IDKMesh

Independent verification of the benchmark publication counts

Refs: issue 541. Date: 2026-09-24.

Source commit

Environment

Baseline checks (done)

Method for the independent count

For each of the four cohort files under benchmarks/*/cohort.json, I did the following by hand, reading the raw JSON (not benchmarks/publication.json or benchmarks/PUBLICATION.md, and not tools/benchmark_publication.py):

  1. Counted the number of task objects inside the top-level "tasks" array.
  2. For each task, read evidence.attempts and counted the entries (an empty array means zero attempts for that task).
  3. For each task, read the evidence.status field ("verified", "pending", or "excluded") and counted a task as verified only when status == "verified". This is the field the schema itself uses to mark verification; I did not infer verification from the presence of an outcome value.
  4. Compared each cohort’s definition_digest (when present) against the digest listed for that cohort in benchmarks/PUBLICATION.md, using grep.
  5. Summed tasks, attempts, and verified counts across the four cohorts.

Counts I derived

Cohort Tasks Verified Attempts definition_digest matches PUBLICATION.md
phase-b2-successor-five 5 5 5 yes
phase-b2-first-five-v2 5 0 0 yes
phase-b2-successor-v2 5 0 0 n/a — cohort is stage: scaffold; its description field states it intentionally has no definition_digest until it is frozen
phase-b2-first-five 5 0 0 yes (matches the predecessor_burned_definition_digest recorded in phase-b2-successor-five/cohort.json)
Total 20 5 5  

These totals match the published summary in benchmarks/PUBLICATION.md: “4 cohorts · 20 tasks · 5 with a verified outcome · 5 attempts”.

PYTHONPATH=. python tools/benchmark_publication.py --check reports pass (“benchmark publication is up to date”).

Discrepancies or ambiguities

None found. My independently derived counts match the published totals exactly, for all four cohorts.

Additional observations (not discrepancies in the counts)

AI assistance

I used an AI assistant (Claude) to help me understand the project’s documentation and JSON schema, translate terminology, and plan the verification method. The counting itself — reading each cohort.json, counting tasks/attempts, checking each status field, and comparing digests — was done by me, manually, cohort by cohort. No script or model performed the count.