Materials · corvid/triage.md

# Triage of Corvid review findings (orchestrator)

Judged against the record, Jordan's writing preferences, and how much each point matters to a reader. F = factual, U = unfamiliar reader, E = editorial. Numbers follow each review's own ranking.

| # | Finding | Decision | Reason |
|---|---|---|---|
| F1 | Letter implies the *Management Science* paper contains the judge validation | Apply | The record says only that the paper reports the Relay experiments. Reword to say exactly that. |
| U1 | Opening claim about Corvid has no source | Apply | It is the first thing the reader checks. Name the March 2026 engineering blog post. |
| U2 | Letter proposes measuring rater agreement but never says whether Jordan measured it | Apply as placeholder | Only Jordan can answer. Keep the placeholder and do not delete it. |
| U3 | The Relay story stops at the failure | Apply as placeholder | Same reason as U2. This is the most valuable missing fact in the letter. |
| U4 | Add agreement statistics | Set aside | Rule 3 keeps numbers out of letters, and the record has no statistics to give. |
| U5 | "Strong Python" is thin | Set aside | The record says only that Python was used. Anything more would be invented. |
| U6 | Analogy assumes the graders are wrong, which contradicts the rest of the letter | Apply | This is a real logic flaw, and Corvid's readers would catch it. Recast so the open question is which reference to trust. |
| U7 | "Model-access route" and "hidden context" are unexplained | Apply | Say it in plain words using only what the record states. The record does not say how the error was found, so the letter does not either. |
| U8 | Human-subjects evidence is old and indirect | Set aside | The letter states it accurately. The record has nothing newer, and padding it would overstate. |
| U9 | "Before a client ran it" sits next to "no employer" | Partly apply | Drop "no employer" from the résumé, since the title already conveys it. Ask Jordan how the client work relates to self-funding. |
| U10 | Kestrel credit unclear (Jordan led a team of three) | Set aside | The record attributes the finding and the rebuild to Jordan. The résumé shows the team. |
| U11 | "Code," "qualification," and "judges" may be misread | Apply in part | Change "code" to "label" and gloss qualifications as hedges and caveats. Leave the judge models unnamed because the record does not name them. |
| U12 | PhD may not read as quantitative. Parcel, the testbed link, and the talk content are not connected | Set aside | The posting allows equivalent experience. The brief chose not to lead with Parcel. Calling Relay an evaluation environment would claim an equivalence the record does not support. The record does not give the talk's content. |
| E1 | Kestrel paragraph restates the résumé and is not tied to Corvid | Apply | Rule 10. Tie it to the head of research's "measuring the wrong thing" and stop repeating the résumé wording. |
| E2 | Hidden-context sentence copies the résumé word for word | Apply | Rule 10. Rewrite it together with U7. |
| E3 | Greeting says research team, but rule 10 says hiring team | Apply | Jordan's rule controls. My prompt caused the error. |
| E4 | Reinforcement-learning sentence may raise a gap without cause | Set aside | The posting lists reinforcement-learning experience as preferred, so readers will notice it. Rule 5 says to state it once. |
| E6 | Talk venue differs between the letter and the résumé | Apply | Cheap fix, and the two documents should match. |
| E7 | Months, AoM month, unused Helvetica resource | Months go to Jordan. Set aside the rest | The record has no AoM month. Helvetica draws no text, so this is not a rule breach. |
| E8 | Letter is 83% full before the placeholders are answered | Apply | Tighten so that Jordan's answers fit on one page. |
| Own | "Under review" lacks "first round" in the letter | Apply | Rule 8 calls for exact status. |