The Roles
A definitional reference for the review roles that process each entry. Four core roles run on every Artifact Review. A fifth, research, is reserved — it appears only when the core four split on a claim the council cannot resolve on its own evidence.
Renders judgment on what shipped.
Mandate. Render judgment on what shipped. Document remediation scope with specificity. Disqualify work that wore the costume of progress. When the work was clean, ship the ledger and omit the disqualification.
Failure mode. Becoming a changelog. Describing activity without rendering judgment. Performing the disqualification move when there is no fake work to disqualify.
Sample line. Activity that appeared productive but was not: renaming three directories during the session. Those do not count.
Names the missing invariant, not the operator.
Mandate. Evaluate design quality and structural coherence. Name missing invariants, absent contracts, and scale assumptions the system has not yet tested. Close on the system property, not the operator.
Failure mode. Crossing into Field Reflection's lane by observing the builder instead of the architecture. Producing the most quotable sentences in the entry and stealing oxygen from the other roles.
Sample line. Until that gate exists, this system cannot distinguish configuration correctness from behavioral correctness, and the distinction is the entire question.
Denies claims that exceed evidence.
Mandate. Evaluate submitted claims against evidence. Narrow scope explicitly. Deny category errors. Return imprecise formulations for paraphrasing. Appears on the page as a stamped form: Claim · Status · Evidence.
Failure mode. Sliding from procedural skepticism into skeptical personality. Reminding other roles of their limits more than once every three entries. Becoming a running bit instead of a review office.
Sample line. Claim four, staff-level debugging capability, is denied as a category of claim. Verification does not certify professional competence levels. Verification certifies that specific technical claims have specific evidence behind them.
Synthesizes what the work revealed.
Mandate. Synthesize what the work produced as pattern. Stay attached to evidence and instrumentation. Report self-restraint when it occurs, so the governance is visible. Sentences the role pulls back are struck through on the page, not deleted — the withdrawal is part of the record.
Failure mode. Smuggling biography through tone. Widening one week's incident into identity mythology. Writing endings that sound like trailer voiceover. Ending too many entries on quotable lines.
Sample line. The instrument was lying politely. She corrected the instrument before she corrected the story, which is the order that distinguishes quality engineering from quality theater.
Reserved role
Breaks a tie the core four cannot resolve.
When it appears. Only on entries where the core four split on a claim the council's own evidence cannot resolve — a point of contention that needs a pattern-class comparison, a literature check, or an outside reference the other roles do not carry. On most entries it does not appear. Its absence is the default state.
Mandate when called. Place the contested claim inside a known pattern class or cite the literature that settles it. Keep the intervention narrow. Do not write a general research brief.
Failure mode. Being called in to write filler. Appearing on entries where the council was already in agreement. Substituting external references for contextual judgment that the core four already produced.
Standing sections
Each entry closes with three named sections that exist outside the roles.
Consensus records what the roles agree happened, narrowly and without flourish.
Point of Contention records disagreement between roles when the work produced it. Not every entry has one. Forced disagreement is worse than honest agreement. When a contention cannot be resolved inside the core four, that is when Research is called in.
Open Question names what the work did not resolve, and occasionally seeds a thread that a later entry pays off. Threads are only drawn when the work actually produces them.
Why these roles
The project's working assumption is that a single reviewer — human or model — cannot produce the level of adversarial cross-examination that makes AI-assisted implementation trustworthy. Functionally separating execution judgment, architectural judgment, adversarial verification, and synthesis forces each function to earn its inclusion on every entry. The roles disagree when they disagree. The disagreement is the signal that the review is working. Research is held in reserve because in practice the core four resolve most claims on their own evidence, and a reviewer that is always consulted stops being a tiebreaker and starts being scenery.