itpp-infrastructure/proposals/verdicttank/consolidation-reasoning-overlap-kimi.md

VerdictTank Pipeline Consolidation Recommendation

Lens: Reasoning / Logic Overlap Detection (Contradiction Detector)

Author: Kimi K2.6 (Reasoning-Verification specialist) Spec: judge-pool-spec.md v2.1 Target: ~15% reduction Date: 2026-08-12

---

Executive finding

ONE recommendation: Collapse Phase 10 (Reasoning-Verification) and Phase 11 (Audit Agent) into a single post-scoring synthesis seat: Synthesis and Integrity Gate.

These two roles ask the same meta-question with different labels:

Role	Stated question	Actual job
Phase 10 Reasoning-Verification	Where do judges contradict?	Read all score sets, find disagreements, synthesize
Phase 11 Audit Agent	What did the panel miss / groupthink?	Read finished review, find blind spots / groupthink, adjust confidence

Both are non-scoring meta-reviewers of the panel's own output. Neither sees the raw proposal as primary input in a unique way that the other does not eventually consume. Both produce a post-hoc integrity adjustment, not an independent domain score. Under a contradiction-detection lens, they are statistically and logically collinear seats.

---

1. Changes

Collapse

Remove	Absorb into
Phase 10: Reasoning-Verification (standalone)	New Phase 10: Synthesis and Integrity Gate
Phase 11: Audit Agent (standalone)	(same merged seat)

New merged role definition

Field	Value
Phase	10 (final)
Role	Synthesis and Integrity Gate
Model	Kimi K2.6 (Moonshot) — keeps contradiction specialty as primary
Fallback 1	Claude Fable 5 (Anthropic) — former Audit primary
Fallback 2	GPT-5.6 Sol (OpenAI) — former Reasoning F2
Scores?	No (preserves non-scoring fence)
Input	ALL score sets from phases 2-9 + assembled draft review
Output	(a) Contradiction / alignment report, (b) synthesized dimension scores via section 4.1 formula, (c) blind-spot / groupthink notes, (d) single confidence adjustment +/-0.5

Prompt contract (must stay explicit)

Merged prompt has two ordered sections, not a vague do-both:

1. Contradiction pass (former Phase 10): pairwise/cluster disagreement map across Primary, Validation, Cross-Check avg, Financial, Team, Market, Execution, Legal. Emit synthesized scores.
2. Integrity pass (former Phase 11): given the synthesis above, scan for groupthink (panel sigma too low), missed domain blind spots, and apply one confidence adjustment (+/-0.5).

Single model, single call, single output schema. No second serial hop.

Spec section edits required

Section	Edit
section 1.1 Pipeline Phases	Drop row 11; rewrite Phase 10 as Synthesis and Integrity Gate; phase count 11 to 10
section 1.2 Enterprise Lineup	13 seats to 11 seats (12 unique models still available in pool; one fewer active seat)
section 1.2 Model double-ups	Remove Fable 5: Team + Audit double-up. Fable becomes Reasoning/Audit fallback only (Team primary only)
section 2.3 Fallback table	Delete Audit Agent row; replace Reasoning-Verification row with merged role chain above
section 3.1 Hard constraints	Replace "Audit reads FINISHED review only" with "Synthesis and Integrity Gate is post-scoring, read-only, cannot open new domain scores"
section 3.3 Low-complexity Enterprise	Already drops Audit — update text to drop CC-C, Team, Legal (Audit no longer separate)
section 4.1 Scoring algorithm	Merge the two non-scoring bullets into one: Synthesis and Integrity Gate produces contradiction report + synthesized scores + confidence +/-0.5
section 4.2 Cost ceiling rule	Swap Tier A to B for non-scoring roles first (Research, Synthesis) — drop separate Audit
section 4.4 Latency	Enterprise sequential phases lose one hop — recalibrate target downward slightly (e.g. 180s to ~165s aspirational; keep 300s max)
section 6.3 Feature matrix	Replace Audit Agent (Phase 11) row with Synthesis and Integrity Gate (includes audit) — still Enterprise/WL only if desired

What this is NOT

• Not dropping Cross-Check C (that is diversity, not logic-overlap).
• Not merging Primary with Validation (Validation is deliberately contaminated by Primary; different question).
• Not merging Market with Execution (different input scopes and vertical triggers).
• Not touching Research (unique tool surface; never scores).

---

2. Reduction %

Seat / phase accounting (Enterprise standard panel)

Metric	Before (v2.1)	After	Delta
Pipeline phases	11	10	-1 phase (-9.1%)
Active seats	13	11	-2 seats (-15.4%)
Serial post-score hops	2 (P10 then P11)	1	-50% of meta-tail latency
Non-scoring roles	3 (Research, Reasoning, Audit)	2 (Research, Synthesis)	-33% meta overhead
Unique models in default lineup	11 distinct (Sonnet x2, Fable x2)	11 distinct (Sonnet still x2 Validation+Legal; Fable Team-only)	pool size unchanged
Anthropic share of active seats	4/13 = 30.8%	3/11 = 27.3%	more headroom under 33% cap

Cost reduction (directional)

• Eliminate one full Tier-A (or Tier-A-fallback) completion on every Enterprise review.
• v2.1 cited about $1.30-2.10 per standard full review — expect ~12-18% COGS cut on the judge tail (one fewer expensive synthesis call), landing near ~15% all-in when weighted by token volume of P10+P11 vs earlier short scoring calls.
• Latency: remove one sequential dependency (~15-40s depending on model) from the critical path after phase 9.

