Autopost Summary

How much of this company's document load the AI booked by itself (CompanyMember).

rate = auto_posted / codable_items, where the denominator is every archived item of a bookable document type and the numerator is the AI codings that posted with no accept audit at all. It is null, never 0, when there is nothing to divide by.

Read it beside autonomy_mode and min_confidence: a company whose autonomy is off reports 0 % as a matter of policy, not of AI quality.

The state counters (auto_posted / accepted / held / shadow / manual / skipped / expense_receipt / bank_attached) count decisions, not items, so a document the AI held and a human then coded by hand appears in two of them; only codable_items, items_without_decision and triage_parked count items. triage_parked is a disclosed confound inside the denominator: a human who marks an already-archived document unrelated stops the coder for ever, so those can never be posted. accepted_by_kind splits the sign-offs by actor kind, which — not accepted_by — is what tells an auto-post from a human approval; accepted_kind_unknown holds the accepts made before that column existed and is reported as unclassified rather than counted as either.

Why documents are NOT posted comes on three axes: by_cause classifies the hold reason, while by_risk_flag and by_coding_tier report the accountant's own structured fields verbatim — the guardrail collapses every blocking flag into a single reason and never names the resolver tier, so those two are the only way to see them. control_set is the held-then-accepted population grouped by its original hold cause: proposals a human reviewed and posted unchanged.

since/until are whole UTC days matched against the item's archive date, both inclusive; omit them for all time. An archived item with no archive timestamp cannot appear in a windowed answer.

Read-only. See doc/pipeline.md §"4.2 AI accountant autonomy".