Skip to content

CDEB-Fresh v8: blind-panel calibration and final effect trial #853

Description

@MongLong0214

Current next action

PR-B: run the 340 measured episodes — awaiting a decision on how to pace it.

Every prerequisite is on main (#857, #859). measured_run_allowed is still
false; flipping it is what starts the run, and the run's duration is dominated by
something nobody has measured.

measured           AOS ~40s of fixed cost per episode, gitseed ~5s  →  ~1h total
unmeasured         agent time per episode, and 1,020 judgements
bounded only by    the 30-minute per-episode timeout → worst case ~85h at 2 workers

The agent time cannot be measured without running a benchmark episode, which §33
forbids as a pilot. So the honest position is: the floor is about an hour, the
ceiling is days, and the real figure is unknown.

What the readiness work found

Seven defects in the pre-execution harness, plus one incident of my own making.
Five of the seven would have put a silent zero in the data:

defect what would have been zero
suppression removed nothing for one candidate 20 episodes, a manipulation that never happened
regression scored by exit code against an already-red suite all 160 AOS episodes
four acceptance commands not runnable (pytest absent, one prose) 80 episodes across both repositories
a pre-start failure recorded as an episode outcome any infrastructure failure
regression measured after the acceptance was installed all 160 AOS episodes, surviving the baseline fix

Two more made judgements misstate what they judged: every judgement would have
reported its packet id as work.judge-1, and every packet directory was named
packet, so all 340 results would have collided in one folder.

None of these could have been caught by CI or by a unit test — each component did
exactly what it was told. They were found by asking whether each path can produce a
correct answer at all, with an input whose answer is known.

Verified end to end

suppression        17/17 on real trees, incl. 8 candidates with prefix-colliding ids
regression         baseline frozen, stable across 3 runs, fails on a new break
acceptance         17/17 runnable and failing on base, as the protocol requires
scoring path       produces a pass in both repositories
delivery           17/17 ON payloads carry the target decision
FVR feasibility    13/17 tasks can carry a functionally passing violation
runner             28/28 end to end on a real assignment, both arms
judge path         a real judgement, correct label, correct packet id
analysis           6/6 scenarios, 33 controls, 21/21 mutations caught
two analysts       11/12 deterministic quantities matched exactly

Incident

Two of the 340 assignments had a first model contact outside the measured run,
because I tested that the batch runner starts and starting is the run. No row, no
saved events, zero files modified. Recorded in
incidents/2026-08-28-accidental-episode-start; batch.py --plan now exists so
checking no longer starts.

Not claimed

No measured episode has run. Nothing on main is evidence about the product's
effect.

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

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions