| name | provenance |
| description | Audit the tamper-evident decision chain behind yeaboi's signals: every practice nudge, blocker flag, confidence adjustment, conflict card, and performance prep/review is recorded with its evidence in a hash-chained local log. Use when the user asks why yeaboi said something, whether a signal's history has been tampered with, for an audit/compliance trail of automated decisions, or to trace the evidence behind a specific decision. |
Provenance audit workflows with yeaboi
Every deterministic signal yeaboi surfaces is recorded as a hash-chained
decision with the evidence it rested on. The chain is append-only: retracting a
decision appends a tombstone record, and deleting or editing a stored row is
detectable, never hidable. Everything is deterministic and local — no LLM is
involved anywhere in this pipeline.
-
Verify and summarise with provenance_audit. The default window is 30
days (window_days, 1–365). Lead with chain_valid: true means every
record's checksum, link, and sequence number verified across the WHOLE
chain, not just the window. If it is false, breaks lists each failure —
checksum_mismatch means a record was edited in place, chain_break means
one was deleted or renumbered, and truncated_tail means the newest
records were removed (the walk fell short of the head anchor every append
updates in the same transaction). Say so plainly; that is the finding. One
honest caveat if asked: the chain is a local file with no external anchor,
so an adversary who rewrites the rows and the anchor together is beyond
what it can prove — it makes tampering visible, not impossible.
-
Read what was decided: recent lists the window's decisions newest
first (capped at 50 — a warning names the overflow), and records_by_type
counts the whole chain by kind (practice-signal, blocker-signal,
confidence, adjudication-drop, conflict, one-on-one-prep,
one-on-one, six-month-review).
-
Trace a "why" with provenance_trace and an entity_id from the audit
(e.g. standup:2026-08-16:practice:wip-sprawl:alice). The trail includes
the entity's own records — retractions included — plus the latest record
behind each piece of evidence it used, up to depth hops (default 2).
-
An empty chain is normal on a fresh install: decisions start recording
when a standup or performance workflow runs. Relay the warning rather than
treating it as an error.
-
Privacy. Records carry decision ids, rule/model names, evidence keys
and counts — never transcript content, ticket bodies, or 1:1 notes. If the
user asks what is IN the chain about a person, the honest answer is:
which rules fired, when, and on what evidence keys.
Error handling
Every tool returns {ok, llm_mode, warnings, data}. If ok is false, relay
error.message and its hint; don't retry blindly. provenance_trace on an
unknown entity returns found: false with a hint to list ids via
provenance_audit — offer that instead of guessing ids.
