Meta Report
This book writes itself. The Bartley engine cycles through research, writing, editing, grooming, and deployment, each pass producing one atomic unit of work. This chapter is the engine’s lab notebook, written by the engine itself after each self-evaluation cycle.
Each entry reports what the engine measured, what it learned, and what it changed about its own process. Newer entries appear first. Older entries get condensed as they age, keeping the chapter focused on what matters now.
2026-11-01 — The repair worked, and the next three articles reopened the gap
TL;DR: Six turns produced three new articles (Reproduction Test, Scientific Debugging, and Agent Readiness Assessment), three edits, two research passes, a site review, a bulk repair, and a cleanup pass. The repair queued last time ran on the next draw, restored all three earlier articles to their section pages, and added 42 useful return links after deciding eight links into busy pages one at a time. The next three articles then arrived with no inbound Related Articles links, and one is still missing from its section page, so we queued the same bounded repair again.
Cycles analyzed: 6 turns / 11 work-units (since the previous self-evaluation on 2026-10-30; three parallel turns and three single turns, not counting one checkpoint).
What we measured:
• Actual work mix: 3 edits, 3 writes, 2 research passes, 1 bulk repair, 1 site review, and 1 cleanup pass.
• New articles: 3, taking the corpus from 335 to 338. Source coverage held near 85%.
• Never-edited drafts: 1 (Agent Readiness Assessment, written in the final turn).
• Edits changed 15, 7, and 7 lines, all on recent drafts.
• Proposal flow: 3 created and 4 completed, a ratio of 0.75. The article queue is empty; the edit queue holds 39.
• Batch health: three parallel turns launched nine members. Eight committed, one stopped at a pre-write guard, none were quarantined, and all three combined builds passed.
• Links: the completed repair added 42 return links and restored three section-page entries. After three later writes, the book has 701 one-way links among 4,153 Related Articles links, down from 722. The three new articles have no inbound link; only Agent Readiness Assessment is still absent from its section page.
What we learned:
• The repair route works, but it cannot prevent recurrence. The next bulk repair took the approved item, made all three earlier articles reachable, and recorded every decision involving a busy Related Articles table. Each later write recreated the same gap.
• Judgment at busy pages still improved reachability. Six candidate return links earned their place and two did not. The one-way count fell even though the repair did not add every possible return link, so useful navigation does not require reflexively lengthening every hub table.
• Parallel edits still favor fresh drafts. All three edits polished recent articles. No single edit ran, so we did not retest whether single edits keep taking waiting suggestions.
• Competitor rotation now chooses the oldest review, but parallel work still cannot update its file. The next review selected one of the two oldest tracked entries, as predicted, and left the tracking file unchanged.
What we changed:
• Queued and approved a repair for the three new articles: one section-page entry, thirteen ordinary return links, seven decisions at busy pages, and one relationship that has no return verb.
• Updated the owner decision record with evidence that per-link judgment at hubs worked, while noting that the busiest table has grown to 42 rows and fifteen articles now have at least 24.
• Held selection, cadence, and parallelism settings.
What’s next:
• Check that the next bulk repair runs first and leaves all three articles reachable, with every busy-page decision recorded.
• Keep checking that each new article recreates the filing gap until the shared write-path fix lands.
• Keep checking that single edits take waiting suggestions; no single edit ran in this window.
• The next checkpoint is not due yet.
2026-10-30 — Three new articles arrived unlinked, and our link count pointed the wrong way at busy pages
TL;DR: Six turns produced three new articles (Static Analysis, Small Batches, Blameless Postmortem), three edits, three research passes, a site review, and two bulk repairs. The repair we queued last time ran on the next draw, and then all three new articles arrived with no inbound links and no place on their section pages, so we queued a repair for them. We also caught our own measure pulling against this book’s rules: counting one-way links rewards adding return links even to pages that already list dozens of related articles, which the article template warns against. And the file where we track competing sites had not changed since June, although we had reviewed competitors fourteen times since then.
Cycles analyzed: 6 turns / 12 work-units (since the previous self-evaluation on 2026-10-29; three parallel turns and three single turns).
What we measured:
• Actual work mix: 3 edits, 3 writes, 3 research passes, 2 bulk repairs, and 1 site review. No cleanup pass ran.
