Systems, evidence, and what comes next.

Latest evaluation Guard off → on, as summarized in my August 2026 resume. The 7B open model's unsafe-action rate fell from 0.35 to 0.0. The frontier model already had a 0.0 unsafe-action rate; its over-refusal rate fell from 0.375 to 0.0.

My work Case schema, four-way action scoring, data-driven policy, out-of-model runtime guard, and evaluation harness. A committed response cache supports offline reproduction without credentials. This is instantiation and measurement of prior enforcement approaches, not a new architecture.

Earlier browser demo The illustration and linked sandbox show a separate 14-case version with deterministic stand-in agents, checked against the original harness over 112 episodes. They do not run language models or reproduce the 20-case results above.

Honest limit These are small synthetic-benchmark results, not evidence of clinical safety or generalization. For the browser demo, the cases and policy share an author; zero policy error shows self-consistency. Independent case authoring and held-out validation remain important next steps.

My work Backend architecture, LLM integration, prompt contracts, shared memory, World Book retrieval, and decision state.

Next Structured user testing of the public branching demonstration.

My work Rule engine, legal-move model, MCTS opponent, responsive interaction design, accessibility, and deployment.

Next Export the trained policy/value checkpoint for browser inference and compare it against the search-only bot.

Finding The binding constraint was how the budget was allocated, not how large it was. Holding back the characters the rendezvous would need, instead of spending them first-come on map facts, moved the working threshold from 32 characters down to about 12.

Next Price patrol danger as a cost that scales with the remaining clock rather than treating it as impassable; the residual failures are agents orbiting a guard with an unspent risk budget.

What decides the claim Substituting random preference pairs for targeted ones degrades every metric, so the gain comes from what is contrasted, not from adding a stage.

Stated limits Single-run (seed 42), model-selected on the same validation split the metrics are reported on, with no separate held-out test split. Monotonicity is guaranteed by construction; the VideoPrism backbone is frozen, so only the projections, fusion, and ordinal head are trained.

Status A complete 9-page manuscript exists. No venue is claimed.

Finding Prediction checks are useful as a dense feature, but too brittle to serve as the whole reward.

Next Validate goal-conditioned value differences on held-out trajectory prefixes.

Design choice Local Qwen summarization is the default; paid APIs remain optional.

Next Add citation-aware chaptering and an evaluation set for summary fidelity.