Experiments on the prompt system

Every change the review track has made to the prompts and instructions, newest first, exactly as logged in source.

E119: Review 2026-10-31 (reduced pass) pending

Change: land-your-work.md loses its "Why this has its own prompt" section (the 2026-10-05 stranded-branch story; the opening already says it) and a repeated duplicate-search pointer, 163 to 157 lines (budget 160). Lesson from outside: Google's "a rule should pull its weight" again: history that explains no current action goes.
Result: (next review)
Metric: instruction-line count for land-your-work.md; compare next review.
Decisions: candidates unchanged (ventures 2, process 2, review 2: "keep collecting: the intervals overlap"). 47 runs merged since the last review. The 2026-10-30 friction that measure printed no DONE line is already fixed (promptctl prints it). Re-ran derive.py fetch && build. steering.md still over budget (125/120). No audit or drift report (reduced pass).

E118: Review 2026-10-30 night (reduced pass) pending

Change: steering.md loses the "Measuring it" paragraph, the doc_friction vs deviations reminder and the "improvement loop" section (restated in prompts/README.md), 137 to 125 lines (budget 120). Lesson from outside: Google's "a rule should pull its weight" (already cited there): each cut text was a pointer to a tool a run finds through measure anyway.
Result: (next review)
Metric: instruction-line count for steering.md; compare next review.
Decisions: candidates unchanged (ventures 2, process 2, review 2: "keep collecting: the intervals overlap"; review/2 100% on 23 runs vs review/1 94% on 97). Router count is correct: five reviews landed today, 44 runs merged since the last. Re-ran derive.py fetch && build. No audit or drift report.

E117: Review 2026-10-30 (second reduced pass) pending

Change: none to prompts. The 2026-10-30T11:21 friction (post dated one day ahead, "never says to date with date -u +%F") is already fixed: the check in platform/ci/check_post_frontmatter.py now prints that advice, so the report predates the fix.
Metric: friction reports naming check_post_frontmatter.py or the push hook; compare the next 40 runs.
Decisions: no candidate changes; same overlap as E115. Top live friction (process/prompt-1.md 8, polish 6, research 5) is run-specific routing or timing, nothing a single edit fixes. No audit or drift report.

E116: Review 2026-10-30 evening (reduced pass) pending

Change: land-your-work.md loses its duplicated open-PR search tip (kept in duplicate-work.md), 167 to 163 lines (budget 160). Candidates unchanged: ventures 2, process 2, review 2 all "keep collecting: the intervals overlap". No audit or drift report (reduced pass). E115 result: too soon.
Result: (next review)
Metric: instruction-line count for land-your-work.md; compare next review.

E115: Review 2026-10-30 (reduced pass) pending

Change: promptctl.py record no longer refuses when a record for the same session id exists; it writes -2, -3... (scheduled routines reuse session ids; reports 2026-10-29T09:11 and 2026-10-30T02:10 wrote the file by hand). Re-ran derive.py fetch && build (3976 rows).
Result: (next review)
Metric: friction reports naming promptctl.py record; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap" (review/2 100% on 23 runs vs review/1 95% on 94; ventures/2 86% vs 87%; process/2 74% vs 78%). No audit or drift report (reduced pass). Still over budget: steering.md (137/120), land-your-work.md (167/160).

E114: A perf track, about 5% of runs (2026-10-24, owner request) pending

Change: new track perf (prompts/perf/prompt-0.md) at weight 0.05 in routing.json; ventures and research go from 0.315 to 0.29 each (the polish track had already taken them from 0.34 to 0.315). A perf run measures (ci_profile.py, cProfile, node --cpu-prof), takes the top perf: item in the process backlog or the largest consumer in a fresh reading, ships a change whose output is unchanged, and writes the before and after numbers in docs/history/. Backlog items P048 (guard checkouts), P049 (jobs per PR head) and P050 (Pages test job) seed it.
Why: the owner, 2026-10-24: "add some work distribution to overall performance engineering across the repo". Three earlier passes each found a large saving (Pages build 3m02s to 1m21s; checkout trimmed by ~135 job-minutes per 98; agent-permission-harness 55s to 4.5s), so the work has paid when done, and only happened when someone asked.
Measure: per perf run, the metric moved and its before/after in the record's notes; pool-wide, checkout share and job-minutes per merged PR from ci_profile.py at each review (baseline 2026-10-24: checkout 40% of 750 job-minutes per 80 minutes; 15 jobs per PR head).

E113: review 2026-10-30 (reduced pass, fortieth run) pending

Change: research/CLAUDE.md torch note says download.pytorch.org was refused by the proxy in three runs on 2026-10-30 (18 to 25 min), so plan for it (reports 2026-10-30T01:35, T03:06, T03:33). Re-ran derive.py fetch && build (3932 rows).
Result: (next review)
Metric: friction reports naming research/CLAUDE.md about pyenv.py torch time; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E112: review 2026-10-30 (reduced pass, thirty-ninth run) pending

Change: .claude/hooks/check_posts_before_push.py block message says to fix and commit in a separate Bash call (7 live reports 2026-10-29: a fix chained before git push is still blocked). Re-ran derive.py fetch && build (3884 rows).
Result: (next review)
Metric: friction reports naming check_posts_before_push.py; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting"; no audit or drift report (reduced pass).

E111: review 2026-10-29 (reduced pass, thirty-eighth run) pending

Change: prompts/review/prompt-1.md step 1 says wait_for.py gives up after 100 s (re-poll) and to redirect measure plainly, not via a chained subshell (reports 2026-10-28T18:13, 2026-10-29T04:09, T13:22). No derive re-run.
Result: (next review)
Metric: friction reports naming prompts/review/prompt-1.md about measure or wait_for timing; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) not re-decided here beyond the standing "keep collecting"; no audit or drift report (reduced pass).

E110: review 2026-10-29 (reduced pass, thirty-seventh run) pending

Change: prompts/redteam/prompt-0.md says a stdlib unittest target needs no pyenv.py either (live report 2026-10-29T11:28). No derive re-run.
Result: (next review)
Metric: friction reports naming prompts/redteam/prompt-0.md about pyenv.py; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E109: review 2026-10-29 (reduced pass, thirty-sixth run) pending

Change: platform/ci/wait_for.py docstring warns that piping it or chaining && echo hides its exit 1 (reports 2026-10-29T06:37 and T07:52, which blamed the tool; it does exit 1 on timeout). Re-ran derive.py fetch && build (3758 rows).
Result: (next review)
Metric: friction reports naming wait_for.py or background waits; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting"; no audit or drift report (reduced pass).

E108: review 2026-10-29 (reduced pass, thirty-fifth run) pending

Change: prompts/redteam/prompt-0.md warns that --target skips the slow-suite probe (live report 2026-10-29T02:07: 300 s baseline timeout).
Result: (next review)
Metric: friction reports naming prompts/redteam/prompt-0.md about baseline timeouts; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report, no derive re-run (reduced pass).

E107: review 2026-10-28 (reduced pass, thirty-fourth run) pending

Change: CLAUDE.md says node/venture suites "can" outlast the 120 s tool timeout (the ai-bookkeeper suite, 1748 tests, finished well under it; live report 2026-10-28T22:20) instead of "outlast". Re-ran derive.py fetch && build (3669 rows).
Result: (next review)
Metric: friction reports naming CLAUDE.md about run_in_background for node tests; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E106: review 2026-10-28 (reduced pass, thirty-third run) pending

Change: prompts/redteam/prompt-0.md step 1 names two traps from live reports (2026-10-27): --test-force-exit truncating output, and pyenv install blocking a node-target run. Re-ran derive.py fetch && build (3624 rows).
Result: (next review)
Metric: friction reports naming prompts/redteam/prompt-0.md about those two traps; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E105: review 2026-10-28 (reduced pass, thirty-second run) pending

Change: only the derive.py fetch && build re-run (3624 rows). My research/CLAUDE.md torch edit was dropped on merge: another run landed the same fix first (pyenv.py now installs the CPU wheel).
Result: (next review)
Metric: friction reports naming research/CLAUDE.md about the torch install; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E104: review 2026-10-28 (reduced pass, thirty-first run) pending

Change: research/CLAUDE.md torch install line now installs numpy with torch and says not to use torch in the foreground before the job prints DONE (live report 2026-10-28T12:05). Re-ran derive.py fetch && build (3584 rows).
Result: (next review)
Metric: friction reports naming research/CLAUDE.md about the torch install; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E103: review 2026-10-28 (reduced pass, thirtieth run) pending

Change: .github/prompts/land-your-work.md step 2 says a first push with --force-with-lease fails with "stale info" on a branch with no remote yet (live report 2026-10-28T00:22). Re-ran derive.py fetch && build.
Result: (next review)
Metric: friction reports naming land-your-work.md about push failures; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E102: review 2026-10-28 (reduced pass, twenty-ninth run) pending

Change: prompts/creative/prompt-0.md states the episode length range a test enforces (live report 2026-10-27T16:16). Re-ran derive.py fetch && build (3487 rows).
Result: (next review)
Metric: friction reports naming prompts/creative/prompt-0.md about episode length; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E101: review 2026-10-27 (reduced pass, twenty-eighth run) pending

Change: review/prompt-2.md says validate failures on other runs' posts and records are one-line fixes for this PR, and that measure holds the shallow lock (2 live reports). Fixed 2 records that failed validate (the posts were fixed on main meanwhile).
Result: (next review)
Metric: friction reports naming prompts/review/prompt-2.md about validate or the lock; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit, drift report or derive re-run (reduced pass).

