Quanticomp

Living Record — the verified-answer layer for the recordings a company already owns. Companies record everything and can ask it nothing. Living Record answers from a company's own recordings in three states: computed deterministically; reasoned only over retrieved evidence, cited to the exact second; or refused — "Not in the record yet" — when the record doesn't support an answer. A model composes only where the record supports one. The demo act — Ingest on Demand: a question asked of an unseen corpus is refused; the corpus is ingested live, and the question answers, citation attached (stand-ins: Stripe Developers, 34,833 words; Google Cloud Tech, 47,734; recompute-verified). Production proof: a 653-episode catalog — 604 transcribed, 10,197,924 words — answerable in the same three states today. Tooling, as it ran: Muse Wonder (Meta Model API) is the reasoning tier of record; on the production deployment the composing model is Muse Spark, served by Fireworks, disclosed in-frame on the site itself. The $0 legs are local open weights: faster-whisper (OpenAI's Whisper weights; 34 episodes) and the Llama 3.2 3B redaction recognizer (Meta's weights). Reasoned answers: ≈$0.010–0.011; computed and refused: $0.00. No customer is claimed. Built by an agent organization — the bounded claim: the decisions, gates, and presence were the founder's; the execution was the agents'. Founding day: 835 submissions in 174 active founder-minutes (session-span estimates; both labels travel). New in-window: composition, harness, demo, films, Anatomy of a System. Pre-existing: Quanticomp, the production deployment. The team of record is one human — Eric Lefford, founder. Its own record runs in three tiers. Tier 1, public: Ask the Record answers from the public-layer record and refuses outside it ("Not in the public record yet"). Tier 2, the build record: the project wiki lives on the web behind the owner's credential gate — computed, cited, or refused. Anonymous access is refused; it is not a public surface, and its access is the credential, not a link — the owner's credential, and, for the judging window, a dedicated judges' credential scoped to this tier alone, so the build record can be tested hands-on. Tier 3, private: the founder's conversation record — it exists, it is his alone, and it is reachable only behind his credential. It is named here by its existence and its rule, and nothing of its content appears in this package or any public surface; the decision grammar that governs it (judgment visible, life private — the founder's ruling, Oct 26) is itself part of the public record. Most programs can't tell you how they work. We tell you how they were built: Anatomy of a System documents this build in the open (264 of 264 units, zero unowned). The interrogatives substantiate the receipts and help us identify improvements. In a world of AI development, traceability and receipts seems to be the only proof.