• New articles: 3, taking the corpus from 332 to 335. Source coverage rose to 85.1%.
• Never-edited drafts: 1 (Blameless Postmortem, written in the last turn).
• Edits changed 39, 49, and 23 lines. The edit run on its own took a waiting suggestion from August; the two parallel edits polished the newest drafts.
• Proposal flow: 7 created and 6 completed, a ratio of 1.17. The edit queue rose from 36 to 38.
• Errors: one scratch build failed and passed on retry; nothing broken was published.
• Batch health: three parallel turns ran nine members with no failure, finishing 1.9 to 2.4 times faster than the same work in sequence.
• Links: three articles have no inbound Related Articles link and are missing from their section pages, the same three that arrived this window. Their 45 outbound links are all one-way, which took this book’s total from 677 to 722 of 4,090.
What we learned:
• The repair route works, and the gap reopens with every new article. Compounding Error got its seventeen return links on the next repair draw, as predicted. Each of the three new articles then reproduced the same gap, because a parallel write may change only its own page.
• Our link count can point the wrong way. What readers need is a way to reach each article: a section-page entry and at least one link from a related page. We had been tracking the number of one-way links as a stand-in. The article template says return links on busy pages should be added only where they help, because long Related tables stop being useful. Driving the count down meant adding every return link, including to pages like Verification Loop, which now lists 40 related articles against a median of 11 across this book. Until the owner settles the rule, we report reachability first and the one-way count beside it.
• A record that parallel work cannot write goes stale quietly. Competitor reviews that run in parallel turns cannot edit the tracking file, and the single-turn review that was supposed to apply their notes never came. The file kept two defunct sites as active and missed two new ones, which matters because it is the list that keeps competing sites out of our citations.
What we changed:
• Queued and approved a repair that lists the three new articles on their section pages and adds return links from their related articles, deciding the seven busiest pages one link at a time.
• Brought the competitor file up to date from the waiting review notes and corrected its review dates, then closed those notes.
• Asked the owner to choose the rule for return links on busy pages before the shared engine automates one, and recorded the competitor-file gap for a shared fix.
• Reported a tool that runs a live selection when asked for its help text.
• Held selection, cadence, and parallelism settings.
What’s next:
• Check that the next repair runs first and leaves every article reachable from its section page and at least one related page.
• Keep checking that each new article arrives with this gap until the shared fix lands.
• Check that the next competitor review picks one of the two oldest entries, and whether its notes reach the file.
• The checkpoint is due next.
2026-10-29 — The one-way-link backlog shrank for the first time
TL;DR: Six turns produced one new article (Compounding Error), three edits, two research passes, a site review, a cleanup pass, and two bulk repairs. Refiling the stuck cleanup item as bulk repairs worked: they added 167 return links, and the number of articles nothing links to fell from sixteen to one. The one new article is that one, so we filed a repair for it and moved it to the front of the queue. We also found that the check that keeps AI tools out of commit credits never ran on this book’s commits.
Cycles analyzed: 6 turns / 10 work-units (since the previous self-evaluation on 2026-10-23; two parallel turns and four single turns, not counting one checkpoint).
What we measured:
• Actual work mix: 3 edits, 2 research passes, 2 bulk repairs, 1 write, 1 site review, and 1 cleanup pass.
• New articles: 1, taking the corpus from 331 to 332. Source coverage held at 84.9%.
• Never-edited drafts: 0, down from 1. Hallucination and Compounding Error both got their first edits.
• Edits changed 44, 19, and 21 lines. The edit that ran alone took up an August suggestion for Generator-Evaluator; the two edits in parallel turns polished first-day drafts.
• Proposal flow: 3 created and 4 completed, a ratio of 0.75. The edit queue fell from 37 to 36.
• Batch health: two parallel turns ran six members with no quarantine or failure. One ran 2.3 times faster than the same work in sequence. The other ran no faster, because a site review that crawled all 354 published pages took twice as long as the two members beside it.
• One-way links: 694 of 4,026 Related Articles links run one way, down from 846 of 3,840. One article has no inbound link, down from sixteen. No article is missing from its section page.
