The panel that produced the validation data ran at 8 of 11 seats because two seats hit a provider credit wall mid-run and one model family could not be dispatched at all. v2.3 is the spec written in response to that failure. Full detail lives in the judge pool specification v2.3.
| Band | Seat | Model | Vendor | Scores |
|---|---|---|---|---|
| 0 | Research Agent | Grok 4.5 | xAI | No |
| A | Primary Reviewer | Claude Opus 5 | Anthropic | Yes |
| A | Cross-Check A | DeepSeek V4 Flash | DeepSeek | Yes |
| A | Cross-Check B | Gemini Pro Latest | Yes | |
| A | Cross-Check C | DeepSeek V4 Pro | DeepSeek | Yes |
| A | Legal / Regulatory | Claude Sonnet 5 | Anthropic | Yes |
| B | Financial Integrity | MiniMax-M3 | MiniMax | Yes |
| B | Team / Founder | Claude Fable 5 | Anthropic | Yes |
| B | Market Reality | Qwen3.7 Plus | Alibaba | Yes |
| B | Execution Feasibility | GPT-5.2 Pro | OpenAI | Yes |
| C | Synthesis & Integrity Gate | Kimi K2.6 | Moonshot | No |
Nine distinct vendors across eleven seats. Nine distinct scoring models. Zero model double-ups: no single model occupies two scoring seats, which is the constraint that keeps correlated failure out of the panel mean. Maximum vendor concentration is Anthropic at 3 of 11 (27.3%), comfortably inside the 40% ceiling. DeepSeek holds 2 of 11 (18.2%). Every remaining vendor holds exactly one seat.
Before any scoring begins, the orchestrator pings every rostered model with a 5-second probe and writes the result to a per-model health file. Any model returning HTTP 400, HTTP 429, or a no-healthy-deployments error is swapped for its pre-assigned failover before a single scoring call is spent. The 2026-08-12 run burned roughly 12 dispatches discovering dead models at runtime. That failure mode is now closed.
Every seat carries a named failover from a different vendor, resolved at gate time rather than improvised mid-run. Failover selection preserves both the vendor-diversity ceiling and the no-double-up rule, so a degraded panel is still a valid panel rather than an accidentally correlated one.
The provider credit exhaustion that cost two seats mid-run is now detected at the gate and treated as an availability failure, not an error. Anthropic capacity has been restored and the Primary Reviewer seat runs Claude Opus 5 as specified.
One frontier model family proved structurally incapable of running as a panel seat under our orchestration and is permanently excluded from the roster, not merely deprioritized. Excluded models cannot be selected as a failover target either.
v2.3 targets a critical path of roughly 113 seconds, against 218 seconds measured on the v2.1 architecture. The improvement comes from band parallelism: Band A and Band B seats execute concurrently rather than sequentially, and the Synthesis seat is the only stage that must wait for all scoring seats to return.
Every scoring seat rates the proposal 1 to 10 on the same ten dimensions, so panel spread is computed dimension by dimension and not only in aggregate:
The Band C seat never scores. It reads all scoring output and performs a fixed checklist: verify score arithmetic, compute panel means and per-dimension spread, compute the delta against the solo baseline, flag every score more than 1.5 standard deviations from the panel mean with a written rationale, and confirm the verdict follows from panel evidence rather than from the Primary Reviewer alone. That gate is what turns eleven opinions into one auditable report.
This is the clearest case in the validation set, because it is the one where score elevation delivered exactly zero and error detection delivered everything. Real scores from the 2026-08-12 run.
Median 6.45 · standard deviation 1.758 · spread 5.4 (min 2.8, max 8.2)
Spread 2.4 · delta +0.04 · statistically tied with the panel
| Seat | Model | Score | Sigma from mean | Flag |
|---|---|---|---|---|
| Band B · Financial Integrity | MiniMax | 8.2 | +1.17 | Within 1.5 sigma |
| Band B · Market Reality | Gemini | 7.7 | +0.89 | Within 1.5 sigma |
| Band A · Primary Reviewer | Opus 5 | 6.9 | +0.43 | Within 1.5 sigma |
| Band A · Legal / Regulatory | Qwen | 6.5 | +0.21 | Within 1.5 sigma |
| Band A · Cross-Check C | DeepSeek V4 Pro | 6.4 | +0.15 | Within 1.5 sigma |
| Band A · Cross-Check B | Gemini | 6.2 | +0.04 | Within 1.5 sigma |
| Band B · Team / Founder | Kimi | 4.4 | -0.99 | Within 1.5 sigma |
| Band B · Execution Feasibility | DeepSeek V4 Pro | 2.8 | -1.91 | OUTLIER |
The 2.8 was not a grumpy model. The Integrity Gate challenged it at 1.91 sigma and it survived the challenge on evidence: a contractor budget broken 4x to 7x, an architecture that regressed from v1 with no schema and no API contract, and a founder workload of 37 engagements plus 950 hours alongside an MSP day job. The Team seat (4.4) and the Band A generalists (6.2 to 6.9) all acknowledged the same workload problem. They weighted it less severely. The Execution specialist is the only seat that forced it into the verdict.
Panel verdict: CONDITIONAL GO. Estimated rework 15 to 20 hours. The solo baseline returned a 6.10 and none of the six conditions above.