ARGUS — I predicted that removing code execution would cause hallucinated numbers. It did not; groundedness held at 1.00. What disappeared was 100% of the derived analytical content — answers stayed accurate and got shallow. Two agents make zero LLM calls on purpose (contradiction detection and the release gate), because a check should not grade its own kind of work. Runs entirely on the free Gemini tier, so it costs $0 to reproduce.

Meraki — latency came down one measured change at a time: clause-level chunking instead of sentence-level (4.0 → 3.2 s), then a streaming TTS endpoint (3.2 → ~1.3–2.2 s). The first was a surprise: the persona asks for one-sentence replies, so a sentence-only rule meant pipelining never engaged at all. Barge-in is task cancellation — one asyncio.Task per turn. Tests include a regression guard for a real bug where a global dict handed one visitor's API keys to the next.

Arbiter — built solo. Benchmarked XGBoost head-to-head against a real LLM (gpt-oss:20b, given a fair shot) on identical data: XGBoost wins by 3.65x on PR-AUC and ~60x on latency, so the LLM only writes the explanation — delete it and every decision stays byte-identical. Rebuilding the Kaggle-winning solution's future-leaking version priced what refusing leakage costs: +0.0066 PR-AUC. Honesty is nearly free, and now that is a measurement instead of a claim.

NutriBot — refuses foods that conflict with a stated allergy or condition even when they appear in retrieved context, verified by the eval suite rather than assumed from the prompt. The LLM judge deliberately runs on a different provider than generation. Retrieval is all-mpnet-base-v2 with cross-encoder reranking, not raw vector similarity straight into the prompt.

Credit Risk — each customer appears ~8 times, so a naive row split puts the same customer on both sides. The leaky split reports macro-F1 0.8171; the honest customer-grouped split gets 0.6974. Causal shift(1) features, 10 variants ranked on out-of-fold score only, holdout touched exactly once. Two hypotheses that failed are kept in the notebook on purpose.

CrossFuse — SYNC_MIN_AUC = 0.70 was fixed in advance so the sync head's failure (AUC 0.507) could not be quietly reported as a working signal, with the root cause diagnosed: ~97% of the corpus's fakes are Wav2Lip, whose objective is correct lip-sync. Identity-disjoint splits enforced by assertion, bootstrap 95% CIs and explicit n on every number.