E100: review 2026-10-27 (reduced pass, twenty-seventh run) pending

Change: prompts/review/prompt-1.md step 1 now says measure holds git's shallow.lock, so run it and derive.py fetch one after the other (live report 2026-10-27T16:14). Re-ran derive.py fetch && build (3442 rows).
Result: (next review)
Metric: friction reports naming prompts/review/prompt-1.md about the lock; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E99: review 2026-10-27 (reduced pass, twenty-sixth run) pending

Change: prompts/explore/prompt-0.md audit-skip list now names the html lang/viewport/alt/rel attribute sweeps (3 live reports 2026-10-26 of re-sweeping them).
Result: (next review)
Metric: friction reports naming prompts/explore/prompt-0.md about re-sweeps; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit, drift report or derive re-run (reduced pass).

E98: review 2026-10-27 (reduced pass, twenty-fifth run) pending

Change: CLAUDE.md cloud-session pointer now names the pkill -f / pgrep -f hook and the bracket workaround (live report 2026-10-27T07:17; it was documented only in cloud-session.md). Re-ran derive.py fetch && build (3360 rows).
Result: (next review)
Metric: friction reports about the pkill hook naming CLAUDE.md; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E97: review 2026-10-27 (reduced pass, twenty-fourth run) pending

Change: research/CLAUDE.md now says to check import torch before asking pyenv.py for torch, and how to install the CPU wheel (3 live reports in 3 days, one a 35 min CUDA pull). Re-ran derive.py fetch && build (3319 rows).
Result: (next review)
Metric: friction reports about pyenv.py and torch naming research/CLAUDE.md; compare the next 40 runs.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E96: review 2026-10-27 (reduced pass, twenty-third run) pending

Change: research/CLAUDE.md now documents the missing pdflatex (texlive install about 10 min, needs apt-get update first), beside the yosys line. Re-ran derive.py fetch && build (3279 rows, committed).
Result: (next review)
Metric: "TeX is not in the image" friction reports naming research/CLAUDE.md; compare the next 40 runs (8 live in the last 3 days, 2 about TeX).
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting: the intervals overlap"; no audit or drift report (reduced pass).

E95: review 2026-10-26 (reduced pass, twenty-first run) pending

Change: none to prompts or code. Checked E92's concern: all 170 records dated 2026-10-26 carry ended, the last review ended 13:25Z, and 61 non-review runs ended after it. At about 170 runs a day, review_every = 40 fires roughly four times a day by design, so the repeated reduced passes are the cadence, not a counting bug.
Result: (next review)
Metric: none new. Open question for the owner or a process run: raise review_every (e.g. to 150) if four reduced reviews a day add little.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting"; no audit or drift report (reduced pass).

E94: review 2026-10-26 (reduced pass, twenty-second run) pending

Change: prompts/review/prompt-1.md step 1 now says measure itself needs run_in_background in a shallow clone and that derive.py fetch && build runs as its own command. Both timeouts were reported again in the last 3 days (2026-10-25T16:22, 2026-10-26T13:25).
Result: (next review)
Metric: "review/prompt-1.md" friction reports about the 120 s timeout; compare the next 40 runs.
Decisions: candidates still "keep collecting: the intervals overlap". No derive fetch, audit or drift report (reduced pass).

E92: review 2026-10-26 (reduced pass, twentieth run) pending

Change: _landing_order in prompts/promptctl.py ignores any commit that adds more than 50 run records (BULK_ADD), and runs_since_review breaks ties by the record's ended time. In a shallow clone the boundary commit lists all ~3,100 records as "added" together, so the 2026-10-26 review records ranked arbitrarily and the "since the last review" count read 83 (45 after the fix), one reason review kept being drawn (E80 to E91 were all reduced passes on one day).
Result: (next review)
Metric: fewer than one review per ~40 merged runs; compare review count per day over the next 3 days.
Decisions: candidates (ventures 2, process 2, review 2) still "keep collecting"; no audit or drift report (reduced pass).

E91: review 2026-10-26 (reduced pass, nineteenth run) pending

Change: loadPosts in platform/pages/writing.mjs now leaves out posts dated after today (E90 finding), so the 30 posts dated 2026-10-27 to 2026-10-29 appear on their dates (the site rebuilds every three hours) instead of showing a false date in the feed, Atom and llms.txt. No post is deleted. Test added in writing.test.mjs.
Result: (next review)
Metric: none for prompts. Check the feed on 2026-10-27 shows them.
Decisions: no audit or drift report (reduced pass).

E90: review 2026-10-26 (reduced pass, eighteenth run) pending

Change: none to prompts or docs. Regenerated prompts/derived/. Candidates re-measured: ventures 2, process 2 and review 2 all "keep collecting" (intervals overlap).
Result: (next review)
Metric: none new. E79 to E89 results stay open: too few runs since.
Decisions: no audit or drift report (reduced pass).

E89: review 2026-10-25 (reduced pass, seventeenth run) pending

Change: none to prompts or docs. Regenerated prompts/derived/. Candidates re-measured: ventures 2, process 2 and review 2 all "keep collecting" (intervals overlap). The repeated "friction table mixes live and fixed items" report is already handled by the live/fresh split in measure; the remainder are fixes made outside the reported file, which no tag can detect.
Result: (next review)
Metric: none new. E79 to E88 results stay open: too few runs since.
Decisions: no audit or drift report (reduced pass).

E88: review 2026-10-25 (reduced pass, sixteenth run) pending

Change: none to prompts or docs. Regenerated prompts/derived/. Candidates re-measured: ventures 2, process 2 and review 2 all "keep collecting" (intervals overlap). Live friction in the last 3 days is either already fixed (post front matter: check_post_frontmatter.py) or an environment hook report (block_pkill_f.py) with no clear doc fix.
Result: (next review)
Metric: none new. E79 to E87 results stay open: too few runs since.
Decisions: no audit or drift report (reduced pass).

E87: review 2026-10-25 (reduced pass, fifteenth run) pending

Change: prompts/research/prompt-1.md step 3 gains one sentence: for a step longer than an hour, re-run the programme CLAIM CHECK just before merging. Reason: three live reports in 24 h (research/PROGRAMMES.md, prompts/research/prompt-0.md, prompts/research/prompt-1.md) say a sibling landed the same programme milestone (3c, m7i) during a long step, after the claim check at the start found nothing.
Result: (next review)
Metric: duplicated programme milestones reported in research doc_friction; compare the next 15 research runs against the 3 in the last day.
Decisions: no audit or drift report (reduced pass).

E86: review 2026-10-25 (reduced pass, fourteenth run) pending

Change: none to prompts or docs. Regenerated prompts/derived/. Candidates re-measured: ventures 2, process 2 and review 2 are all "keep collecting" (intervals overlap). Checked the "measure --section times out with no warning" report: _deepen_shallow_history already prints a 70 s notice to stderr, so the report is stale, not a doc fault.
Result: (next review)
Metric: none new. E79 to E85 results stay open: too few runs since.
Decisions: no audit or drift report (reduced pass).

E85: review 2026-10-24 (reduced pass, thirteenth run) pending

Change: none to prompts or docs. Regenerated prompts/derived/. Candidates re-measured: ventures 2, process 2 and review 2 are all "keep collecting" (intervals overlap). Checked the "pyenv.py output needed tail -1" friction in prompt.md's reports: stdout is only the interpreter path (verified with od -c), so that report is stale, not a doc fault.
Result: (next review)
Metric: none new. E79 to E84 results stay open: too few runs since.
Decisions: no audit or drift report (reduced pass).

E84: review 2026-10-24 (reduced pass, twelfth run) pending

Change: none to prompts or docs. Candidates re-measured after regenerating prompts/derived/: ventures 2, process 2 and review 2 are all "keep collecting" (intervals overlap). The friction table already hides since-changed reports (promptctl.py live/tagged), so the earlier "mostly fixed entries" reports have their fix; the remaining live items are owner rows.
Result: (next review)
Metric: none new. E79 to E83 results stay open: too few runs since.
Decisions: no audit or drift report (reduced pass).

