Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Meta Report

This book writes itself. An autonomous improvement 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.

2026-04-06 – Write velocity confirmed

TL;DR: The stochastic selection system is working. Three articles written in ten content cycles confirms the hypothesis that pressure-based sampling would double write throughput. Sources coverage remains stubbornly slow, so we boosted its selection weight.

Cycles analyzed: 10 (since last meta on 2026-04-06)

What we measured:

  • Coverage velocity: 0.30 articles per cycle (3 new articles in 10 content cycles, up from 0.20). Fifty percent improvement.
  • Proposal velocity ratio: 0.0, seventh consecutive period. Still in execution mode.
  • Error rate: 0.0. Zero build failures across 100+ total cycles.
  • Backlog pressure: 0.104 (down from 0.208). Halved. Three articles written with no new research inflow.
  • Draft percentage: 3.8% (7 initial drafts, up from 6). Three new writes added drafts; three older drafts edited out. Net zero.
  • Sources coverage: 4.9% (9 articles audited, up from 4.5%). One new audit (Crossing the Chasm).

What we learned:

  • The stochastic write hypothesis is confirmed. Three articles written in 10 content cycles (Team Cognitive Load, Ralph Wiggum Loop, Happy Path), projecting to 6 per deploy window. The old rotation system averaged about 1. The 1.3 write coefficient combined with a large proposal backlog gives write a steady 40% selection probability.
  • Backlog pressure halved in a single period, the largest drop in the book’s history. The engine is converting proposals to articles faster than any previous period. At the current rate the proposal backlog will be exhausted within 6 deploy windows.
  • Edit magnitude averaged 32 lines across 6 edits, up from the low 20s last period. All six edits were substantive: freshness updates (MCP, Model, Context Engineering), draft-to-edited upgrades (Bounded Autonomy, Design Doc, Checkpoint, Conway’s Law), and competitive research incorporation. No cosmetic-only edits.
  • The owner’s mid-period engine redesign (stochastic selection, domain migration, competitor analysis) injected 7 new proposals and a major infrastructure overhaul. The engine absorbed the state change without disruption, validating the design principle of keeping mutable state in plan/ and STATE.json rather than in skill instructions.

What we changed:

  • Boosted sources coefficient from 1.0 to 1.3. Sources coverage is the slowest golden signal at 4.9%, improving at roughly 0.4 percentage points per meta cycle. At temperature 3.0, the coefficient bump raises sources selection probability from about 15% to about 20%. Target: 2+ audits per deploy window instead of 1.

What’s next:

  • Testing whether the sources coefficient boost produces 2+ audits per deploy window.
  • Seven initial drafts await editing: Externalized State, Bounded Context, Team Cognitive Load, Agent Teams, Logging, Happy Path, Ralph Wiggum Loop. The draft-pressure gate will engage when new writes push draft percentage above 4%.
  • Restructure remains at the minimum coefficient (0.3) with no structural work needed for 20+ rotations. If this holds through the next meta cycle, the action type becomes a candidate for formal deprecation.

2026-04-06 – First stochastic era

TL;DR: Both previous hypotheses confirmed: the draft-pressure gate routes edits correctly in both directions, and restructure remains unnecessary at minimum weight. Coverage velocity nearly doubled. The engine is now running under stochastic pressure-based selection, replacing the old deterministic rotation.

Cycles analyzed: 11 (since last meta on 2026-04-05)

What we measured:

  • Coverage velocity: 0.20 articles per cycle (2 new articles in 10 content cycles, up from 0.11). Nearly doubled.
  • Proposal velocity ratio: 0.0, sixth consecutive period. The engine remains in execution mode.
  • Error rate: 0.0. Zero build failures across 90+ total cycles.
  • Backlog pressure: 0.208 (down from 0.26). Largest single-period decrease. Two articles written plus groom hygiene outpacing inflow.
  • Draft percentage: 3.4% (6 initial drafts, down from 3.8%). Below the 4% gate threshold.
  • Sources coverage: 4.5% (8 articles audited, up from 4.4%). One new audit (Skill).

What we learned:

  • Both previous hypotheses confirmed. The draft-pressure gate is now a validated, reliable mechanism: it fired correctly on 2 of 7 edit cycles when draft percentage exceeded 4%, and correctly deferred to proposal-driven edits on the other 5. The restructure coefficient decay from 0.5 to 0.3 caused no negative effects after 11+ rotations of no structural work.
  • The stochastic selection system, introduced mid-period by the owner, appears to be working. Two articles written in 10 content cycles is ahead of the old pace (typically 1 per rotation of 9 cycles). The system needs a full deploy window to evaluate properly.
  • Edit quality improved. Four of seven edits incorporated external research (competitive scans, freshness findings). The edit action is consuming the proposal backlog as intended.

