Files
waggle-os/docs/briefs/2026-04-24-cc1-manifest-v6-phase1-kappa-recal-brief.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
moving
2026-09-02 10:10:29 +02:00

12 KiB

CC-1 Brief — Manifest v6 Phase 1: Emission + Config Amendment + κ Re-Calibration

Date: 2026-04-24 Status: §2.0 v6 emission + §2.1 κ re-cal (Phase 1 of 2; Phase 2 = N=400 execution, emitted post-PM-RATIFY-V6-KAPPA) Authorized by: Marko Marković ("prihvatam tvoje preporuke, idemo dalje") Predecessors: Full §1.3g + §1.3h + §1.3h-C validation sequence CLOSED; MiniMax M2.7 primary + Kimi K2.6 backup selection ratified PM: claude-opus-4-7 (Cowork)


§0 Overview

Manifest v6 supersedes v5 (fc16925) as authoritative pre-registration for Stage 3 N=400 re-kick. Three deliverables in Phase 1:

  1. v6 emission — MD + YAML with full delta log from v5, new trio declaration, swap rationale, supersession of §11 frozen paths
  2. Config amendmentlitellm-config.yaml editing to add MiniMax + Kimi aliases under v6 authority (explicit supersession of v5 §11 freeze per PM ratification)
  3. κ re-calibration — full 100-instance three-way κ on new trio (Opus + GPT + MiniMax), success criterion ≥ 0.70 substantial agreement

Phase 2 (N=400 execution) gated on PM-RATIFY-V6-KAPPA. CC-1 HALTS after Phase 1 completion, does NOT self-advance to N=400.


§1 Manifest v6 emission specification

§1.1 Structure

MD + YAML twin with SHA-pinned cross-reference (same pattern as v4→v5 transition). Emit to:

  • benchmarks/preregistration/manifest-v6-preregistration.md
  • benchmarks/preregistration/manifest-v6-preregistration.yaml

§1.2 Required sections

Copy v5 structure with these substantive changes:

§0.1 Anchor: v6 anchor = this commit SHA (set at commit time).

§0.2 Parent: v5 anchor fc16925. Full parent chain documented.

§0.5 Delta log (expand from v5):

  • Judge ensemble swap rationale (Google preview quota block → Chinese GA flagship evaluation → correctness-driven selection)
  • §1.3f-§1.3h-C sub-gate summary with anchor SHAs
  • MiniMax primary selection rationale (86% correctness on splits + operational profile)
  • Kimi backup selection rationale (80% correctness + per-instance failover)
  • Abandoned candidate documentation: Gemini 3.1 Pro Preview (quota 250 RPD infeasible), Zhipu GLM-5.1 (100% GPT-echo), DeepSeek V4 Pro (GPT-alignment escalates at higher reasoning budget)
  • §11 supersession note: v5 §11 freeze on litellm-config.yaml supersedes here; new §11 in v6 pins post-amendment state

§1 Ensemble declaration (new from v5):

  • Primary judges: Opus 4.7 + GPT-5.4 + MiniMax M2.7 (openrouter routing)
  • Backup judge: Kimi K2.6 (direct routing, per-instance failover)
  • Backup activation policy: per-instance failover on MiniMax failure (API error / parse failure / timeout); if Kimi also fails → judge_ensemble_fail marker, instance excluded from final analysis
  • Disqualified candidates documented with verdict rationale

§5.2 Consistency constraint (updated from v5):

  • One judge call per instance per primary; backup activated only on primary failure
  • No prompt-level batching (preserves judge protocol)
  • Identical prompt template per failure-mode-judge.ts:245-258 verbatim

§6 κ re-calibration methodology (new from v5):

  • Sample: full 100-instance calibration set from v5 (same instances used for original κ=0.7458)
  • Three-way measurement: Opus vs GPT, Opus vs MiniMax, GPT vs MiniMax; take minimum for conservative trio κ
  • Success criterion: conservative trio κ ≥ 0.70 substantial agreement
  • If κ < 0.70: trio validity compromised, v6 re-evaluates (PM decision required)
  • If κ ≥ 0.70: PM-RATIFY-V6-KAPPA → Phase 2 authorization

§7 Throttle + concurrency (amended from v5):

  • concurrency: 1 (preserved)
  • rpm per alias: MINIMAX_M27 per OpenRouter tier (verify at probe time), KIMI_K26 per Moonshot tier (verify), Gemini rpm:20 retained but unused (orphan declaration OK, not routed)

§11 Frozen paths (supersession):

  • Supersedes v5 §11
  • Pins litellm-config.yaml state after MiniMax + Kimi additions (this commit's tree state)
  • All other frozen paths retained from v5: runner.ts, judge-runner.ts, failure-mode-judge.ts, health-check.ts

Other v5 sections (Gate D parameters, budget envelope, Fisher one-sided test specification, halt criteria, output schema) — preserve verbatim unless substantively changed.