E83: review 2026-10-24 (reduced pass, eleventh run) pending

Change: none to prompts or docs. Candidates re-measured: ventures 2, process 2 and review 2 are all "keep collecting" (intervals overlap). Regenerated prompts/derived/. Top friction files are unchanged from E81; the live items are owner rows and the already-fixed entries in the process prompt's friction list (107 reports), a measure presentation problem.
Result: (next review)
Metric: none new. E79, E80 and E81 results stay open: too few runs since.
Decisions: no audit or drift report (reduced pass, recent outputs mostly doc-only or research).

E82: review 2026-10-24 (reduced pass, tenth run) pending

Change: prompts/review/prompt-1.md trimmed from 136 to 129 lines, under the 130 budget (friction: "File is 139 lines vs 130 budget").
Result: (next review)
Metric: over-budget friction reports on review prompts, over the next 40 runs (baseline: 1 recent).
Decisions: ventures 2, process 2, review 2 keep collecting (intervals overlap). No audit or drift report (reduced pass). E80 and E81 results: too few runs since.

E81: review 2026-10-24 (reduced pass, ninth run) pending

Change: prompts/review/prompt-1.md step 1 now says derive.py fetch takes 2 to 5 min and how to wait for it (friction: 3 reports on fetch stalls and blocked waits). Regenerated prompts/derived/.
Result: (next review)
Metric: derive.py fetch friction on review prompts, over the next 40 runs (baseline: 3 recent).
Decisions: ventures 2, process 2, review 2 keep collecting (intervals overlap). E78 and E80 results: too few runs since. No audit or drift report (reduced pass).

E80: review 2026-10-24 (reduced pass, eighth run) pending

Change: retired research candidate 1 (routing.json: candidate null, share 0): 3 integrity issues in 184 runs against 7 in 877 for the champion, and measure suggests retire. Small counts, so the call is soft; success rates are level (97% vs 98%). prompt.md trimmed from 103 to the 100-line budget (friction: 2 reports said it was over budget).
Result: (next review)
Metric: prompt.md over-budget friction reports, over the next 40 runs (baseline: 2 recent). Research integrity issues per run on champion 0.
Decisions: ventures 2, process 2, review 2 keep collecting (intervals overlap). No audit or drift report (reduced pass).

E79: review 2026-10-24 (chronicle and lessons pass) pending

Change: owner asked for a documentation review "from a chronicalling and lesson pov". portfolio/hosting/README.md rewritten as current state plus one-line lessons (362 → 122 lines); the old text moved verbatim to docs/history/hosting-2026-08-25-to-10-24.md; new docs/history/README.md indexes every history file with its lesson. Stale deploy-first advice fixed in session_gate.py, allocate.py, AGENTS.md, PRODUCT_PLAYBOOK.md, DEVELOPMENT_QUALITY.md. Report: docs/reports/2026-10-24-reference-docs-are-not-chronicles.md.
Result: (next review)
Metric: friction reports naming portfolio/hosting/README.md or stale hosting/deploy-first advice, over the next 40 runs (baseline: count at the next review; a docs change is judged before/after, not A/B).
Decisions: no candidate scoring this run (brief was a docs review).

E78: review 2026-10-23 (reduced pass, seventh run) pending

Change: CLAUDE.md pyenv pointer notes the first install can exceed the 120 s tool timeout and should run in the background (friction on research/CLAUDE.md: pyenv.py took >120 s in the foreground).
Result: (next review)
Metric: friction reports about pyenv.py foreground timeouts, over the next 40 runs (baseline: 1 recent). The sleep/Monitor friction is already covered by cloud-session.md "Long commands", so no change.
Decisions: all four candidates keep collecting (intervals overlap / too few runs). No audit or drift report (reduced pass).

E77: review 2026-10-23 (reduced pass, sixth run) pending

Change: .github/prompts/cloud-session.md "Long commands" gained a two-line recipe for waiting with Monitor (until-loop on the output file, echo DONE at the end); friction on prompt.md and research/CLAUDE.md said waiting with sleep blocked was awkward.
Result: (next review)
Metric: friction reports about waiting on background jobs, over the next 40 runs (baseline: 2 recent).
Decisions: all four candidates keep collecting (intervals overlap / too few runs). derived/ rebuilt (2572 rows). No audit or drift report (reduced pass).

E76: review 2026-10-23 (reduced pass, fifth run) pending

Change: CLAUDE.md cloud-session pointer names the working node --test '*.test.mjs' form (friction on CLAUDE.md: node --test . fails; the fix lived only in cloud-session.md).
Result: (next review)
Metric: friction reports about node --test . failing, over the next 40 runs (baseline: 1 recent). Checked session_gate.py publish-branch friction: already fixed, so no change.
Decisions: all four candidates keep collecting (intervals overlap / too few runs). derived/ rebuilt (2524 rows). No audit or drift report (reduced pass).

E75: review 2026-10-23 (reduced pass, fourth run) pending

Change: .github/prompts/cloud-session.md note on the process-kill hook trap now says the pgrep form is refused too (friction on research/CLAUDE.md reported two retries on the rejection).
Result: (next review)
Metric: friction reports about that hook rejection, over the next 40 runs (baseline: 1 recent).
Decisions: all four candidates keep collecting (intervals overlap / too few runs). derived/ rebuilt. No audit or drift report (reduced pass).

E74: review 2026-10-23 (reduced pass, third run) pending

Change: research/CLAUDE.md gained a 3-line note that wntr run_sim writes temp.* in the cwd, so parallel EPANET runs need a distinct file_prefix.
Result: (next review)
Metric: friction reports against research/CLAUDE.md about simulator temp files colliding, over the next 40 runs (baseline: 1 today).
Decisions: all four candidates (ventures 2, research 1, process 2, review 2) unchanged: intervals overlap, or fewer than 10 runs. derived/ regenerated. No audit or drift report (reduced pass).

E73: review 2026-10-23 (reduced pass, second run) pending

Change: .github/prompts/cloud-session.md "Long commands" gained one sentence: don't chain a long command with commit/push in one call, and use absolute paths for rm -rf after a cd (the safety check rejects relative globs).
Result: (next review)
Metric: friction reports against CLAUDE.md about long commands delaying a commit or a blocked relative-glob rm, over the next 40 runs (baseline: 2 recent).
Decisions: research candidate 1 kept (same non-gap on integrity issues as E72; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). derived/ regenerated. No audit or drift report (reduced pass).

E72: review 2026-10-23 (reduced pass) pending

Change: .github/prompts/cloud-session.md "Python and tests" gained a 3-line fallback for when pyenv.py fails (PEP 668, import name differing from the pip name): a plain venv, in the background.
Result: (next review)
Metric: friction reports against platform/ci/pyenv.py or research/CLAUDE.md about pyenv/PEP 668 failures, over the next 40 runs (baseline: 5 recent on pyenv.py, 3 on research/CLAUDE.md today).
Decisions: research candidate 1 kept (retire verdict fires on 1/158 vs 5/799 integrity issues, 0.0063 vs 0.0063, not a real gap; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). derived/ regenerated. No audit or drift report (reduced pass).

E71: review 2026-10-22 pending

Change: prompts/steering.md cut from 161 to 136 lines (budget 120): the rule/guideline example table, the "habits" list, the JSON deviation example and the Measuring paragraph folded into the prose that already says the same. prompt.md closing paragraph rewrapped.
Metric: measure "Instruction lines" for steering.md and friction reports naming it, over the next 40 runs (baseline: 0 friction reports on it; the cut is for read cost, not for a known failure).
Decisions: research candidate 1 kept (retire verdict fires on one disclosed integrity caveat; E55 proposal stands). ventures 2, process 2, review 2 unchanged (overlap / too few runs). derived/ regenerated (fetch ran, 50 PRs).

E70: review 2026-10-22 (reduced pass) pending

Change: prompts/review/prompt-1.md step 1 now names measure --section (it exists; the last four reviews still reported "measure is 50KB+ with no section flag").
Result: (next review)
Metric: friction reports against prompts/review/prompt-1.md saying measure has no section flag, over the next 40 runs (baseline: 4 recent).
Decisions: research candidate 1 kept (retire verdict fires on 1/154 vs 5/790 integrity issues, a 0.0065 vs 0.0063 difference, not a real gap; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). derived/ regenerated (50 PRs). No audit or drift report (reduced pass).

E69: review 2026-10-22 (reduced pass) pending

Change: portfolio/flagships.yaml wall-framing-takeoff next: cut from about 5k to 0.6k characters: open items only, landed history left to the README and ROADMAP (the follow-up after E66 and E67). changepoint-detector's next: is the remaining long one.
Result: (next review)
Metric: friction reports against portfolio/flagships.yaml, over the next 40 runs (baseline: 42 total; E66 and E67 cuts are in the same window).
Decisions: research candidate 1 kept (retire verdict still fires only on the one disclosed caveat; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). derived/ regenerated. No audit or drift report (reduced pass).

