agentii.challenge

The buyer-side IC verb: challenge the thesis. Not audit — that name was taken three times (audit-xls, quality-audit.yaml, G3) — challenge is the thing that happens in an investment committee.

Method bodies (the four angles)

1. Cross-run contradiction — the entity index (reduce_journals.build_entity_index): same (entity, metric, period) with different values. Same retrieved_at → true contradiction; different → suspected restatement. The validator never auto-resolves — candidates go to the IC agenda.
2. Pre-mortem — assume the loss already happened; reconstruct the path backwards. Which pillar failed first? What data would have to be wrong?
3. Inversion — what would make this thesis a sell? What would the bear's best case look like? What would falsify the conviction, not just the claim?
4. wrong_if falsifiability — is each falsifier mechanically checkable (metric + threshold + source)? A prose falsifier is not a falsifier (Q8-4).

Economics (Q40/Q68)

• Findings: ≈50 cap, severity-sorted; overflow aggregates by category.
• Finding IDs are content-derived (hash(entity+metric+period+gap_type)) — unchanged re-runs yield byte-identical IDs (eval corpus comparability).
• Scope is incremental (artifacts whose pins/as_of changed since last run) with three backstops: every Nth converge / constitution MINOR-MAJOR / subscription change (new subscriptions create new comparison pairs — Q9).
• Lifecycle hooks: pillar-complete (cheap, early) / thesis-complete (final line) / pre-reduction (claims entering the portfolio are adversarially verified). Unchallenged claims never enter the knowledge base (Q64).

Invocation

python3 scripts/challenge.py --nth-converge 10