Headline reduction: ~15% (seats -15.4%; cost band -12-18%; phases -9%).

---

3. What survives

Intact domain + scoring spine

1  Research Agent               (Grok 4.5)           - grounding, never scores
2  Primary Reviewer             (Claude Opus 5)      - 10-dimension critique
3  Validation Reviewer          (Claude Sonnet 5)    - challenges Primary
4a Cross-Check A                (GPT-5.6 Sol)        - blind re-score
4b Cross-Check B                (Gemini Pro Latest)  - blind re-score
4c Cross-Check C                (DeepSeek V4 Pro)    - blind re-score
5  Financial Integrity          (MiniMax-M3)         - cap table / unit econ
6  Team/Founder Assessment      (Claude Fable 5)     - founder-market fit
7  Market-Reality               (Qwen3.7 Plus)       - TAM / competitive / non-Western
8  Execution-Feasibility        (GPT-5.6 Terra)      - ops / timeline realism
9  Legal/Regulatory             (Claude Sonnet 5)    - multi-jurisdiction risk
10 Synthesis and Integrity Gate (Kimi K2.6)          - contradictions + audit + +/-0.5

Preserved invariants

• Research never scores
• Three blind cross-checks, isolated from Primary/Validation
• Domain specialists (Financial, Team, Market, Execution, Legal) remain separate scoring questions
• Section 4.1 weighted aggregation formula unchanged
• Cross-vendor fallback rule unchanged (merged chain still cross-vendor at every hop)
• Vendor cap still satisfied (Anthropic drops to 27.3%)
• Build validation gate (section 5.5) unchanged — still require >=0.5 panel advantage
• Fenced post-score mode retained (merged seat cannot reopen domain scores)

Why these roles were NOT collapsed

Pair	Why keep separate
Primary vs Validation	Validation is score-conditioned (sees Primary). Different epistemic job than independent re-score.
Validation vs Cross-Checks	CCs are blind; Validation is not. Collapsing destroys the blind control.
Cross-Check A/B/C	Same question by design — but vendor diversity is the product. Overlap is intentional ensemble, not accidental duplication. Do not collapse under this lens.
Market vs Execution	Different section 3.2 triggers, different evidence (competitive density vs ops timeline).
Financial vs Execution	Cap-table math is not shipping realism.
Legal vs Audit	Legal is domain-scoring on statutes; Audit was meta. Legal stays; Audit folds into Synthesis.
Team vs Audit (same model Fable)	Team scores founders; Audit was meta. After merge, Fable stops double-duty — removes same-model Team-to-Audit soft contamination path.

---

4. Risks

Risk	Severity	Mitigation
Single-call attention split — model skimps on contradiction map OR groupthink scan	Medium	Enforce two-section JSON schema; reject output missing either contradictions[] or integrity.confidence_delta.
Loss of independent second meta-reader — two weak meta-passes can catch what one strong pass misses	Medium	Keep dual-section prompt; optionally run Fable fallback as shadow audit on 10% of Enterprise reviews for 90 days (not on critical path).
K2.6 empty-output / reasoning-token burn (section 4.5) now single-points the entire meta-tail	High	Raise min tokens to 768-1024 for merged seat; on empty go immediate F1 (Fable). Circuit breaker already in section 4.2.
Confidence +/-0.5 applied by same model that synthesized scores — self-grading bias	Medium	Constrain: confidence delta may only move on documented groupthink (panel sigma under threshold) or explicit missed-claim list; ban free-form "I feel +/-0.5".
Pro tier marketing currently sells no Audit; matrix row rename may confuse	Low	Feature matrix: Synthesis and Integrity (Enterprise) — capability retained, seat count honest.
Low-complexity Enterprise drop list referenced Audit explicitly	Low	Update section 3.3 text only.
Outcome-tracking granularity — cannot attribute T+90 errors to Reasoning vs Audit separately	Low	Accept; meta-tail was always under-identified. Track merged seat as one accuracy series.
Over-collapse temptation — if 15% feels good, next PR kills CC-C or Validation	Process	This lens only authorizes logic-collinear non-scoring merges. Ensemble scorers are out of scope.

Explicit non-risks

• Vendor diversity of scoring panel unchanged (still 8 vendors available; scoring seats still multi-lab).
• No new same-vendor Primary/meta self-audit (Kimi is not Anthropic Primary/Validation).
• Fable no longer audits after sitting Team — contamination risk decreases.

