Answer — of the dossier's — rows
before you book a bench.

One product description — composition, claims, capabilities, computed biology, provenance — travels to every surface above it, so a bench, a rule pack and a dossier all judge the same object. Rules are written in a closed vocabulary of — predicates and — actions: a regulation becomes data you can diff, review and replay, instead of code that can quietly do anything.

Lit classes carry dossier requirements today. The contract already describes — beyond them, so widening the ledger is a data change, not a schema migration.

Korea MFDS/NIFDS guidance, collected and curated, then cut into chunks and pinned by sha256. Retrieval is lexical FTS5 by default, hybrid with a —‑dimension vector sidecar on request. RAPL retrieves and structures. It never generates — the connected agent does that, and the boundary is a test, not a promise.

The figures a corpus is supposed to keep at zero. They are the receipt, and they are re‑measured every time this page is built.

Every engagement opens in one of — operation modes — analysis‑only, assistive‑edit, full‑ops — and a mode it does not recognise resolves down to — instead of erroring open. Before the first tool call it clears — preflights and proves the writable workspace lies outside the target project root; if it does not, the run is blocked and told which two roots to set. The boundary is computed at entry, not promised in a prompt.

Work is split across a team rather than a longer prompt: — flow topologies, with subagents pinned to — isolation modes. The orchestrator may orchestrate, review, approve and promote — and is barred from executing tests, writing runtime state or running a subagent itself, so the reviewer can never quietly become the author. Long runs survive on a compaction pass that protects head and tail, prunes stale tool output first, and hands the middle forward marked as background rather than as fresh instructions.

It runs on your own hardware by default. Of — providers, — needs no credential at all; cloud sits in the same registry with its price per million tokens attached, and the tool surface follows the bill — a self‑hosted model is handed the composed arsenal loop, a cloud model the full one, because on a local box every schema token is paid in wall clock. The ambient UI reads the same run live across — stores: topology, cost and token usage as they accrue.

And it compounds. A run that took real work is distilled into a skill candidate, which cannot be promoted until — kinds of evidence are attached: a verification receipt, a source review, and a human approval. The agent proposes a capability; it never grants itself one.

— sibling bridges carry all of it to the benches, each with its own runnability probe, because a bridge that reports available while every call fails is worse than one that admits it is down. Composed, they close the loop in a single call: arsenal_assess scores a molecule on the mechanistic oracle and hands that payload — verbatim, in process — to the regulatory ledger. Design to body effect to dossier gap, with no human relaying numbers between two tools.