Scirium v2 — VerdictTank Panel Verdict
11-seat multi-model panel · v2.3 specification · executed 2026-08-17 · billed to VerdictTank-Key
CONDITIONAL GO
Panel mean 5.82 / 10 · spread 1.5 · 1 outlier (optimistic) · proceed with 4 mandatory conditions
Panel Composition
9 scoring judges across 8 distinct models, plus 1 non-scoring Research Agent (Band 0, xAI Grok 4.5) that fact-checked every material claim against live web sources before scoring began.
| Band | Seat | Model | Score | Verdict |
|---|
| Band A (blind) | Primary | Claude Opus 5 | 5.4 | Conditional Go |
| Band A (blind) | Cross-check A | DeepSeek V4 Flash | 6.35 | Conditional Go |
| Band A (blind) | Cross-check B | Gemini Pro Latest | 6.35 | Conditional Go |
| Band A (blind) | Cross-check C | DeepSeek V4 Pro | 6.6 | Conditional Go |
| Band A (blind) | Legal / Regulatory | Claude Sonnet 5 | 6.1 | Conditional Go |
| Band B (informed) | Financial Integrity | GPT-5.2 Pro* | 5.6 | Conditional Go |
| Band B (informed) | Team / Founder | Claude Fable 5 | 5.5 | Conditional Go |
| Band B (informed) | Market | Qwen 3.7 Plus | 5.4 | Conditional Go |
| Band B (informed) | Execution / Feasibility | GPT-5.2 Pro | 5.1 | Conditional Go |
* Financial seat fell back to GPT-5.2 Pro per spec §2.3 — the rostered MiniMax-M3 model was unavailable upstream (insufficient balance, HTTP 402). This is the spec's own F1 fallback chain.
Panel Statistics
Product Thesis Test — FAIL
The panel must beat the solo Opus 5 baseline by >0.5 points to justify premium multi-model pricing. It beat it by +0.42, so the thesis formally fails. This is expected and normal: the panel's value is adversarial error detection, not score inflation. It found four recurring critical flaws that a single model would have soft-pedaled.
Outlier Flag (1.5σ)
CC-C (DeepSeek V4 Pro) at 6.6, +1.55σ — mild optimistic deviation. Retained, not discarded; its rationale still lands in the CONDITIONAL GO band and does not shift the mean above 7.0.
Cross-Cutting Critical Flaws
These four flaws were independently flagged by 7–9 of the 9 judges. They are the mandatory conditions on the verdict.
Critical · 9/9 judges
Unvalidated churn assumption
The §A10 churn rate (≤3.5%/mo) is a GUESS with no cohort data. It compounds to ~35%+ annual logo churn — high for sticky enterprise software. The launch gate is statistically underpowered at 3 tenants / 6 months.
Critical · 8/9 judges
Cold outbound CAC violates the 3:1 ceiling
Cold CAC of $6,205 exceeds the 3:1 LTV:CAC ceiling ($4,918) at every modeled churn rate — yet the recommended plan implicitly budgets cold additions. Condition 3 bans the channel the plan quietly funds.
High · 8/9 judges
Unresolved Scirium / Cirium trademark collision
Scirium vs Cirium (RELX / LexisNexis) is a single-letter phonetic collision, unassessed at professional level. It gates the core white-label reseller channel (Motion B) pending professional clearance.
High · 7/9 judges
Double-booked execution capacity
Technician hours are double-booked between the acquisition ceiling (3.4 warm wins/mo) and fleet operations (1.9 hrs/tenant/mo). The 1,667h build figure is unsourced, and the warm pool is exhausted by month 5.
Integrity Gate
- ✓ Score arithmetic verified across all 9 files
- ✓ Panel means, baseline deltas, and 1.5σ outlier flags computed
- ✓ Verdict derived from panel mean, not solo baseline
- ✓ Product thesis tested; failure documented honestly
- ✓ All 9 judges accounted · 100% coverage · 0 missing models
Final Recommendation
Proceed with 4 conditions. Scirium v2 is a genuinely honest, well-engineered document that self-falsifies rather than advocates — the panel's standout finding was its intellectual honesty (it labels its own v1 fictions and rebuilds them bottom-up). Resolve the four flaws above before scaling: (1) instrument churn with real cohort data before widening the launch gate, (2) reconcile the cold-outbound CAC against the 3:1 ceiling or cut the channel, (3) obtain professional trademark clearance for Scirium vs Cirium, (4) reconcile technician capacity and source the 1,667h build estimate.
VerdictTank v2.3 · band-c synthesis by Kimi K2.6 · full judge JSON and research brief on disk at /root/.hermes/cache/verdicttank/scirium-v2/