---

5. New diagram

Before (v2.1) — 11 phases / 13 seats

[Proposal]
    |
    v
+-----------------+
| 1 Research      |  Grok 4.5          (no score)
+--------+--------+
         v
+-----------------+
| 2 Primary       |  Opus 5            (scores)
+--------+--------+
         v
+-----------------+
| 3 Validation    |  Sonnet 5          (scores, sees Primary)
+--------+--------+
         v
    +----+----+
    v    v    v
+----+ +----+ +----+
|4a  | |4b  | |4c  |  Sol / Gemini / DeepSeek  (blind scores)
+--+-+ +--+-+ +--+-+
   +-----+-----+
         v
+-----------------+
| 5 Financial     |  MiniMax-M3
+--------+--------+
         v
+-----------------+
| 6 Team/Founder  |  Fable 5
+--------+--------+
         v
+-----------------+
| 7 Market        |  Qwen3.7 Plus
+--------+--------+
         v
+-----------------+
| 8 Execution     |  Terra
+--------+--------+
         v
+-----------------+
| 9 Legal         |  Sonnet 5
+--------+--------+
         v
+-----------------+
|10 Reasoning     |  Kimi K2.6         (contradictions + synth)
+--------+--------+
         v
+-----------------+
|11 Audit         |  Fable 5           (groupthink + +/-0.5)
+--------+--------+
         v
    [Final Report]

After (recommended) — 10 phases / 11 seats (~15% seat cut)

[Proposal]
    |
    v
+-----------------+
| 1 Research      |  Grok 4.5          (no score)
+--------+--------+
         v
+-----------------+
| 2 Primary       |  Opus 5            (scores)
+--------+--------+
         v
+-----------------+
| 3 Validation    |  Sonnet 5          (scores, sees Primary)
+--------+--------+
         v
    +----+----+
    v    v    v
+----+ +----+ +----+
|4a  | |4b  | |4c  |  Sol / Gemini / DeepSeek  (blind scores)
+--+-+ +--+-+ +--+-+
   +-----+-----+
         v
+-----------------+
| 5 Financial     |  MiniMax-M3
+--------+--------+
         v
+-----------------+
| 6 Team/Founder  |  Fable 5           (Team only — no Audit double)
+--------+--------+
         v
+-----------------+
| 7 Market        |  Qwen3.7 Plus
+--------+--------+
         v
+-----------------+
| 8 Execution     |  Terra
+--------+--------+
         v
+-----------------+
| 9 Legal         |  Sonnet 5
+--------+--------+
         v
+-------------------------------------+
|10 Synthesis and Integrity Gate      |  Kimi K2.6
|   - contradiction / alignment map   |
|   - section 4.1 synthesized scores  |
|   - groupthink / blind-spot notes   |
|   - confidence delta +/-0.5 (gated) |
|   F1: Fable 5 · F2: Sol             |
+------------------+------------------+
                   v
              [Final Report]

Overlap evidence (why this pair, not another)

                    INPUT SCOPE
                 proposal  scores  finished-review
Primary            ###      .        .
Validation         ##       ###      .
Cross-Checks       ###      .        .     intentional ensemble (KEEP x3)
Financial          ##       ##       .
Team               ##       ##       .
Market             ##       ##       .
Execution          ##       ##       .
Legal              ##       #        .
Reasoning (P10)    .        ###      ##    ]
Audit (P11)        .        ##       ###   ]  SAME META QUESTION
Research           tools    .        .     (unique — KEEP)

Overlap score P10 intersect P11:
  question similarity   ~0.85  (contradiction equiv missed-agreement/groupthink)
  input Jaccard         ~0.80  (both post-domain, panel-output-primary)
  output collinearity   ~0.75  (both emit integrity adjustment, not domain scores)
  -> COLLAPSE CANDIDATE under reasoning lens

Opus empirical echo (from v2.1 review results)

Opus 5 panel-vs-single test already showed cross-checks can be statistically tight (market dimension spread 0.00 across three CCs). That finding pressures ensemble scorer count — out of scope for this lens. The safer 15% cut is the duplicated meta-tail, not the diversity scorers the product sells. Killing Audit+Reasoning redundancy reduces COGS without touching the scoring majority (still 9 scoring judges + Research + 1 meta).

Scoring judges retained: 9 (Primary, Validation, 3xCC, Financial, Team, Market, Execution, Legal) Meta retained: 1 (merged) + Research

---

Decision line

Collapse Reasoning-Verification and Audit Agent into one Synthesis and Integrity Gate (Kimi K2.6). -2 seats (-15.4%), -1 serial phase, ~15% meta-tail cost, cleaner vendor math, removes Fable Team-to-Audit double-duty. Domain ensemble untouched. Ship as v2.2 delta on top of v2.1 blocker fixes.

---

Implementation checklist (for conductor)

1. Patch judge-pool-spec.md sections listed in section 1
2. Bump version header to v2.2; status remains FINAL DRAFT pending 3-5 proposal gate
3. Update feature matrix + latency table
4. Rewrite section 2.3 fallback row; delete Audit row
5. Add output JSON schema for two-section merged seat
6. Re-run vendor % assertion (expect 3/11 Anthropic)
7. Do not change Pro 7-judge composition in this PR (Enterprise matrix row rename only)
8. Optional: 90-day 10% shadow Fable integrity pass off critical path