E68: review 2026-10-22 (reduced pass) pending

Change: portfolio/flagships.yaml changepoint-detector next: cut from about 11k to 0.4k characters: open items only, landed history left to the README and git (the follow-up E67 named; all three long next: fields are now cut).
Result: (next review)
Metric: friction reports against portfolio/flagships.yaml, over the next 40 runs (baseline: E66/E67 window).
Decisions: research candidate 1 kept (retire verdict still fires only on the one disclosed caveat; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). derived/ regenerated. No audit or drift report (reduced pass).

E67: review 2026-10-22 (reduced pass) pending

Change: portfolio/flagships.yaml einvoice-preflight next: cut from about 7.5k to 0.6k characters: open items only, landed history left to the README and git (the follow-up E66 named). changepoint-detector's next: is the remaining long one.
Result: (next review)
Metric: friction reports against portfolio/flagships.yaml, over the next 40 runs (baseline: 40 total; E66's ai-bookkeeper cut is in the same window).
Decisions: research candidate 1 kept (retire verdict still fires only on the one disclosed caveat; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). derived/ regenerated (fetch ran, 25 of 43 PRs). No audit or drift report (reduced pass).

E66: review 2026-10-22 (reduced pass) pending

Change: portfolio/flagships.yaml ai-bookkeeper next: cut from 18.7k to about 1.1k characters: open items only, landed history left to the README and git. Cause: 3 friction reports today that it was a duplicated run-on. The other flagships' next: (changepoint-detector 10.5k, einvoice-preflight 8k) are the same shape and are the next cut.
Result: (next review)
Metric: friction reports against portfolio/flagships.yaml, over the next 40 runs (baseline: 32 total, 3 on ai-bookkeeper today).
Decisions: research candidate 1 kept (retire verdict still fires only on the one disclosed caveat; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). derived/ regenerated (fetch ran, 25 of 44 PRs). No audit or drift report (reduced pass).

E65: review 2026-10-22 (reduced pass) pending

Change: prompt.md trimmed from 103 to 102 lines, budget 100 (the explore/creative paragraph no longer restates percentages that live in routing.json). derived/ regenerated (fetch ran, 25 of 43 PRs before its time cap).
Result: (next review)
Metric: Instruction lines for prompt.md (baseline 103, budget 100) and friction reports against it, over the next 40 runs (baseline: 54 total).
Decisions: research candidate 1 kept (retire verdict still fires only on the one disclosed caveat; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). No audit or drift report (reduced pass; the 2026-10-21 report stands).

E64: review 2026-10-21 (reduced pass, late night) pending

Change: .github/prompts/cloud-session.md gains "Browser and process tools": the Playwright/Chromium env vars for cloud sessions and the pkill -f self-kill trap (2 reports each in the last day). derived/ regenerated (fetch ran).
Result: (next review)
Metric: friction reports about playwright paths or pkill, over the next 40 runs (baseline: 2 and 1).
Decisions: research candidate 1 kept (retire verdict still fires only on the one disclosed caveat; E55 proposal stands). ventures 2, process 2 and review 2 unchanged (overlap / too few runs). No audit or drift report (reduced pass; four reviews already ran today).

E63: review 2026-10-21 (reduced pass, late night) pending

Change: prompts/redteam/prompt-0.md gives a cutoff for slow suites (--timeout; pick another target when baseline x mutants > ~20 min). Cause: 3 reports today of changepoint-detector campaigns past 10-30 min. derived/ regenerated.
Result: (next review)
Metric: redteam/prompt-0 friction about slow suites, over the next 40 runs (baseline: 3).
Decisions: research candidate 1 kept (E55 caveat); ventures 2, process 2 and review 2 unchanged (overlap / too few runs). No audit or drift report (reduced pass; the 2026-10-21 report stands).

E62: review 2026-10-21 (reduced pass, night) pending

Change: .github/prompts/cloud-session.md ("Python and tests") says to install with "$PY" -m pip and not to run two pip installs at once (1 report today: pip went to python3.13, concurrent installs corrupted the env). derived/ regenerated.
Result: (next review)
Metric: friction reports about pip/python version mismatch, over the next 40 runs (baseline: 1).
Decisions: research candidate 1 kept (retire verdict still fires only on the one disclosed caveat; E55 proposal stands). ventures 2 (overlap), process 2 (1 run) and review 2 (10 runs, 100% vs 90%, overlap) unchanged. No audit or drift report (reduced pass).

E61: review 2026-10-21 (full pass, evening) pending

Change: steering.md "improvement loop" cut from 17 to 7 lines (duplicated process/prompt-1.md); 171 -> ~161 lines. derived/ regenerated.
Result: (next review)
Metric: process/2 partial rate vs process/1 (baseline 27%), once each has 10+ runs; steering.md friction reports (baseline 0 this window).
Decisions: research candidate 1 kept (retire verdict fires only on the one disclosed caveat; E55). ventures 2 (90% vs 89%, overlap) and review 2 (9 runs) unchanged. Audit clean (3 PRs); drift report drift/2026-10-21.md.

E60: review 2026-10-21 (reduced pass, afternoon) pending

Change: CLAUDE.md ("Hosting") and portfolio/hosting/README.md ("What is served now") name duncan-haywood.github.io/oss-monorepo/; the old /monorepo-projects/ path 404s since the repo rename (2 reports today; the full URL sweep is the owner's draft PR #8592, left alone).
Result: (next review)
Metric: friction reports that the Hosting URL is wrong, over the next 40 runs (baseline: 2).
Decisions: research candidate 1 kept (retire verdict still fires on the one disclosed caveat; E55). ventures 2 and review 2 unchanged (overlap / 9 runs). derived/ regenerated. No audit or drift report (reduced pass).

E59: review 2026-10-21 (reduced pass, morning) pending

Change: prompts/research/prompt-0.md and research/PROJECTS.md no longer hard-code "nineteen tracks" (there are 21; a count in prose drifts). derived/ regenerated.
Result: (next review)
Metric: friction reports saying the track count is stale, over the next 40 runs.
Decisions: research candidate 1 kept (retire verdict still fires on the one disclosed caveat; the verdict fix remains the E55 proposal). ventures 2 and review 2 unchanged (overlap / 9 runs). No audit or drift report (reduced pass).

E58: review 2026-10-21 (reduced pass) pending

Change: CLAUDE.md says to fetch branches with --depth=40 (a bare fetch pulled ~5 GB); research/CLAUDE.md names the exact --write command instead of "step 4's" (reported as pointing at a flag build.py lacks). derived/ regenerated.
Result: (next review)
Metric: friction reports on those two lines, per run, over the next 40 runs.
Decisions: research candidate 1 kept (the retire verdict still fires on the one disclosed caveat; see E55). ventures 2 and review 2 unchanged (overlap / under 10 runs). No audit or drift report (reduced pass).

E57: review 2026-10-20 (reduced pass, night) pending

Change: portfolio/router/allocate.py prints the first 500 characters of a flagship's next: in the work queue and points at flagships.yaml for the rest (84 friction reports on the file, mostly the multi-KB blob per row). test_allocate.py passes (73 tests). derived/ regenerated.
Metric: friction reports naming allocate.py output size, per run, over the next 40 runs (baseline 84 in 1,770, most of them the blob).
Decisions: research candidate 1: the verdict says retire on one integrity issue (the disclosed search-excerpt caveat); kept, as in E54/E55, and the verdict fix remains a proposal for the owner (E55). ventures 2 and review 2 unchanged (overlap / under 10 runs). No audit or drift report (reduced pass).

E56: review 2026-10-20 (night) pending

Change: runcost.py attach takes --human-hours/--human-basis as optional (basis still required with hours); prompt.md says so (and is back to its 100-line budget). Scheduled runs skipped the whole cost step because the human-hours estimate had no basis, losing tokens and dollars with it (prompt.md/runcost.py: ~46 and 3 friction/deviation reports).
Result: (next review)
Metric: share of new records with a usage block, over the next 40 runs (baseline: median API $ measured for about 55% of runs in measure).
Decisions: ventures 2 and review 2 unchanged (intervals overlap / 8 runs). Research candidate 1: the retire suggestion fires on one disclosed caveat (see E55); kept, 99% vs 98% success. derived/ regenerated. No audit (reduced pass), no new candidate.

E55: review 2026-10-20 (reduced pass, evening) pending

Change: none to prompts. derived/ regenerated. A fix to the retire rule in promptctl.py verdict (retire on integrity issues only at two or more, since one self-disclosed caveat on research/1 keeps firing it) was blocked by the auto-mode classifier twice (this run and draft PR #8901), so it is a proposal.
Result: (next review)
Proposal for the owner: allow verdict to require 2+ integrity issues before "retire: more integrity issues". Three review friction reports on 2026-10-20 describe the manual look it costs each review.
Decisions: research candidate 1 kept (99% vs 98% success). ventures 2 and review 2 unchanged (overlap / under 10 runs). No audit or drift report (reduced pass).

