Skip to content

MIT-618: ledger lines for the 2026-11-11 GitHub-task runs - #175

Merged
ZachFara merged 2 commits into
feature/eval-setupfrom
chore/MIT-618-pilot-ledger
Nov 12, 2026
Merged

ZachFara merged 2 commits into
feature/eval-setupfrom
chore/MIT-618-pilot-ledger

Conversation

@ZachFara

@ZachFara ZachFara commented Nov 11, 2026 •

Copy link
Copy Markdown
Owner

What changed

Adds 48 lines to docs/eval/trials.jsonl, one per eval session run on 2026-11-11. No code changes.

These sessions started before the leak scan (MIT-678, #174) merged, so the harness wrote their lines without flags. The second commit adds flags to each line from the merged leaks.scan over its saved transcript, in the position append_ledger now writes it. Every list is empty.

Condition Sessions What it was
github-pilot 4 pip-tools, all four arms, 1 run each, fresh empty databases
github-same10-seed 10 Commons arm on all 10 tasks, seeding commons_eval_store_same10
github-haiku-none-check 10 Haiku, none arm, all 10 tasks (side check, outside the decision record)
github-same10 24 pip-tools and table-extraction, 4 arms × 3 runs, on the seeded store with 2 approved entries

These conditions use their own names, not the record's github-arms*, so they never enter its verdict.

Why

Ticket: MIT-618. The ledger is the committed trial log. These sessions are the pilot and a same-task best-case run for the 10 tasks.

How it was tested

  • agent_eval.py flags over all four transcript directories reported no flagged session, and the same scan produced each line's flags field.
  • Against feature/eval-setup the diff only adds lines (48 additions, 0 removals).
  • I read all 619 tool calls by hand. None reached the network, GitHub or the repo cache.
  • The pre-commit hook's tools/check.py passed.

Learned

  • server/tests/test_hosted_hooks.py::test_a_session_end_to_end_over_http failed once in the full pre-commit run (outcome unknown, expected tests_passed). It passed 5 of 5 times on its own and on the retry. It looks flaky and needs its own ticket.

  • All 10 tasks pass with no help, on both Sonnet and Haiku.

  • Commons sessions search only when they get stuck. pip-tools had an approved entry, and 0 of 6 sessions searched for it. Table-extraction: 6 of 6 searched and found it at rank 1.

  • 4 of the 6 table-extraction sessions copied the entry's workaround fix rather than the upstream approach. Whatever fix an approved entry carries spreads.

  • One Haiku pip-tools session spent most of its 31 turns waiting on a background pytest run (Monitor and sleep), so its token count overstates how hard the task was.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qc1CQTb8r67J32HpkhS43y

ZachFara and others added 2 commits November 11, 2026 17:16
48 sessions for MIT-618: github-pilot (4), github-same10-seed (10),
github-haiku-none-check (10) and github-same10 (24). None is leak-flagged.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Qc1CQTb8r67J32HpkhS43y
These sessions ran before the leak scan merged, so their lines had no
flags field. Each line now carries the flags the merged scan finds in its
saved transcript (none).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Qc1CQTb8r67J32HpkhS43y
@ZachFara
ZachFara merged commit 9e362b8 into feature/eval-setup Nov 12, 2026
2 checks passed
@ZachFara
ZachFara deleted the chore/MIT-618-pilot-ledger branch November 12, 2026 02:09
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant