Skip to content

REX-12: B2 loop: 27B teacher logits (gx10, top-k, m=1) over train pool; B2a/b/c acceptance tests wired to qwen35 finetune/distill/merge #4367

Closed as not planned

Description

@noahgift

owner: rex
Epic: #4354 (REX-001). Spec: docs/specifications/review-experiment-protocol.md §7 row REX-12.

Done when (all must hold): teacher dataset receipted (sha, count, 0 test hashes); B2 rows read NotRun{VerbRefused} until verbs land, then run §5.4 automatically

Binding: §6 R-1..R-12, §8 STOPs. Ticket id = PMAT-.

Activity

  1. added this to the 0.71.0 milestone on Sep 25, 2026
  2. noahgift commented on Sep 25, 2026

    @noahgift
    MemberAuthor

    REX-12 gate landed on rex/001 @ da7b638 (to be folded into B3 as the cop directs).

    • rex b2 status writes one review-b2-loop-v1 row per B2 tier (B2a distill, B2b finetune, B2c merge, B2d quantize) from evidence/verbs/refusals.json (refusal-receipt-v1). A tier is ready only with an acceptance receipt; a ledger row gives verb_refused{removed_by}; with neither the row is unledgered (NotRun). An absent refusal is not acceptance. Exit 11 while any row is NotRun.
    • rex b2 teacher-receipt records sha, count, k, m=1, sealed-index size and 0 test hits. The whole file is refused on a test item, an unknown or duplicate id, wrong k, a NaN, a sealed diff or diff sha in any string field, or an empty sealed index.
    • Contract review-b2-loop-v1 (FALSIFY-RXB-001..003) passes pv validate; 4/4 planted mutants turn RED.
    • At HEAD all four rows read unledgered. The refusal ledger carries only bench, so qwen35 distill/finetune/merge have no measured refusal row (noted on 0.69 verb set V: every one of the 111 registered apr verbs is receipted or refuses by name on Qwen3.5 #3597; the verbs are EPIC EPIC E8 0.75: CRUX Fine-Tune/Distill #4002, 0.75.0). The 27B teacher run is NotRun{NoDeclaredExecutor} until infra#1088 declares a gx10 executor, so the teacher-dataset acceptance is open.
  3. noahgift commented on Oct 4, 2026

    @noahgift
    MemberAuthor

    Folded into #4354 as a checklist line (ISSUE-TREE-001: a family of rows becomes one parent with checklist lines). Work and status are tracked there. aprender-wtix-even

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

    P2Medium prioritysize:MK̂ 3-4 h

    Type

    No type

    Projects

    No projects

      Milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions