Adversarial-refinement pipeline
Courtroom.
Three opposing intents on one shared base model — Proposer (ADVOCATE), Defender (SKEPTIC), Judge (WEIGH) — review meaningful model outputs before they reach users.
Foundation · Qwen 3.6-27B (Apache 2.0)
Every meaningful model output is argued by Proposer, defended by Skeptic, weighed by Judge.
The honest description.
Three opposing intents on one shared base model — Proposer (ADVOCATE), Defender (SKEPTIC), Judge (WEIGH) — review meaningful model outputs before they reach users.
Run the model on your own input.
Demo runs in your browser against a mocked edge — no signup, no production data. The same shape as the live route.
Try the Courtroom (mock)
Loading the demo…
How it works.
Tier-routed adversarial verification of model outputs. DEEP (T1/T3) runs all three; STANDARD (T2) is Judge-only; LIGHT (T4) is rule-based; BYPASS (T5) is log-only.
What it learned from, and how we watch it.
No training data of its own. Inherits the licensed base-model weights (Qwen 3.6-27B). Prompt templates ship in-repo with the framework.
Drift baseline not yet documented for this model.
- All 8 registered slugs are at rollout DISABLED — the framework ships dormant pending per-slug activation PRs.
- Adversarial verification reduces but does not eliminate hallucination on long-form output.
- Circuit breaker falls back to tier-differentiated handling on adapter failure; T1 hard-fails, T3-advisory soft-REFINEs, T2 passes with a flag.
- Adds latency proportional to tier depth; SHADOW rollouts pay the cost without surfacing the verdict.
Across the house and the API.
No public routes documented yet.
Every release, traceable.
Data Mukodi holds about you is subject to DPDP §12 erasure (use /consent in the app), and any concern about this model can be raised via [email protected] under DPDP §13 grievance, or via /security.