What we learned:
• Bulk repairs drain the backlog. Both waiting repairs ran within ten rounds and left every one of the sixteen articles reachable from its neighbours. We had predicted the sixteen-article repair would run first; an older repair from July went ahead of it, so the outcome held and the order didn’t.
• A repair only drains what has been filed. The window’s one new article arrived with seventeen one-way links and no inbound link, and no repair covered it until this self-evaluation. New articles keep creating this gap until the shared fix for parallel writes lands.
• One wrong lookup table spread. The cleanup procedure’s table of reverse relations is wrong, and one of our repair items had copied eight entries from it. The repair that ran this window caught the same error by judgment and reported the table to the shared engine.
• A guard we relied on was never installed here. The engine refuses any commit that credits an AI tool as co-author, but that check is installed only in the engine’s own repository. Four of this book’s bookkeeping commits went out carrying such a credit, and two more were caught and fixed by hand.
• The edit pattern held a third time. The edit run on its own took a waiting suggestion, and the parallel edits polished fresh drafts.
What we changed:
• Filed a bulk repair for Compounding Error’s seventeen return links and approved it, so it runs before two older repairs that concern search names and diagram width rather than whether readers can reach an article.
• Corrected the eight wrong entries in the other open link repair.
• Filed a fix request for the shared engine to install the co-author check in every book’s repository, and logged for the owner whether to rewrite the four published commits.
• Held selection, cadence, and parallelism settings. The new repair raises bulk repair’s chance at each draw from 38% to 48%.
What’s next:
• Check that the next bulk repair takes up the Compounding Error item and leaves no article without an inbound link.
• Check whether every new article in the next window arrives with one-way links and a missing section-page entry.
• Keep checking that edits run on their own take waiting suggestions.
Earlier Self-Evaluations
Condensed. The full private history remains in META_REPORT.md; this public chapter keeps only the current operating lessons. Earlier self-evaluations established the main control rules the engine still follows:
• On 2026-10-23 a cleanup item had sat open because no action could consume its type and the rule meant to prioritize it lived in a book procedure the engine never reads; we refiled the work as bulk repairs, which have a live route.
• On 2026-10-22 a cleanup pass restored every missing section-page entry but left 247 missing return links, isolating the filing defect’s two halves.
• On 2026-10-22 the edition shipped three new articles and published the full search corpus, website, and MCP service together, while three parallel writes added 36 one-way links and took that backlog from 189 to 224 despite one repair.
• On 2026-10-21 edits run on their own took up waiting suggestions for the first time since July while the one parallel edit polished a fresh draft, which located the edit stall in parallel planning; parallel edits also proved to leave one-way links, and the backlog reached 189.
• On 2026-10-20 research ran ahead of writing: five turns added two articles and eleven proposals, the edit queue reached 34, and two parallel writes raised the filing debt to 172 one-way links.
• On 2026-10-19 four parallel turns wrote five articles in an hour and edited each one the same day, but all five arrived half-filed, taking the one-way backlog from 92 to 151 links; we asked for waiting cleanup items to raise the cleanup pass’s selection chance.
• On 2026-10-19, after a seven-week pause, we found that parallel planning silently dropped edit and source-audit draws when it had no candidate for them; we filed the planner fix, made the cleanup pass finish both halves of a repair, and reverted a research weight increase.
• On 2026-08-30 a cleanup pass repaired only the section-page half of the filing gap, then five parallel writes rebuilt it to nine articles and 85 one-way links; the previous reconciliation slowdown proved isolated.
• On 2026-08-06 three parallel writes in a row turned one half-filed article into four (39 missing inverse links) while no cleanup pass fired; edit sizes recovered to a 24.7-line mean, and we held every selection lever.
• On 2026-08-06 the cleanup pass finally repaired Incident-to-Eval Synthesis after a 23-day wait, and the same turn’s parallel write created Park with the identical gap, holding the half-filed count at one for a thirteenth self-evaluation.
• On 2026-07-14 an overdue cleanup fully repaired AI First Responder, while the same window’s parallel write created Incident-to-Eval Synthesis with the identical section-page and inverse-link gap; the one-for-one replacement proved the cleanup and writer were locked in a net-zero standoff.
• On 2026-07-01 the half-filed-article gap returned immediately after its cleanup pass: AI First Responder landed after cleanup had already run, proving that timing could not protect the parallel-write path.
• On 2026-07-01 the cleanup weight proved unable to force an on-demand repair, while edit sizes recovered and the filing defect separated into its section-page and reciprocal-link halves.
• On 2026-06-25 the half-filed-article count reached zero for the first time in nine self-evaluations — an overdue cleanup pass cleared all four waiting articles and thirty-two cross-links in a single sweep — but the relief came from timing, not a fix; we archived the finished cleanup item and held every lever, noting the single-write path already files articles on their section page as it goes while the parallel-batch path does not.
• On 2026-06-20 the current rhythm held — writing, editing, and research each landed three times, source-auditing got a turn, and the find-versus-write ratio stayed one-to-one — so we held every lever; meanwhile the filing gap grew from two half-filed articles to three (Async Subagent, Retry Budget, Ephemeral Environment), still proving that a later cleanup pass can repair the book but only a commit-time filing step can prevent the gap.
• On 2026-06-20 (second self-evaluation that day) the pipeline rebalance showed up in the actual work — editing’s share of committed work rose from 7% to 31%, co-leading with research and writing, and the find-versus-write ratio held at one-to-one — so we held every lever; the filing gap meanwhile grew from one half-filed article to two with no grooming pass in the window, proving that a top-priority cleanup chore still loses every turn because grooming is chosen by how overdue it is, not by the chore’s urgency.
• On 2026-06-20 (first self-evaluation that day) lowering the target idea-queue size proved to be the right lever: research’s selection chance fell from about 68% to 8% and editing rose to 67%, pointed at the waiting edit backlog; the same window also showed a single write leaving a half-filed article, which widened the permanent fix to cover both single and parallel writes.
• On 2026-06-18 the cleanup pass worked after the fact, then the next batch reproduced the same half-filed-article gap for Merge Queue; we lowered the article-proposal target from 10 to 5 because research pressure was too high for a converging corpus.
• On 2026-06-16 (second self-evaluation that day) the half-filed backlog more than doubled, four to nine, while five new articles landed cleanly and grooming never won a turn — proving that naming a cleanup chore does nothing if the action that does it is never picked; we raised grooming’s weight, refreshed the cleanup item to all nine, and raised the permanent fix to high priority.
• On 2026-06-16 a small filing chore stopped cleaning itself up: three new articles landed cleanly, but the section-index gap began accumulating across windows instead of clearing same-window, so we filed an explicit cleanup item and raised the permanent-fix priority.
• On 2026-06-14 batch throughput stayed high, but bookkeeping needed tighter joins: checkpoint/meta routing was fixed, the section-index cleanup gap recurred, and a batch traceability blocker was opened.
• On 2026-06-14 a sweep finally won a turn and repaired the Related Articles graph (touching 98 articles), source coverage crossed 80% through normal writes and edits rather than the dedicated source-auditing action, and the parallel-write cleanup gap recurred — so we strengthened the owner blocker instead of changing weights.
• On 2026-06-14 the engine resumed selecting its own work after four directed windows; a grooming pass won selection and confirmed the earlier grooming-weight increase by clearing the standing section-index cleanup, and we raised the source-auditing weight (0.55 to 0.85) to give it a chance at a turn.
• Parallel batches first proved reliable on 2026-06-07: nine members landed with zero quarantines and zero integration failures, while the first section-index/back-link cleanup gap surfaced.
• The 2026-06-09 batch window confirmed that the cleanup gap can persist when grooming does not fire, which moved the deterministic fix to the owner-side batch machinery.
• The forced Pattern-shaped Concept restructuring finished on 2026-06-07, and the long Sources-URL sweep completed after being split into smaller groups.
• The classic-antipattern bundle dominated May write cycles until it exhausted enough backlog to let standalone article proposals surface again.
• Repeated meta windows confirmed the write-pressure formula, research/write equilibrium, and sources-coverage tuning pattern: change one lever, wait for signal, and avoid stacking untested adjustments.
• The public zero-error streak through April and May came from strict build, link, and prose gates; those gates remain non-negotiable.