The stages

A stage is one step of a workflow. A model-backed stage follows the same sequence every time: start on the run's provenance entity (skip if already done for these inputs), call the agent, decode the typed reply, apply the gate, write the accepted records, complete. A failed gate records the attempt with a classified defect; if the defect is retryable and the budget (two retries by default) is not spent, the step runs again, otherwise the stage is recorded as failed.

Ingestion and verification

Stage	Reads	Model	Produces	Check	On failure
fetch	an adapter and an identifier	no	the source's metadata and content	adapter errors classified transient or permanent	transient: two retries, then the failure is recorded and the workflow pauses
record	the fetched content	no	a source version, unless the hash is already a version	duplicate by content hash	a duplicate ends the workflow; nothing is added
parse	the version's content	no	sections with stable offsets, parse issues, completeness	corrupt or text-less content	the version is quarantined with its reason; the content is kept
chunk	the sections	no	chunks with embeddings, into the chunk view
measurements	internal rows	no	measurements with units	the unit table	a unit outside it is kept as written and marked invalid; nothing is converted
eligibility	the source and version	no	an eligibility judgement with reasons	retracted, superseded, restricted, quarantined, failed parse	the source is ineligible, which every later stage reads
extract	the sections, keyed for quotation	evidence agent	findings and claims, each with quotations	each quotation must be found in its section	retryable: up to two retries
validate	one proposed record	no	a validation result	every span resolves to its text; required fields present; no self-asserted status	quarantined with the defect; no model is called
verify	the record beside its resolved passages	evidence agent, separate prompt	a label and five checks: negation, scope, species, treatment, endpoint	a failed check forbids "entailed"	critical defects reject; ambiguous or thin goes to the review queue
retry	a record with a retryable defect	evidence agent	a new attempt re-extracted from the source	the same validation and verification	quarantined once the budget is spent

The investigation

Stage	Reads	Model	Produces	Check	On failure
extract facets	the request text	request agent	eight facets with origin and importance	an explicit facet cites a span of the request that resolves	retry up to two
round trip	the request and the facets	request agent, separate prompt	a reconstruction with omissions, additions, contradictions	none; the result is shown to the reviewer
await facet review		no	pause	every mandatory facet confirmed, assumed or left unresolved and flagged	the confirm command is refused
derive questions	the confirmed facets	request agent	questions linked to facets, with answer fields	every mandatory facet investigated or named as uninvestigated	coverage gap: retry up to two
plan	one question	search agent, small model	a search plan	at least one lexical and one semantic query; targets; stopping criterion	retry up to two
execute	the plan	no	one execution per query and target, and one fused execution, each with the snapshot searched		adapter transient failures: two retries
eligibility	the round's executions	no	eligible results and exclusions, after ranking	eligibility rules per source version	nothing eligible opens a coverage gap
supplemental	a coverage gap	no	another round with the plan's expansions, linked to the gap	the gap's budget, two rounds	refused once spent
resolve	the eligible hits	no	each chunk's context; the verified records of the hit sources
assess	the question and verified records	evidence agent	relationship and applicability per record	applicability rules: a species or endpoint mismatch is never direct; unknown species is unknown	retry up to two
sufficiency	the question, assessments, gaps, one decision	evidence agent	a status, assumptions, critical gaps, coverage judgement	a record that is the only affirmative evidence sends the judgement to review	retry up to two
build packages	each question's assessments	no	one package per question	spans resolve, contradictions included, only verified and eligible records, no critical trim	the package is invalid, and says why
candidates	the packages	design agent	at least two approaches and a comparison on eight criteria	every assumption labelled; every criterion rated with a rationale; no probability of success	fewer than two or unlabelled: retry; a probability: refused and paused
await decision		no	pause	a configured reviewer, a reason for each rejected approach	the decide command is refused
design	the decision, the approach, verified claims	design agent	a twelve-section design	every cited claim verified, every span resolves, mandatory facets accounted for	retry up to two
await approval		no	pause	a configured reviewer, a rationale	changes requested run the design stage again with the changes

What a reviewer sees

The review pages under /review show the same records the API returns. The facet page lists each facet with its origin, importance and history, and the round-trip check's findings. The queue shows each record in review beside the passage it cites and the verifier's checks. The candidate page shows the approaches side by side on the eight criteria. The design page shows the rendering, marked as a proposal until a version is approved, with the approval history.