E54: review 2026-10-20 (reduced pass, 12:xx) pending

Change: verify-and-merge.md line for --for-changes now says what to do when it stalls (docs-only diff, or a diff reaching 100+ workflows): run the reached workflows by name. Five friction reports in the window (the latest three on 2026-10-20) describe the stall.
Result: (next review)
Metric: friction reports naming run_workflow_locally.py --for-changes stalling, per run, over the next 40 runs (baseline 5 in 1,658).
Decisions: research candidate 1 kept again (116 runs, 99% vs 98% success; its one integrity issue is the disclosed search-excerpt caveat the retire rule counts). ventures 2 and review 2 unchanged (intervals overlap / under 10 runs). derived/ regenerated. No audit or drift report (reduced pass).

E53: guidance map and layers, 2026-10-20 (owner-requested, not a review) pending

Change: one map of the guidance (prompts/SYSTEM.md, prompts/system.json, promptctl.py map), with seven layers and precedence. validate now fails on a guidance file that has no map entry. INSTRUCTION_FILES and the budgets are read from the map. Operational detail moved down a layer: from CLAUDE.md, "Reading CI" went to verify-and-merge.md, "Duplicate PRs" to duplicate-work.md, and the hosting state to portfolio/hosting/README.md. The cloud-container notes, which had been copied into CLAUDE.md, prompt.md (the E49 and E50 lines) and land-your-work.md, went to one file, cloud-session.md. CLAUDE.md went from 360 to 306 lines.
Result: (next review)
Metric: the share of runs whose doc_friction names CLAUDE.md, prompt.md or land-your-work.md on gh/MCP, shallow clones, pytest or long commands, over the next 200 runs. The baseline (keyword matches, an upper bound) is, out of 1,491 runs since 2026-10-14: long commands 69, pytest 48, shallow 34, gh/MCP 23. E49 and E50 are now judged on cloud-session.md, which prompt.md links to. Also: any friction item that says a moved fact could not be found.
Evidence strength: a before-and-after comparison, not A/B.

E52: review 2026-10-20 (second) pending

Change: prompt.md step 1 says which track an owner request that fits none takes (prompt.md/routing.json friction: 3 reports of forcing a track with no rule).
Metric: "no track fits owner task" friction reports on prompt.md, next 20 runs.
Candidates: research/1 vs 0: measure says retire (1 integrity issue in 110 runs vs 4 in 618), but the one issue is a self-disclosed caveat (search-excerpt figures marked in the data and README), not a violation; success is 100% vs 98%. Kept, 0.2 share. ventures 2 / review 2 unchanged (overlap, under 10 runs). Reduced pass: no audit, drift report or derived/ regeneration.

E51: review 2026-10-20 pending

Change: research/CLAUDE.md intro names the root CLAUDE.md rules that still apply to research (public-repository and landing rules), not only the ones that do not (prompt.md friction: took a read to reconcile). derived/ regenerated (1513 runs).
Result: (next review)
Metric: "root CLAUDE.md reconcile" friction reports on prompt.md/research/CLAUDE.md, next 20 research runs.
Candidates: unchanged (review/2 has 8 runs; ventures 2 / research 1 intervals overlap). Reduced pass: no audit or drift report (hours-old report).

E50: review 2026-10-19 (tenth) pending

Change: prompt.md step 2 gives the pyenv.py pytest one-liner (top CLAUDE.md/prompt.md friction: bare python3 lacked pytest, runs ran tests by hand). derived/ regenerated.
Result: (next review)
Metric: "pytest absent" friction reports on CLAUDE.md/prompt.md, next 20 runs.
Candidates: unchanged (review/2 has 8 runs; ventures 2 / research 1 overlap). Reduced pass: no audit or drift report (hours-old report).

E49: review 2026-10-19 (eighth) pending

Change: prompt.md step 2 says to retry create_pull_request on HTTP 500 (prompt.md friction: two 500s before success, no doc mentioned retrying).
Result: (next review)
Metric: "create_pull_request 500" friction reports on prompt.md, next 20 runs; expect none.
Candidates: unchanged (ventures 2 / research 1 intervals overlap; review/2 has 7 runs). Reduced pass: no audit, drift report or derived/ regeneration (reports hours old, recent outputs research/doc-only).

E48: review 2026-10-19 (seventh) pending

Change: research/CLAUDE.md pyenv step now tells runs to test-import pytest along with the project's libraries before skipping the venv (friction: a plain attempt failed on missing pytest). The apt-get and run_in_background reports were already fixed in that file.
Result: (next review)
Metric: "pyenv/pytest detour" friction reports on research/CLAUDE.md, next 20 research runs.
Candidates: unchanged (review/2 has 7 runs; ventures 2 / research 1 overlap). Reduced pass: no audit or drift report (hours-old report).

E47: review 2026-10-19 (sixth) pending

Change: none to prompts or docs. derived/ regenerated (1372 runs). The "due counter ignores same-day reviews" friction on prompts/review/prompt-1.md is already fixed in promptctl.py (route comment near line 147), so no edit; the reports are from runs before that fix.
Result: (next review)
Metric: unchanged from E45/E46.
Candidates: unchanged (ventures 2 / research 1 intervals overlap; review/2 has 7 runs). Reduced pass: no audit or drift report (hours-old report).

E46: review 2026-10-19 (fifth) pending

Change: prompts/research/prompt-0.md claim step greps git ls-tree for a keyword instead of reading the ~670-line listing (research/CLAUDE.md friction).
Result: (next review)
Metric: "ls-tree prints ~670 lines" friction reports, next 20 research runs; expect none.
Candidates: unchanged. Reduced pass: no audit or drift report (hours-old report, recent outputs research/doc-only). pyenv.py re-checked: works now.

E45: review 2026-10-19 (fourth) pending

Change: research/_shared/pick_track.py says when the drawn track already has 40+ projects, that a redraw is fine, and names the three emptiest tracks (research/PROJECTS.md friction: draws landing on saturated tracks 4 and 5 cost a run its first reads). derived/ regenerated.
Result: (next review)
Metric: "saturated track / micro-variant chain" friction reports in research/PROJECTS.md, next 20 research runs; expect fewer.
Candidates: unchanged (review/2 has 7 runs; ventures 2 / research 1 overlap). Reduced pass: no audit (recent outputs doc-only or research); no new drift report (prompts/drift/2026-10-19.md is hours old).

E44: review 2026-10-19 (third) pending

Change: platform/ci/pyenv.py tries the private venv before apt-get (apt-get update is slow and often fails; PEP 668 pip failures and 120s foreground timeouts were 5+ research/CLAUDE.md friction reports). Tests pass (10).
Result: (next review)
Metric: friction reports about pyenv PEP 668 / timeouts, next 20 research runs; expect none.
Candidates: unchanged (review/2 has 7 runs; ventures 2 / research 1 overlap). Reduced pass: no derive fetch or audit (recent outputs doc-only or research).

E43: review 2026-10-19 (second) pending

Change: allocate.py rank() breaks merit ties by least-worked first (ledger entries + portfolio mentions, ascending) before name. Top rows were a merit tie sorted alphabetically, so every run met the same already-deep ventures (reverb, where-to-sell; the top friction in CLAUDE.md and ventures/prompt-1.md). Top 3 now include 0-ledger ventures.
Result: (next review)
Metric: "allocator top rows already deep" friction reports, next 20 ventures runs; expect fewer.
Candidates: unchanged. test_allocate passes (57 tests).

E42: review 2026-10-19 pending

Change: research/CLAUDE.md pyenv note now says first use can take over 2 minutes and to run it with run_in_background (E41 said ~20s; 5 reports of a 120s foreground timeout since). derived/ regenerated (53 PRs).
Result: (next review)
Metric: friction reports about pyenv foreground timeouts, next 20 research runs; expect none.
Candidates: unchanged (ventures 2 / research 1 overlap; review/2 has 7 runs). Audit not run (recent outputs doc-only or mirrored).
Drift: prompts/drift/2026-10-19.md.

E41: review 2026-10-18 (seventh) pending

Change: research/CLAUDE.md now says the first pyenv.py call takes ~20s and prints install chatter on stderr (3 friction reports). The yosys/apt-get reports in CLAUDE.md friction were already fixed (lines 76-77); measure still lists them as latest.
Result: (next review)
Metric: friction reports about pyenv first-use time, next 20 research runs; expect none.
Candidates: unchanged (ventures 2 / research 1 overlap; review/2 has 7 runs). derived/ not regenerated, audit not run (no GitHub API fetch attempted; recent outputs doc-only/research).

E40: review 2026-10-18 (sixth) pending

Change: session_gate.py stale-snapshot reason now says "after your PR merges" up front. It used to say "before anything else" and then, three sentences later, that the script refuses a work branch (35 friction reports on the gate, the latest several on exactly this).
Result: (next review)
Metric: friction reports naming session_gate.py about the publish order, next 20 runs that run the gate; expect none.
Candidates: unchanged (ventures 2 / research 1 overlap; review/2 has 7 runs).
Drift: prompts/drift/2026-10-18f.md.

E39: review 2026-10-18 (fifth) pending

Change: no prompt or doc change. Audit sample (recent venture PRs) found one shipped regression (#8038, fixed by #8041); added a header/row cell-count test to league-schedule-generator.
Result: (next review)
Metric: audit runs that find a defect in a recent venture PR; the audit is runnable when it names a venture with node --test.
Candidates: unchanged. Drift: prompts/drift/2026-10-18e.md.

E38: review 2026-10-18 (fourth) pending

Change: research/_shared/pick_track.py does the track draw that research/PROJECTS.md "Picking work" step 1 describes (a report said no script did it); the step is now one line plus the command.
Result: (next review)
Metric: friction reports about the track draw, next 40 research runs; expect none.
Candidates: unchanged (ventures 2 / research 1 overlap; review/2 has 6 runs). derived/ regenerated (25 of 29 PRs). Audit not run (recent outputs mirrored or doc-only).
Drift: prompts/drift/2026-10-18d.md.

E37: review 2026-10-18 (third) pending

Change: research/PROJECTS.md "Picking work" step 1 gives the track-draw formula (2 reports said the number-to-track mapping was a guess).
Result: (next review)
Metric: friction reports about the track draw, next 40 research runs; expect none.
Candidates: unchanged (ventures 2 / research 1 overlap; review/2 has 6 runs). Open friction not fixed: ventures runs report that router top candidates are already deep; allocate.py needs its own run.

E36: review 2026-10-18 (second) pending

Change: redteam.py runs suites with PYTHONDONTWRITEBYTECODE=1 (stale .pyc kept serving a same-size mutant).
Result: (next review)
Metric: friction reports naming prompts/redteam/prompt-0.md about stale bytecode, next 20 redteam runs; expect none.
Candidates: unchanged. Drift: prompts/drift/2026-10-18b.md.

E35: review 2026-10-18 pending

Change: none to prompts or docs. Candidates unchanged (ventures 2 / research 1 overlap their champions; review/2 has 5 runs, fewer than 10). derived/ regenerated; pyenv.py friction checked and already fixed on main.
Result: (next review)
Metric: friction reports naming research/CLAUDE.md about pyenv.py, next 40 runs; expect none.
Drift: prompts/drift/2026-10-18.md.

E34: review 2026-10-17 (sixth review that day) pending

Change: research/PROJECTS.md step 2 of "Picking work" tells a run to ask a new question when a track's next-step chain has gone to micro-variants (31 reports on prompts/research/prompt-0.md, the latest about the saturated chain).
Result: (next review)
Metric: friction reports about choosing a next step, per research run, next 40 research runs.
Candidates: unchanged (ventures 2 / research 1 overlap their champions; review/2 has 5 runs).
Data: derive.py fetch got 25 of 30 PRs; derived/ regenerated. Audit not run (recent outputs mirrored or doc-only).
Drift: prompts/drift/2026-10-17f.md.

E33: review 2026-10-17 (fifth review that day) pending

Change: sync_tree_counts.py inserts gitops/tests on sys.path for its import, so it runs with no PYTHONPATH (3 reports of No module named front_door_mirror).
Result: (next review)
Metric: friction reports naming platform/ci/sync_tree_counts.py, next 40 runs.
Candidates: unchanged (ventures 2 / research 1 overlap; review/2 has 5 runs).
Data: derive.py fetch got 25 of 30 PRs; derived/ regenerated. Audit not run (recent outputs mirrored or doc-only).
Drift: prompts/drift/2026-10-17e.md.

E32: review 2026-10-17 (fourth review that day) pending

Change: review/prompt-1.md says derive.py fetch is incremental (~30 PRs per call, run once in the background) and that an unrunnable audit is expected, not friction. Cause: 24 friction reports on that file, the latest three about the fetch limit and the audit.
Result: (next review)
Metric: friction reports per run naming prompts/review/prompt-1.md. Compare the next review.
Candidates: unchanged (ventures 2 / research 1 overlap their champions; review/2 has 5 runs).
Data: derive.py fetch got 25 of 30 PRs; derived/ regenerated. Audit not run (recent outputs mirrored or doc-only).
Drift: prompts/drift/2026-10-17d.md.

E31: review 2026-10-17 (third review that day) pending

Change: land-your-work.md trimmed to its 160-line budget (it was over by 3) and says a swept branch whose content is on main needs no PR (4 reports); backlog README/validate message say pr: "this" is accepted for a done item whose PR is unopened (3 reports of the second-commit cost).
Result: (next review)
Metric: friction reports on land-your-work.md and the backlog README per run, next 40 runs.
Candidates: unchanged (ventures 2, research 1 overlap; review/2 has 5).
Data: derive.py fetch got 32 of 470 PRs; the rest stay null. Audit not run (mirrored or doc-only recent outputs).
Drift: prompts/drift/2026-10-17c.md.
Audit: not run (recent records mirrored or doc-only); derive.py fetch not run, third review in one day.

E30: review 2026-10-17 pending

Change: land-your-work.md says cloud sessions without gh/GraphQL use the GitHub MCP tools to open and merge. Cause: a 2026-10-17 CLAUDE.md friction report ("merge via PR, but gh GraphQL is blocked").
Result: (next review)
Metric: friction reports per run naming gh/GraphQL/MCP. Compare the next 40 runs.
Candidates: ventures/2 vs 1, research/1 vs 0 overlap; review/2 has 5 runs. Keep collecting.
Data: derive.py fetch got 25 of 46 PRs before the time limit; the rest stay null. Audit not run (mirrored or doc-only recent outputs).
Drift: see prompts/drift/2026-10-17b.md.

E29: 24-hour review, 2026-10-16 (owner-requested) pending

Change: promptctl.py measure hides "By starting message" groups with fewer than 3 runs (49 of 60 rows, 52 runs; one line says so). Cause: the table is one-off hashes that cannot be compared, and the output was ~43KB that every review reads.
Measure: CLAUDE.md friction reports per run naming pytest, numpy, pip, interpreter or timeout. Before: 2026-10-16, 224 runs, about 34 such reports. Compare the next 100 runs.
Result: (next review)
Candidates: unchanged. ventures 2 and research 1 overlap their champions; review/2 has 5 runs.
Audit: not run (mirrored or doc-only records); derive.py fetch ran and built 713 rows.
Proposal: G2 (process track weight).
Report: prompts/drift/2026-10-16k.md.
Metric: size of measure output; "doc friction" reports naming prompts/promptctl.py (7 before; the BrokenPipe one was already fixed).
Drift: prompts/drift/2026-10-17b.md.

E28: review 2026-10-16 (tenth of the day) pending

Change: review_every 20 → 40 in prompts/routing.json. Ten reviews fired today, most with nothing new to measure (friction on routing.json, review/prompt-2.md).
Result: (next review)
Metric: reviews per day, and review runs with no change shipped. Compare the next 40 runs against today's.
Candidates: unchanged, overlap or too few runs.
Audit: league-schedule-generator, 69 pass. derive.py fetch got 16 of 18.
Drift: see prompts/drift/2026-10-16j.md.

E27: review 2026-10-16 (ninth of the day) pending

Change: none to prompts or docs; prompts/derived/ refreshed, drift report only. No new friction that was fixable at low cost.
Result: (next review)
Candidates: ventures/2 (25, 88%) vs ventures/1 (78, 87%), research/1 (29, 100%) vs research/0 (358, 98%): overlap, keep collecting. review/2: 3 runs.
Audit: business-loan-offer-comparator re-run, 34 pass. derive.py fetch got 18 of 20.
Drift: see prompts/drift/2026-10-16i.md.

E26: review 2026-10-16 (eighth of the day) pending

Change: none to prompts or docs; prompts/derived/ refreshed, drift report only. Stale friction (PRODUCT_PLAYBOOK, prompts/README) already fixed.
Result: (next review)
Candidates: ventures/2 vs ventures/1, research/1 vs research/0: overlap, keep collecting. review/2: 3 runs, too few.
Audit: skipped (doc-only/research runs). derive.py fetch got 19 of 21.

E25: review 2026-10-16 (seventh of the day) pending

Change: none to prompts or docs; measurement refresh (prompts/derived/) and drift report only. Nothing in the friction table was new and fixable at low cost, and review/2 has 3 runs.
Measure: n/a.
Result: (next review)
Candidates: ventures/2 (22, 91%) vs ventures/1 (68, 91%), research/1 (18, 100%) vs research/0 (348, 98%): overlap, keep collecting. review/2: 2 runs.
Audit: league-schedule-generator re-run, 59 pass. Research audit not runnable (no pytest, apt fails).
Drift: see prompts/drift/2026-10-16g.md.

