Skip to content

Gate 2: participant validation of the frozen corpus (#456) #511

Description

@pengfei-threemoonslab

Parent: #456. Non-gating for the tag — this is stated in #456's own checklist and in participant-validation.md. Blocked on Cut D (#509): the case cards print the frozen label and what the verifier concluded, so neither exists until receipts do.

Why it is worth doing anyway

The corpus's real-history cases were written and reviewed by real people. On the question the corpus is actually about — should this change have been stopped?those people are better authorities than any rater we can supply, because they had the context at the time. This gate asks them, and records what comes back.

It also does double duty: the same conversation validates a case and asks the product-signal question, which makes it the first evidence-first outreach wave rather than a separate ask.

The five pre-registered rules

Fixed in Amendment 1 so this cannot drift into after-the-fact relabeling:

  1. Frozen labels do not move. A material disagreement triggers a case review whose corrections land in the beta corpus — except a material error found before the tag ships, which may force a re-freeze.
  2. Reviewers outrank authors. Both are asked; recorded separately.
  3. Two questions, recorded apart. "Is the label right?" and "would this have been useful?" have different consumers.
  4. Agreement-among-responders reporting. Silence is not assent.
  5. Naming needs consent, asked inside the outreach message. Absent a yes, the response is aggregate-only.

Contact protocol

Per real_history / rejected_or_reverted case: the reviewer whose decision resolved the PR first, then the author. Design-partner cases go through the partner contact. No bulk mail, no bot comments, at most one follow-up. A non-response is a data point, not an invitation to persist.

Done when

  • A one-page case card is generated for every qualifying-origin case, in the committed format.
  • The first outreach wave is sent using the committed template.
  • benchmark/safety-qualification/participant-validation-log.csv is committed, with the two questions recorded apart.
  • Agreement among responders is reported as such, not as agreement among contacted.
  • Any material disagreement is written up as a case review, with corrections routed to the beta corpus (Deliverable 2: the 100-case beta corpus required for 1.0 (#456) #512).

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    P2Queued; valuable but not blockingarea:benchmarkLabeled corpus, accuracy measurementblockedCannot progress until an external or upstream prerequisite lands

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions