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.

2026-07-14 – One half-filed article cleared, another created, in the same window

TL;DR: This window showed, as cleanly as possible, why the temporary workaround can never win. The cleanup pass finally came due – after about thirteen days of waiting it won a turn and fully repaired the article left half-filed last window (AI First Responder): it added the article to its section page and reciprocated all ten of its cross-links in one sweep. That confirms exactly what we predicted – the cleanup pass runs on how overdue it is, not on any weight we set. But ten minutes earlier, the same window’s parallel-batch write created a fresh, fully half-filed article (Incident-to-Eval Synthesis): missing from its section page and missing all six of its cross-links. So the window cleared one and made one – the count of half-filed articles held at exactly one – and the cleanup clock has already reset, so the new one will wait about two weeks in turn. We also corrected last window’s claim that new articles now reach their section page as they are written: that holds only for single writes; the parallel-batch path still files neither the section-page entry nor the cross-links. We recorded the twelfth straight observation in the owner blocker, closed out the finished cleanup item, filed a new one for the new gap, and changed no levers.

Cycles analyzed: 5 turns / 11 work-units (since the previous self-evaluation on 2026-07-01).

What we measured:

• Actual work mix: 4 research findings, 3 critiques, 2 edits, 1 write, 1 cleanup pass. A scouting-heavy window with one decisive cleanup.
• New articles: 1 (Incident-to-Eval Synthesis); the corpus stands at 304, with source coverage flat at about 81.9%.
• Half-filed articles: 1 – but a new one (Incident-to-Eval Synthesis), not the leftover. The leftover (AI First Responder) was cleared this window.
• Edit sizes: 5 and 5 lines – two deliberately light touches (a paragraph split and a post-write tidy), so read as a small sample, not a sign the corpus is running out of work.
• Balance of finding versus writing ideas: 2 found, 1 written – at the two-to-one level we watch, but the article queue drained to empty with no sign of research crowding out writing.
• Batch health: no quarantines, no integration failures, no build errors.

What we learned:

• The cleanup pass and the parallel-batch writer are locked in a stalemate that nets to zero progress. A batch write creates a fully half-filed article every time; a cleanup pass repairs one only every couple of weeks, when it is finally overdue enough to win a turn. The count will stay at one or more until the permanent engine fix lands.
• The section-page half is not handled as-you-write on the parallel-batch path – only on the single-write path. The batch writer defers both the section-page entry and the cross-links, so the permanent fix has to cover both, at the moment the batch is committed.
• Both of last window’s predictions held: the cleanup pass is chosen by how overdue it is (the thirteen-day wait proves it), and the cross-link half of the gap recurs on its own (this window it recurred alongside the section-page half).

What we changed:

• Recorded the twelfth straight observation in the owner blocker, with the stalemate evidence and the correction that the parallel-batch path files neither half.
• Closed the finished cleanup item (the pass did its work but had not marked itself done) and filed a fresh one for the new half-filed article.
• Held every selection lever: no errors, clean batches, steady coverage, and the cleanup weight is spent.

What’s next:

• Watch whether the new half-filed article is still waiting at the next self-evaluation, unless enough days pass for the cleanup pass to come due or the permanent fix lands.
• Watch whether edit sizes return above ten lines on the next window with more than a couple of edits.

2026-07-01 – The cleanup weight is spent, and the filing gap turns out to be two problems

TL;DR: Two findings this window settle a long-running debate. First, the cleanup pass never ran once – across the whole window – even though its selection weight was already raised specifically to force it and a top-priority cleanup chore was waiting the entire time. The reason is now plain: the cleanup pass is chosen by how overdue it is, not by how urgent the waiting work is, and a recent cleanup had just reset that clock. So the article left half-filed last window stayed half-filed, and no weight we can set will change that. Second, the gap is really two separate problems. This window’s own new articles (Context Editing and GraphRAG) landed correctly on their section pages – that half is fixed as the article is written now – but their cross-links to neighboring articles were still left one-directional. Fixing one half does not fix the other. We recorded the eleventh straight observation in the owner blocker, confirmed that edit sizes recovered (last window’s dip was a small sample), and changed no selection levers, because another cleanup-weight bump would plainly do nothing.

Cycles analyzed: 5 turns / 11 work-units (since the previous self-evaluation on 2026-07-01).

What we measured:

• Actual work mix: 4 edits, 3 research findings, 2 writes, 2 critiques, and zero cleanup passes. Editing led, as its saturated backlog and dominant selection chance predicted.
• New articles: 2 (Context Editing, GraphRAG); the corpus stands at 303, with source coverage flat at about 81.8%.
• Half-filed articles: 1 – but a leftover (AI First Responder), not a new one. It stayed because the cleanup pass never ran.
• Edit sizes: 9, 32, 13, and 6 lines – an average of 15 across four edits, back above the ten-line mark and up from last window’s two-edit average of 7.5.
• Balance of finding versus writing ideas: 4 found, 2 written – 2 to 1, at the level we watch, but with the article queue nearly empty and no sign of research crowding out writing.
• Batch health: no quarantines, no integration failures, no build errors.

What we learned:

• The cleanup weight is spent. Raising it last window did not make the cleanup pass run even once this window, because that pass is picked by how overdue it is, not by how important the waiting chore is. A recent cleanup reset the clock, so the top-priority chore lost every turn. No weight change can force an on-demand cleanup; only the permanent engine fix will.
• The filing gap is two problems, not one. New articles now reach their section pages as they are written – that half is handled. But their cross-links to neighbors are still left one-directional, so half the gap remains even when the section pages are right. The permanent fix has to close both.
• The edit-size dip was a small sample. Four edits averaging fifteen lines confirm last window’s 7.5 was noise, not a sign the corpus is running out of substantive work.

What we changed:

• Recorded the eleventh straight observation in the owner blocker, with the evidence that the cleanup weight is exhausted and the framing that the gap is now two distinct problems.
• Held every selection lever. With no errors, clean batches, steady coverage, and edit sizes recovered, another tweak would only add noise.

What’s next:

• Watch whether the leftover half-filed article stays that way until the cleanup pass is overdue enough to win a turn on its own – more evidence for how the pass is chosen.
• Watch whether the next new article again leaves its cross-links one-directional even when its section page is correct.

2026-07-01 – The half-filed article came back the very next window, and timing did not save it

TL;DR: Last window’s clean result held for exactly one grooming pass. This window a grooming pass cleared the three articles left half-filed by the previous batch, and two minutes later the next batch write left a new one: AI First Responder landed in the table of contents but not on its section page, with all ten of its cross-links still one-directional. The telling detail is the timing. The gap opened after the grooming pass had already run, so being overdue for grooming is no defense at all – the parallel-batch machinery re-creates the gap faster than any cleanup pass can close it. That settles the last doubt: only the permanent engine fix will do, and this is the tenth self-evaluation to say so. We filed a fresh cleanup item, updated the owner blocker with the tenth observation, and confirmed two of our three standing predictions while setting the third aside as a small-sample blip. We changed no selection levers.

Cycles analyzed: 5 turns / 13 work-units (since the previous self-evaluation on 2026-06-25).

What we measured:

• Actual work mix: 4 research findings, 3 writes, 3 critiques, 2 edits, 1 grooming pass. Research led and fed the writing: a concept scan produced Agent Drift, and a competitor scan produced AI First Responder.
• New articles: 3 (Loop Engineering, Agent Drift, AI First Responder); the corpus stands at 301, with source coverage flat at about 81.7%.
• Half-filed articles: 1 (AI First Responder), up from zero – and produced after the window’s own grooming pass.
• Edit sizes: 7 and 8 lines – back to light polish, but on only two edits against an edit backlog that has grown to thirteen.
• Balance of finding versus writing ideas: 5 found, 3 written, which is 1.67 to one – recovered from last window’s 3-to-1.
• Batch health: no quarantines, no integration failures, no build errors; the deploy succeeded.

What we learned:

• Being overdue for grooming is no defense. The gap opened after grooming had already run this window, so the parallel-batch path re-creates it independent of when the cleanup pass comes up. The single-write path is safer now, but this article came through the batch path, where the safeguard does not reach.
• The find-versus-write balance recovered on its own, falling from 3-to-1 back toward even as writing picked up and the article queue emptied – exactly the one-window artifact we predicted, with no lever touched.
• The edit-size dip is a small sample, not a plateau. Two light edits against a growing backlog does not mean the corpus is out of substantive work; editing dominates the next turn, so the coming window will show the real size.

What we changed:

• Filed a fresh cleanup item to add AI First Responder to its section page and reciprocate its ten cross-links.
• Updated the owner blocker with the tenth straight observation and the after-grooming timing evidence, which is the clearest case yet for the permanent fix.
• Held every selection lever. With no errors, clean batches, healthy coverage, and a balanced pipeline, another tweak would only add noise.

What’s next:

• Test whether the next batch write re-creates the gap again while the permanent fix is still pending. It should.
• Test whether edit sizes climb back above ten across more edits as the backlog draws down.

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 2026-06-25 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.