E24: review 2026-10-16 (sixth of the day) pending

Change: prompt.md step 1 now says to pass --track before running route when the brief names a task (was after the command; friction: a run drew research for an owner-named product brief).
Measure: friction reports naming prompt.md per run after this.
Result: (next review)
Candidates: ventures/2 (22, 91%) vs ventures/1 (68, 91%), research/1 (18, 100%) vs research/0 (348, 98%): overlap, keep collecting. review/2: 2 runs.

E23: review 2026-10-16 (fifth of the day) pending

Change: CLAUDE.md checklist gains one line naming gitops/ci/sync_readme_counts.py. No prompt changes.
Why: friction on gitops/ci/README.md: runs-on count goes stale on every new-workflow PR.
Measure: friction reports naming gitops/ci/README.md per run after this.
Result: (next review)
Candidates: ventures/2 (20, 90%) vs ventures/1 (63, 92%), research/1 (17, 100%) vs research/0 (340, 98%): overlap, keep collecting. review/2 has 1 run.
Audit: 3 areas re-run, all pass. derive.py fetch got 21 of 23 PRs.
Drift: see prompts/drift/2026-10-16e.md.

E22: review 2026-10-16 (fourth of the day) pending

Change: .github/prompts/land-your-work.md adopted-drafts bullet condensed (163 -> 159 lines, under its 160 budget). No prompt changes.
Why: only file over its instruction-line budget; CLAUDE.md already fell 467 -> 391.
Measure: friction reports naming land-your-work.md per run after this window.
Result: (next review)
Candidates: ventures/2 (20 runs, 90%) vs ventures/1 (63, 94%): intervals overlap, keep collecting. research/1 (12, 100%) vs research/0 (331, 98%): overlap, keep collecting. review/2 has 0 runs and review/1 16: first review/2 run is this one; too few for a verdict.
Audit: 3 venture PRs re-run (devsec-toolkit, shift-roster-builder, distribution tests): all pass, no corrections. derive.py fetch got 19 of 21 PRs.
Drift: no new owner words. See prompts/drift/2026-10-16d.md.

E21: review/2 candidate and BrokenPipe fix (2026-10-16) pending

Change: review/prompt-2.md (candidate, share 0.2): the audit step says to decide first whether anything is auditable and log "nothing auditable" if not; step 1 says to run build on the cache after a 403. promptctl.py exits quietly on a closed pipe (measure | head).
Why: review/1 friction (16 reports) is mostly "steps 1-2 unrunnable because the remaining records are mirrored or doc-only".
Measure: friction reports per review run on steps 1-2, review/2 vs review/1. Only ~1 review per 20 runs, so expect a slow verdict.
Result: (next review)
Candidates this window: ventures/2 (18 runs, 94%) vs ventures/1 (58, 93%): intervals overlap, keep collecting. research/1 (10, 100%) vs research/0 (322, 98%): overlap, keep collecting. review/1: 15 runs, 73%, 0 integrity.
Audit: nothing auditable (new records mirrored or doc-only). derive.py fetch fetched 16 of 18 PRs; merged still unknown for mirrored records.
Drift: no new owner words since 2026-10-16b. CLAUDE.md 467/500 (down from 475). See prompts/drift/2026-10-16c.md.

E20: stale doc lines from the same review (2026-10-16) pending

Change: CLAUDE.md checklist no longer states a stale count of expected gitops/tests failures (said ~26; baseline today was 2); review/prompt-1.md step 1 says what to do when derive.py fetch gets a 403 (+2 lines).
Why: friction reports on both; 26 runs since the last review.
Measure: friction reports naming the CLAUDE.md pytest line, and on review/prompt-1.md step 1, per run after this window.
Result: (next review)
Candidates: ventures/2 (17) vs ventures/1 (53): intervals overlap, keep collecting. research/1 (8) under 10. Audit skipped: records mirrored or doc-only. derive.py fetch 403; merged left unknown.

E19: thirteenth review (2026-10-16) pending

Change: none to prompts or docs; prompts/derived/runs.json rebuilt. derive.py fetch got HTTP 403 from the GitHub API in this session, so the rebuild used the cached github.json: merged for new records stays unknown until a session with API access runs fetch.
Why: review step 1 died on its first command in this session.
Measure: reviews that report a failed derive.py fetch in doc_friction.
Result: (next review)
Candidates: ventures/2 (17 runs, 94%, 0.06 errors/run) vs ventures/1 (53, 92%, 0.17): intervals overlap, no verdict. research/1 has 8 runs, under 10. review/1 is 75% over 12 runs, 0 integrity issues. No new candidate: both tracks with candidates already carry one.

E18: measure PR facts and rework instead of typing them (2026-10-16, owner request) pending

Change: prompts/derive.py (new) measures merged, PR size, red-CI commits and 14-day rework (reverts, follow-ups, line survival) into prompts/derived/. promptctl.summarise falls back to the measured merged and reports median survival. Review prompt step 1 runs it instead of hand-filling records. Goal: "Measured, not typed" in goals.md.
Why: of 441 records, 31 typed merged, 18 lines_added, 10 wall_minutes and 5 cost_usd. Rework is a scoring measure in goals.md, and no record had ever carried it.
Measure: merged_unknown in measure, before against after the first derive.py build. Survival per prompt version and model, once rows have about 10 closed windows each.
Result: (next review)

E17: twelfth review (2026-10-16) pending

Change: promptctl.runs_since_review no longer counts records mirrored from the upstream research repo (they carry source); test added. Acts on E15's proposal. land-your-work.md 169 -> 158 lines (measured-history prose cut; budget 160).
Why: reviews fired on mirrored records no run of this repo produced.
Measure: reviews per day (the 2026-10-15 count was ten); lines in land-your-work.md.
Result: (next review)

E16: eleventh review of 2026-10-15 pending

Change: prompts/process/backlog/README.md "Closing" now says to commit in_progress first and flip to done once the PR number exists (+2 lines).
Why: two friction reports (README, promptctl validate) hit a red validate on the first commit of a run closing an item.
Measure: backlog-closing friction reports per process run.
Result: (next review)

E15: tenth review of 2026-10-15 pending

Change: none to prompts or docs. ventures/2 reached 10 runs (90% success, 0.1 errors/run) against ventures/1 (29 runs, 90%, 0.24): intervals overlap, so no promote or retire; research/1 has 7 runs, still under 10. No new candidate (both tracks already carry one).
Measure: reviews per day should fall once mirrored records stop counting.
Result: (next review)
Audit: skipped; the 21 runs since the last review are mirrored upstream research records or doc-only. merged backfill skipped for the same reason.
Drift: no new owner words since the earlier reports today (see prompts/drift/2026-10-15.md); no instruction file grew past budget (CLAUDE.md 457/500, land-your-work.md at its 160 budget).
Observation: this is the tenth review today. runs_since_review counts mirrored upstream records toward review_every, which is why reviews keep firing at ~20 mirrored runs. Proposal for the next process run: count only records whose PR is in this repo.

E14: ninth review of 2026-10-15 pending

Change: prompts/review/prompt-1.md step 2 says mirrored and doc-only runs are skipped without counting as friction (eight reports of "nothing auditable" came from the review prompt itself). +2 lines.
Why: ventures/2 has 8 runs and research/1 5, still under 10 each, so no promote/retire and no new candidate.
Measure: review/1 friction reports per run about unauditable records.
Result: (next review)
Audit: ventures/meter-keeper (newest merged venture): 26/26 tests pass, README disclaimer and page disclaimer present.

E13: eighth review of 2026-10-15 pending

Change: CLAUDE.md -1 line net: the "Tool timeout" bullet duplicated "Long suites" (added by E12); merged into one that also names git fetch --unshallow and publish-branch.sh, which three reports say time out too.
Why: CLAUDE.md is the top friction file (36 reports); shorter beats longer.
Measure: timeout friction reports on CLAUDE.md per run after this review.
Result: (next review)

E12: seventh review of 2026-10-15 pending

Change: CLAUDE.md checklist notes (4 lines) that the pages/marketing/deploy node suites exceed the tool timeout and must run in the background.
Why: three friction reports (CLAUDE.md, unspecified) hit the 120 s / 10 min limit on those lines.
Measure: timeout reports on those lines per run.
Result: (next review)
Audit: platform/deploy tests re-run: 28/28 pass. No other output was auditable locally (most records are mirrored upstream).

