ippadmin/verdicttank
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

v4.1 go-live cut, Moonshot to Mistral swap, v5.x superseded

Browse Source 
Mark v5.0/v5.1 SUPERSEDED (error-detection thesis failed at -0.40 delta); v4.1 is canonical. Cut v4.1 proposal with 12 version strings bumped. Moonshot to Mistral across production seats; production worker de-kimi'd 2026-08-18. Data-retention posture corrected 7/9 to 8/9 no-training (DeepSeek sole exception). Reconciled COGS with measured Mistral spend. Committed deployed v4.0 content and research docs to resolve the repo/live fork.
This commit is contained in:
root
2026-08-18 20:28:36 -04:00
parent c1e690a122
commit 3c07727f5c
18 changed files with 3409 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+22
View File
@@ -1,5 +1,27 @@
# VerdictTank — Changelog
## 2026-08-18
### v4.0 Dogfood + Panel Integrity Fixes
- v4.0 proposal run through the v4.0 single-pass panel (9 scoring seats + synthesis gate): **CONDITIONAL** (Proposal Strength 71, Investor Readiness 44, Composite 58)
- Root-caused glm-5.2 "flakiness": reasoning model token-starved on the financial seat (11,998 reasoning tokens at a 12k floor = zero visible output). Not a model defect — budget bug. glm-5.2 stays.
- Fixed VerdictTank-Key allowlist (LiteLLM): was v3.5-era with 11 models, 403'd on 6 v4.0 roster models. Now 17 models (11 existing + 6 v4 additions)
- Fixed MODEL_QUIRKS floors in worker.py: glm-5.2 8k→24k, deepseek-v4-pro 8k→12k, gemini-pro-latest added at 12k
- Replaced fragile extract_json parser with brace-matching (handles nested JSON, trailing commas, embedded braces) — was silently dropping the financial seat
- Full dogfood report: /root/projects/verdicttank/dogfood-v4-report.md
- Fixed three proposal arithmetic errors (Condition 6): ceiling volume 3,940→2,690 reviews/mo; annual-mix MRR reduction $2,946→$2,619; annualized prepaid $187,000→$156,000. Fixed in v4-proposal.md and index-v4.0-new.html.
- Added bottom-up SAM + competitive research (Condition 1, research half): US SAM $1.2M-$3.7M/yr, base ~$1.8M; 8 named competitors. /root/projects/verdicttank/sam-competitive.md
- Added nine-vendor data-retention + training research (Condition 4, research half): 7 of 9 no-training by default; Moonshot trains by default (no opt-out), DeepSeek silent + PRC storage. /root/projects/verdicttank/vendor-retention-facts.md
- Corrected data-handling.html third-party section: "five of nine" → "seven of nine" no-training, precise two-exception wording
### v4.1 Cut + Moonshot→Mistral Swap + v5.x Superseded
- Cut v4.1 proposal as the go-live artifact. Version bumped v4.0 → v4.1 across title, hero badge, document label, all prose references, and footer (12 strings).
- Moonshot→Mistral vendor swap executed across all production seats (market primary + crosscheck_a/c fallbacks → mistral-large-latest). Production worker de-kimi'd 2026-08-18; py_compile clean; systemd verdicttank-worker restarted active. Defensive Kimi/Moonshot VENDOR_PATTERNS scrub lines retained.
- Data-retention posture corrected: 7/9 → 8/9 no-training. Moonshot (trains by default) replaced by Mistral (no-training default, EU-resident). DeepSeek is now the sole training exception. data-handling.html updated ("Eight of the nine" / "One provider does not offer") and deployed to app3.
- Reconciled COGS with measured Mistral token usage (SpendLogs startTime): Team 9,624 prompt / 953 completion / $0.00624; Synthesis Gate 20,675 / 1,841 / $0.01310.
- Marked v5.0 / v5.1 SUPERSEDED (error-detection-density direction deprecated; its validation thesis failed at -0.40 delta). v4.1 is canonical. Lineage reconciled: v3 (08-10) → v5.x pivot (08-12, superseded) → v4.1 (08-18).
- Repo/live fork resolved: deployed v4.0 content committed (index-v4.0.html now matches deployed), v4.1 added, superseded banners on v5 pages.
## 2026-08-10
### v3 Proposal + Technical Architecture