§1.3 Budget envelope update

  • v5 envelope: $30 cap / $28 halt / ~$23 expected (Gemini preview-model premium)
  • v6 envelope: $60 cap / $55 halt / ~$50 expected (κ re-cal $25 + N=400 $25, no preview premium)

§2 LiteLLM config amendment

§2.1 Scope

Edit litellm-config.yaml under manifest v6 authority to add:

  1. MiniMax M2.7 alias via OpenRouter:
- model_name: minimax-m27-via-openrouter
  litellm_params:
    model: openrouter/minimax/minimax-m2.7
    api_key: os.environ/OPENROUTER_API_KEY
    rpm: <verify at probe time; default 60 if unspecified>
  1. Kimi K2.6 alias via Moonshot direct:
- model_name: kimi-k26-direct
  litellm_params:
    model: moonshot/kimi-k2.6
    api_key: os.environ/MOONSHOT_API_KEY
    api_base: https://api.moonshot.ai/v1
    rpm: <verify at probe time>

Verify exact model identifiers via OpenRouter + Moonshot catalog discovery before writing config (use §1.3g artefact memo as starting reference; confirm not deprecated).

§2.2 Retention

  • All existing v5 aliases (Opus 4.7, GPT-5.4, Gemini 3.1 Pro Preview) retained in config — do NOT delete Gemini alias even though unused (preserves audit trail; v6 §11 pins this state)
  • Existing rpm:20 on gemini-3.1-pro alias from Fold-in 3.5a/3.5b retained (orphan OK)

§2.3 Commit

Single commit under v6 authority. Message format: [v6] litellm-config amendment: add minimax-m27 + kimi-k26 aliases per judge swap ratification

Anchor = v6 commit. CC-1 verifies post-amendment config is the exact state v6 §11 pins.


§3 κ re-calibration execution

§3.1 Sample

Full 100-instance κ calibration set from v5 (identifier in benchmarks/results/locomo-mini-n20-retry-2026-04-24T00-02-12Z.jsonl — same authoritative source used for §1.3h split analysis).

Do NOT introduce new instances. Reproducibility-critical that new trio κ is measured on the same instances as original κ=0.7458 three-way measurement.

§3.2 Judge verdicts

  • Opus 4.7: reuse existing verdicts from v5 κ set (no new calls; 100 instances already have Opus verdicts)
  • GPT-5.4: reuse existing verdicts (no new calls)
  • MiniMax M2.7: NEW execution, 100 calls via v6 alias minimax-m27-via-openrouter, verbatim judge prompt from failure-mode-judge.ts:245-258, temperature=0.0, matched max_tokens

Total new API calls for κ re-cal: 100 (MiniMax only).

§3.3 Computation

Three pairwise Cohen's κ:

  • κ(Opus, GPT): should match historical κ baseline (~0.74-0.82)
  • κ(Opus, MiniMax): new measurement
  • κ(GPT, MiniMax): new measurement

Conservative trio κ = min(three pairwise κ values).

Also compute:

  • Raw agreement % per pair
  • Confusion matrix per pair
  • Per-cell breakdown (no-context / retrieval / full-context / oracle-context / agentic)

§3.4 Success criteria

  • κ_conservative_trio ≥ 0.70 → PASS, halt with PM-RATIFY-V6-KAPPA request
  • 0.60 ≤ κ_conservative_trio < 0.70 → BORDERLINE, halt with PM adjudication request
  • κ_conservative_trio < 0.60 → FAIL, halt with swap-path-re-evaluation request (may require Kimi promoted to primary or backup-judge strategy rework)

§3.5 Operational hedge — per-instance failover check

During κ re-cal execution (100 MiniMax calls), log:

  • Parse success rate (target ≥95/100; lower = concern)
  • Latency p50 + p95 (target p50 ≤ 25s)
  • Any openrouter routing errors → document; if rate > 5%, raise PM flag before proceeding to κ compute

If MiniMax parse rate < 90/100, PM adjudicates whether to (a) proceed with κ on valid sample or (b) halt and investigate parse issue before commit.


