Artifacts per attempt Every step attempt — successful, failed, or retried — emits an artifact into the Evidence Ledger: the input, the output, the tool or model that produced it, the policy decisions made along the way, and the verification verdict. Nothing is summarized away at the moment it happens; the run's record is written as it executes, not reconstructed after.
This is what turns "what did the agent do?" from archaeology into a query.
Provenance tiers Every source cited in an evidence dossier carries a provenance tier — a classification of where the claim came from — plus the verbatim passage and its date. A claim without its verbatim passage and URL is not evidence; it is assertion, and the verification prompts reject it as such.
Independent corroboration Corroboration counts independent sources only. Two outlets republishing the same wire story are one source, not two. This rule lives inside the verification prompt of the deep-dive branch, verbatim:
"…strategic move records with their evidence passages, URLs, corroboration counts counting independent sources only, and verification flags (including 'unverified — single source' where applicable)…"
The counting rule matters because the failure mode it prevents is confident synthesis of one weak source — the exact way cascading hallucination launders itself into reports.
Verification flags Claims that survive with a single source are not discarded and not silently merged — they are flagged:
unverified — single source — one source, no independent corroboration Corroborated claims carry their count Competing interpretations get explicit blocks with evidence on both sides, rather than an averaged-away synthesis The report's reader — human or downstream step — always knows the epistemic status of every claim. The evidence dossier in the competitor deep-dive carries: verbatim passages, dates, provenance tiers, the pre-quarter baseline with its quotes, the change list, and the full source list with dates and tiers.
Replay Because every state transition is ledgered — inputs, outputs, policy decisions, verification verdicts, compensations — a failed run is replayable step by step. The audit question ("the run wrote a wrong record; prove what it saw and why it concluded that") is answered from the ledger, not from an incident postmortem.
Verification as the gate The evidence model is enforced at the step boundary, not the report boundary. Each step's verification prompt is the checkpoint — see the SOP reference for the full verification semantics. A step whose output fails verification never becomes the input to the next step; that is where cascading errors die.