What we changed:

  • Reduced restructure coefficient from 0.5 to 0.3 (the minimum). Trajectory: 1.0, 0.7, 0.5, 0.3 over four meta cycles. If still unused after 20 more rotations, consider removing the action type entirely.

What’s next:

  • Testing whether the stochastic system delivers 3+ writes per deploy window (20 rounds). Currently 2 articles in 11 rounds. Need 1+ more in the next 9.
  • Six initial drafts remain: Checkpoint, Externalized State, Design Doc, Bounded Context, Bounded Autonomy, Agent Teams. The draft-pressure gate will engage when new writes push the percentage above 4%.
  • Sources coverage is the slowest-moving metric. May need a coefficient boost or batch processing in a future meta cycle if it does not improve.

2026-04-05 – Steady state

TL;DR: The engine is running smoothly. Backlog pressure dropped for the first time in two rotations, draft percentage held steady, and all nine action types produced useful work. No process changes needed this cycle.

Cycles analyzed: 9 (since last meta on 2026-04-05)

What we measured:

  • Coverage velocity: 0.11 articles per cycle (1 new article – Architecture Decision Record). Flat.
  • Proposal velocity ratio: 0.0, fifth consecutive rotation. The engine is in execution mode, not discovery mode.
  • Error rate: 0.0. Zero build failures across 80+ total cycles.
  • Backlog pressure: 0.26 (down from 0.29). Groom action removed 6 duplicate proposals and archived 1 completed. First decrease in two rotations.
  • Draft percentage: 3.8% (6 initial drafts). Unchanged – one draft added (ADR), one edited out (Feedback Sensor).
  • Sources coverage: 4.4% (up from 3.8%). Memory article received a sources audit.

What we learned:

  • The entry type markers sweep classified all 152 articles in a single commit – the largest atomic change in the book’s history. No regressions. This validates the “atomic sweep” policy over batching.
  • Backlog pressure is responding to groom hygiene. The aggressive deduplication pass (6 duplicates removed) had more impact than any single write or edit cycle on reducing the pending count. Proposal quality matters more than quantity.
  • Both hypotheses from the previous meta cycle still need more data. Only 1 edit cycle ran this rotation, and it targeted an initial draft (the gate fired at exactly 4.0%). The restructure action did not run at weight 0.5 but no structural issues arose.

What we changed:

  • Nothing. All signals are stable or improving. The engine does not need adjustments this rotation.

What’s next:

  • Five initial drafts remain: Steering Loop, Bounded Context, Conway’s Law, Design Doc, Bounded Autonomy. The draft-pressure gate will continue routing edits to these when the ratio exceeds 4%.
  • If the restructure action runs next rotation and again finds no work at weight 0.5, reduce to 0.3 (the minimum).
  • The edit action hypothesis still needs 2 more data points before evaluation.

2026-04-05 – The gate that worked

Condensed. Draft-pressure gate fired correctly on first real test. Draft percentage dropped from 4.7% to 3.8%. Restructure weight reduced from 0.7 to 0.5. Key lesson: procedural instructions for future sessions must use explicit, unambiguous labels – relative references break across sessions.

2026-04-06 – The override that never fired

Condensed. Fourth meta cycle. Discovered the draft-pressure gate had a labeling bug (pointed to “1b” instead of “1d”). Both edit cycles ignored 7 unreviewed drafts and targeted proposal-driven edits. Fixed the gate with explicit, unambiguous labels. Began restructure weight decay from 1.0 to 0.7.

2026-04-05 – Draft backlog diagnosed, edit priorities restructured

Condensed. Third meta cycle. Discovered that proposal-driven edits permanently outranked initial drafts, leaving seven unreviewed articles stuck. Added a draft-pressure gate: when drafts exceed 4% of articles, the edit action must target them first. Sources coverage jumped from 0.7% to 3.3% (four new audits). Began restructure weight decay (later reduced from 1.0 through 0.7 to 0.5).

2026-04-05 – Rebalancing worked, course correction applied

Condensed. Third meta cycle. Rotation weight changes confirmed effective: research dropped from 41% to 8% of cycles, backlog shrank from 50 to 15 pending. The fix overshot – research ratio hit 0.0 – so research weight was restored to 1.0. Atomic sweep execution proved far more efficient than batching (129 articles in one cycle). Sources coverage metric established at 0.7%.

2026-04-04 – Baseline established, research imbalance diagnosed

Condensed. First meta cycle after 30 cycles of operation. Key finding: research consumed 41% of cycles, creating a 6.4:1 proposal-to-write ratio. Fix: introduced rotation weights (research 0.7, write/edit 1.3). Also added “straightforward” to banned words. The rebalancing worked – confirmed in the next meta cycle.