§4 Scope guards

  • HEAD parent = 005a19a (§1.3h-C anchor). Any drift → halt with "HEAD_DRIFTED".
  • Manifest v5 immutable — v5 anchor fc16925 remains audit-immutable; v6 is supersession, not amendment.
  • §11 frozen paths per v5 are active UNTIL v6 emission commits; v6 emission itself edits litellm-config.yaml as explicit supersession authorized by PM.
  • Other §11 files (runner.ts, judge-runner.ts, failure-mode-judge.ts, health-check.ts) remain frozen in v6. Do NOT edit.
  • No N=400 execution in Phase 1. That is Phase 2 post-PM-RATIFY-V6-KAPPA.
  • No Opus/GPT re-verdict generation. Reuse v5 κ set verdicts for those two judges.

§5 Deliverables (commit to feature/c3-v3-wrapper)

Phase 1 artefacts:

  1. benchmarks/preregistration/manifest-v6-preregistration.md (new pre-reg)
  2. benchmarks/preregistration/manifest-v6-preregistration.yaml (new pre-reg twin)
  3. litellm-config.yaml (amended with MiniMax + Kimi aliases)
  4. benchmarks/calibration/v6-kappa-recal/minimax-kappa-responses.jsonl (100 MiniMax verdicts)
  5. benchmarks/calibration/v6-kappa-recal/kappa-v6-analysis.md (three pairwise κ + conservative trio κ + per-cell breakdown)
  6. benchmarks/calibration/v6-kappa-recal/v6-kappa-memo.md (≤250 words: κ values, verdict PASS/BORDERLINE/FAIL, MiniMax operational metrics, cost, wall-clock)

Three sequential commits:

  • Commit 1: v6 emission (manifest MD + YAML only, no config edit yet) — this establishes v6 anchor
  • Commit 2: litellm-config amendment under v6 authority — references v6 anchor in message
  • Commit 3: κ re-calibration artefacts — references v6 anchor

§6 Halt ping format

Emit after Phase 1 completion:

  • phase1_verdict: PASS | BORDERLINE | FAIL | INCONCLUSIVE
  • v6_anchor: <full sha> (commit 1)
  • config_amendment_anchor: <full sha> (commit 2)
  • kappa_recal_anchor: <full sha> (commit 3)
  • v6_manifest_md_sha: <sha>
  • v6_manifest_yaml_sha: <sha>
  • minimax_identifier_used: <string> (exact from catalog)
  • kimi_identifier_used: <string> (exact, for future backup activation)
  • kappa_opus_gpt: <float> (should be consistent with historical baseline)
  • kappa_opus_minimax: <float>
  • kappa_gpt_minimax: <float>
  • kappa_conservative_trio: <float>
  • minimax_parse_success_kappa_set: <int>/100
  • minimax_latency_p50_kappa_set: <int>s
  • minimax_latency_p95_kappa_set: <int>s
  • minimax_routing_errors_count: <int>/100
  • budget_spent_phase1: $<actual> (vs $30 Phase 1 cap)
  • wall_clock_phase1: <duration>
  • next_step_request: PM-RATIFY-V6-KAPPA
  • cc1_state: HALTED

CC-1 does NOT self-advance to Phase 2 N=400. Awaits PM ratification.


§7 Budget (Phase 1 only)

  • Phase 1 cap: $30 (κ re-cal ~$25 expected + overhead)
  • Phase 1 halt: $35 (full escalation if exceeded)
  • Per-call timeout: 60s (MiniMax historical 16s p50; 2-min timeout safe)
  • Phase 1 wall-clock cap: 90 min

Phase 2 (N=400) budget is separate envelope (~$25) and activates post-ratification.


§8 Task #29 trace update (post-Phase 1)

  • §2.0 v6 emission:
  • §2.1 κ re-calibration:
  • PM-RATIFY-V6-KAPPA: <pending/ratified>
  • Phase 2 authorization: <blocked/authorized>

If PM-RATIFY-V6-KAPPA PASS → PM emits Phase 2 (N=400 execution) brief. If BORDERLINE → PM adjudicates with borderline-κ path decision. If FAIL → PM re-evaluates swap path (Kimi promotion, or return to Google waiting, or Branch B reduced coverage).


§9 Authorized by

PM Marko Marković, 2026-04-24 evening. Ratified three-point PM proposal: (1) per-instance backup activation, (2) v6 supersedes v5 §11 freeze, (3) full 100-instance κ re-calibration.

CC-1 may begin immediately. All prereqs in place (keys live, GroupId added, gcloud tooling retained, existing κ set artefacts on disk).