# Docs index Start with **STATUS.md**: what exists, what is proven, what is missing. | doc | purpose | |---|---| | [STATUS.md](STATUS.md) | the current state of the product, evidence, and gaps | | [METHOD.md](METHOD.md) | the drift math, panels, validity gates | | [META-MODELS.md](META-MODELS.md) | how each Meta model was used, plane by plane | | [RESULTS.md](RESULTS.md) | every measurement, with caveats and failed runs | | [HARMFUL-MATRIX.md](HARMFUL-MATRIX.md) | per-request compliance across checkpoints, no harmful text | | [REPLAY.md](REPLAY.md) | the intensity-recovery attempt, its pre-registered gate, and what is blocked | | [CIRCUIT-PLAN.md](CIRCUIT-PLAN.md) | pre-registered circuit fingerprint; two failed runs and the next experiment | | [CROSS-FAMILY.md](CROSS-FAMILY.md) | Qwen, Mistral and Glimmer under one frozen judge | | [JUDGE-BENCH.md](JUDGE-BENCH.md) | judge selection by kappa against a 48-item gold set | | [CLASS-30B.md](CLASS-30B.md) | every runnable Glimmer 30B variation, one table | | [MUSE-CODE.md](MUSE-CODE.md) | what Muse Code built, and how the review trail is visible | | [SPARK_OPEN.md](SPARK_OPEN.md) | the second-family plan, and its gated-access blocker | | [REPLAY_PROTOCOL.md](REPLAY_PROTOCOL.md) | the pre-registered tolerance for the replay test | | [SUBMISSION.md](SUBMISSION.md) | deadline, rubric, required artifacts, ownership | | [VIDEO.md](VIDEO.md) | the demo video: chapters, pipeline, what it must not do | | [VIDEO-DESCRIPTION.md](VIDEO-DESCRIPTION.md) | the public description of the live-run cut | | [DEPLOY.md](DEPLOY.md) | three ways to deploy: CI gate, service, harness seam | | [INTEGRATION.md](INTEGRATION.md) | Meta-internal and Hugging Face integration flows | | [LIFECYCLE.md](LIFECYCLE.md) | label states: pending, verified, review, recalled | | [AGENT.md](AGENT.md) | the watch, triage, scan, alert loop and the agent tool | | [COMPUTE.md](COMPUTE.md) | providers, costs, and why attribution is not on Baseten | | [TRAINING.md](TRAINING.md) | the LoRA environment and the free-GPU path | | [THREAT_MODEL.md](THREAT_MODEL.md) | what the tool assumes an attacker can do | | [REPO_POLICY.md](REPO_POLICY.md) | licensing and what is never committed | ## One-line summary Keel is a checkpoint validity harness: it measures drift between a derivative and its base, refuses to certify a measurement it cannot trust, records the label state, and proves each card by fingerprint.