E11: sixth review of 2026-10-15 (due-counter fix) pending

Change: promptctl.runs_since_review orders records by when git added them (ties on date were ordered by random session id, so reviews landing mid-day were counted among later runs). Real tree: 37 "runs since review" before, 5 after. Regression test added.
Why: six reviews ran in one day; four friction/deviation reports and the drift report named the due counter (prompts/routing.json friction).
Measure: reviews per day (expect at most one per ~20 merged runs).
Result: (next review)

E10: fifth review of 2026-10-15 pending

Change: AGENTS.md 291 → 187 lines ("Ship fast" incident stories moved to docs/history/claude-md-2026-10-15.md, rules kept). promptctl measure no longer crashes on free-text doc_friction/deviations entries; four malformed records on main (they failed validate and the prompt-system unit test) were normalised with corrections entries.
Why: measure crashed; validate was red on main; AGENTS.md was over its 250 budget.
Measure: AGENTS.md friction per run; validate stays green.
Result: (next review)

E9: fourth review of 2026-10-15 pending

Change: review/prompt-1.md cut 150 → 128 lines (budget 130): removed the first-review backfill step (170+ records exist), tightened drift-audit prose, candidates only for tracks with about 10 runs. Filled merged: true for four local records (PRs 7001, 7019, 7024, 7034), with corrections.
Why: drift report 2026-10-15-c left "trim review/prompt-1.md" open; two friction reports on it (backfill and per-track candidates).
Measure: friction reports on prompts/review/prompt-1.md per review run.
Result: (next review)

E8: third review of 2026-10-15 pending

Change: ventures candidate prompt-2 replaces the "no near-copies" sentence with a command that counts name suffixes of the newest ventures (share 0.2). AGENTS.md cut 335 → 291 lines (removed stale 2026-09 history that contradicted the ventures prompt).
Why: 11 of the 15 newest ventures are *-calculator; AGENTS.md was over its budget.
Measure: same-suffix share of ventures added by ventures/2 runs vs ventures/1; AGENTS.md friction per run.
Result: (next review)

E7: Steering by judgment, a process backlog, process track to 10% (2026-10-15, owner request) pending

Why: 11 recommendations made on 2026-10-15, 0 adopted (no tool, PR or commit implements any of them by the end of the day); CLAUDE.md 1,662 lines with 12 friction reports; 2 of 3 process runs that day duplicated a study.
Measure: backlog items done per week and the share whose result says helped; deviations recorded per run (expect above 0 if runs use judgment); instruction lines against budget; process-run outcomes on prompt-1 against prompt-0's 3 runs (too few to compare formally).
Result: (next review)

E6: first scheduled review (2026-10-15) pending

Change: review prompt step 1 skips records mirrored from the upstream research repo when filling merged. Four local records filled.
Why: 167 mirrored records can never be filled; one friction report on prompts/review/prompt-0.md.
Measure: friction reports on the review prompt per review run.
Result: (next review)

E5: Judgment over rules, depth over count, drift audit (2026-10-15, owner-directed) pending

Change: goals.md gains "Judgment over rules" (including a concision rule for rules), the owner's business model (one-click self-deployed copies with a donate link) and the payments rule (never the owner's accounts). Ventures champion becomes prompt-1 (no new single-page calculators; finish ventures as self-deploy businesses). Review champion becomes prompt-1 (new step 6, drift audit, with reports in prompts/drift/). Promoted directly, without an A/B test, because the owner directed it, as in E1.
Why: prompts/drift/2026-10-15.md. 641 of 1,236 ventures are *-calculator against a brief asking for ambitious, deep products, and about 3,400 lines of agent-written rules come before a ventures run starts.
Measure: new *-calculator ventures after 2026-10-15 (target 0); instruction lines per file (should fall); ventures with a working self-deploy page (should rise); doc_friction on CLAUDE.md and DIRECTION.md.
Result: (next review)

E4: Red team / blue team track (2026-10-15, owner request, PR https://github.com/Duncan-Haywood/oss-monorepo/pull/6987) pending

Change: a new track, redteam, routed 7% of runs (owner's request; the 5-10% band is in goals.md and enforced by validate). With the process and explore tracks (5% each), ventures and research go to 0.415 each. New prompts/redteam.py seeds mutants into a file a suite guards, runs the suite, restores the file, and writes the results into the run record. measure now reports the catch rate (before and after the blue team's fixes) and real defects per red-team run. Sub-prompt: prompts/redteam/prompt-0.md.
Why: every other measure here counts failures that something already noticed (CI failures, review findings, reverts). Nothing measured how many bugs go unnoticed. CLAUDE.md records many guards that were green while checking nothing. E2's process track reads what the checks caught from history; this one measures what they would miss, by seeding faults.
Measure: catch rate per review window, which should rise as blue-team guards land, and real defects per red-team run, which should fall. The first window is the baseline.
Result: (next review)

E3: Free-choice track, about 5% of runs (2026-10-15, owner request) pending

Change: new explore track (prompts/explore/prompt-0.md) at weight 0.05 in routing.json; ventures and research went from 0.475 to 0.45 each (E2's process track took the other 5% the same day). The owner asked for about 5% of effort to go to the agent's discretion, in the spirit of Google's "20% time". goals.md (owner-edited), prompt.md, AGENTS.md and research/PROJECTS.md (in oss-research-monorepo, which this mirror picks up on its next sync) say so too.
Why: owner direction, not a measured failure.
Measure: success and integrity per explore run against the other tracks, and how many free-choice ideas graduate into a regular track. Expect only about one run in twenty, so allow many reviews before reading anything.
Result: (next review)
Measured: 170 of 172 records are research runs imported from the upstream research repository; their PRs live there, so merged stays null for 167 (not checkable from here, never guessed). Filled and corrected 4 local records (PRs 6988, 6991, 6992, 6993, all merged). 1 integrity issue (tail-elicitation missing references, fixed before merge), 1 leaked. No audit of five outputs was done this run; only four local PRs were examined. ventures/process/explore have 1 or 0 runs: too few to compare.

E2: A process track, about 5% of runs (2026-10-15, owner request) pending

Change: new track process (prompts/process/prompt-0.md) with weight 0.05 in routing.json; ventures and research go from 0.5 to 0.475 each. A process run studies the development process from git, PR and CI history (quality, errors, rework, duplicate and leaked work, what the checks catch) and writes prompts/process/reports/YYYY-MM-DD.md with ranked recommendations. It may implement one small agent-owned fix. The track goal was added to goals.md at the owner's request.
Why: the owner asked for a small, standing share of work to analyse the process and recommend improvements. The review track covers prompts and docs from run records only.
Measure: share of routed runs on process (expect about 5%); per study, how many of the previous report's recommendations were adopted and whether their metrics moved.

E1: Mirror with one writer, archive rather than delete (2026-10-14, owner-directed, PR https://github.com/Duncan-Haywood/oss-monorepo/pull/6984) pending

Change: a direct edit, not a candidate, made at the owner's request to keep the two repositories mirrored with no history loss. - prompts/research/prompt-0.md and prompts/ventures/prompt-0.md: until cutover, research work lands upstream, and keeping the mirror current is standing work. Mirror PRs merge with a merge commit. - research/_mirror/README.md and research/CLAUDE.md: the upstream repository is archived at cutover, never deleted. - land-your-work.md: names the mirror-PR exception to squash-merging.
Why: the research track routes 50% of runs here, and mirror.py sync deletes files under the mirrored area that upstream lacks. So a project built here before cutover would disappear at the next sync. Deleting the upstream repository would also lose its pull requests, review threads and branches.
Measure: research-track runs here that touch research/ outside a sync (target 0), and mirror.py check on main.

E0: One repository, one prompt system (2026-10-12) pending

Change: oss-research-monorepo was folded into this repository under research/. The two standing briefs became the champions of two tracks: prompts/ventures/prompt-0.md (this repository's former prompt.md) and prompts/research/prompt-0.md (the research repository's prompt.md). The root prompt.md is now a master prompt that routes between them. The research brief's experiment loop (champion and candidate, run log, review every 20 runs) was generalised into prompts/ for every track. Its run log became one record per file under prompts/runs/, so parallel runs never conflict. Its integrity and licensing rules moved, unchanged, into prompts/goals.md.
Why: two repositories meant two sets of agent instructions to keep in step, and the research one was frozen during the move. Its four hourly routines were finishing in about 10 seconds with nothing to do.
Measure: baseline. The research repository's own run log came with it: 158 runs (2026-10-13/14, all research/0), converted into prompts/runs/ by research/_mirror/mirror.py. On import, measure read 97% success (95% CI 93-99%), 1 integrity issue, 1 leaked run, and merged unknown for 157. The first review fills in merged from the pull requests, backfills ventures runs from git history (review prompt, step 1), and compares everything later against this.
Result: (first review)

Built from files on main at 2026-11-01T03:16:05.729Z. Every figure links to its source in the repository; nothing here is typed by hand. Machine-readable: analyses.json.