • Pre-phase-boundary re-projection (§A.10, Amendment 3 NEW BINDING RULE): Pre-Gen-1 halt if projected Gen 1 cost (from actual NULL-baseline per-eval telemetry × 5×3×8) exceeds $78
  • Audit cadence: every 20 evaluations

Amendment 3 rationale (binding): Inherited $0.10/instance generic LLM cost estimate was 170% off vs actual Opus 4.7 cost of $0.27/instance. Cost correction is for inherited error, NOT scope expansion. Quality dimension parity with pilot 2026-04-26 baseline preserved (~6700c materials matches pilot ~5300c) for apples-to-apples Phase 5 comparison. This precedent does NOT apply to scope expansion requests.

§E — 4 mandatory halt-and-PM checkpoints (per Amendment 1 §5)

# Checkpoint Cumulative Trigger PM action
1 Pre-A (NEW) ~$5 Post 50-instance corpus generation + 5-instance random spot-audit Ratify corpus quality + NULL-baseline kick authorization
2 A ~$25 Post NULL-baseline 5 shapes × 8 instances Ratify NULL trio-strict in 18-24% range + κ stability + Gen 1 kick
3 B ~$50-65 Mid-Gen 1 (after 30 evaluations) Ratify intermediate κ + cell-semantic violations review + complete Gen 1
4 C ~$100 Post held-out validation (5 shapes × top-1 × 5 instances) Acceptance verdict per §F + Faza 2 expansion or PHF fallback

Between checkpoints, CC-2 proceeds without further PM interaction unless any sub-rule trigger fires (per Amendment 1 closing line). Sub-rule triggers per brief §5: cost breach, κ drift > 0.05, cell semantic violation, 2 consecutive invalid mutations, API blocker.

§F — Faza 1 acceptance criteria (binding — 4 must-hold conditions + 1 false-positive guard)

Per brief §4 with Amendment 1 §6 + Amendment 2 §5 updates on condition 1 + new §F.5 false-positive guard:

  1. Best GEPA candidate per shape beats NULL-baseline by ≥+5pp on trio_strict_pass rate (where trio_strict_pass = trio_mean ≥ 4.0 per Ask B ratification — primary operationalization (ii)).
    • For Qwen-targeted shapes (qwen-thinking, qwen-non-thinking) ADDITIONALLY: best candidate must have mean retrieval_calls per task ≥ 1.7 (engagement gap closed by ≥50% relative to Qwen baseline 1.33 → Opus parity 2.33). Per Amendment 2 §5.
    • Reported with σ-aware disclaimer per §A.5: +5pp is fitness signal, NOT statistical claim.
    • Supplementary report: operationalization (i) per pilot runner line 657 (judge-mean ≥ 3.5 quorum) reported in parallel column for cross-validation.
  2. At least 3/5 shapes show positive delta (avoids cherry-picking single shape that lucked out).
  3. Trio judge κ remains within ±0.05 of canonical 0.7878 (drift band [0.7378, 0.8378]).
  4. Zero cell semantic violations detected per §A.4 mutation validator audit.

§F.5 (NEW per Amendment 2 — false-positive evolution guard):

If best Qwen-shape candidate achieves +5pp trio_strict delta WITHOUT closing retrieval engagement gap (mean retrieval_calls per task < 1.5), this signals false-positive evolution (improvement via mutation-noise rather than mechanistic fix). Result: candidate REJECTED, shape marked FAIL even if other criteria pass. PM ratifies whether to re-run mutation generation with stronger anti-premature-finalization scaffolding or escalate.

On PASS (all 4 conditions hold AND no §F.5 trigger): Faza 2 expansion brief authoring authorized; CC-1 Phase 5 NULL-baseline run gated on Faza 1 PASS.

On FAIL (any single condition fails OR §F.5 triggers): fallback PHF per decisions/2026-04-26-decision-matrix-self-judge-reframe.md; GEPA work parked; paper claim #2 multiplier teza reframes academically.

§G — Operational sequencing (Faza 1 task ordering)

Per Amendment 1 §7, with Pre-A and progress checkpoints:

  1. ✅ Manifest v7 authored (this LOCK references SHA 1d592a6113c918b7a07fc9aba748c8bdd12a6ce1c6943943c0492678299fa700)
  2. ✅ Launch decision LOCK authored (this file)
  3. NEXT: Build GEPA harness scaffold + tests (≥80% coverage):
    • Selection logic (top-1 per shape selection)
    • Mutation validator (cell semantic preservation)
    • κ audit utility (per-batch κ computation against canonical anchor)
    • Cost governance + super-linear projection tracker
    • Output: benchmarks/gepa/faza-1/ directory + tests
  4. NEXT: Generate 50-instance H3 corpus via Opus 4.7 oracle:
    • 5 task families × 5 personas × 2 company stages = 50 cells
    • Each instance ≥6 source documents + 6-dim Likert rubric
    • Output: benchmarks/results/gepa-faza1/corpus/h3-northlane-cfo-50-instances.jsonl
  5. NEXT: Spot-audit 5 random instances per §A.8; author Pre-A checkpoint report
  6. HALT: Pre-A → PM ratify corpus + NULL kick auth
  7. NULL-baseline run → Checkpoint A halt
  8. Gen 1 → Checkpoint B halt → completion of Gen 1
  9. Held-out validation → Checkpoint C → Faza 1 verdict

§H — LOCK semantics + amendment policy

This decision is LOCKED upon authoring (timestamp 2026-04-28). Subsequent amendments require:

  • New file: decisions/2026-04-28-gepa-faza1-launch-amendment-N.md (where N = sequential integer)
  • New manifest: benchmarks/preregistration/manifest-v7-gepa-faza1-amendment-N.yaml if methodology changes
  • PM ratification recorded in amendment file header
  • This LOCK file remains immutable except for its chain frontmatter list (which appends new amendments)

CC-2 may NOT modify this file mid-run except via the chain extension. Any deviation discovered mid-run triggers halt-and-PM per brief §10.

End of Faza 1 Launch Decision LOCK. Standing READY for §G step 3 (GEPA harness scaffold + tests).

Reference in New Issue View Git Blame Copy Permalink