1. Three numbers to watch

If these move the right way, the path is working. If they don't, no explanation should override the numbers.

1a. Share of work through the new system

Turns handled end-to-end by the new system (strict measure)

No data yet. no published strict end-to-end "turns through new code" metric artifact yet — the stricter full-path number is not generated by the current scoreboard

Honest substitute shown below (writes-metric). The strict number must climb toward 100% once the full-path metric is published.

1.1%
Turns whose write evidence points to the new system (this is the writes-metric, not the strict full-path measure)

26 new-code writes · 705 old-code writes · 1623 not measured · out of 2354 turns in cycle 324f4ab3-c63-20260802. Of turns where code path could be measured: 3.6% new.

Source: latest test-trace cycle, counting which code performed the write (writes-metric — not the stricter full-path metric)

1b. Old vs new pass rates (lines will cross)

Pass rates by which code handled the turn — the "lines will cross" view. New code should catch and then beat old code, category by category. Percents use only graded pass/fail rows in each code group; "no evidence" rows are excluded from the percent.

Test category Overall pass rate New-code pass rate Old-code pass rate Could not classify
Adversarial
w2h-wave3adv-grade.json
30%
30 pass / 53 fail
50% (5 pass / 5 fail) 45% (9 pass / 11 fail) 197
Clarifying questions & declines (conversational)
fanout-conv-w8-20260718-grade.json
11%
20 pass / 2 fail
no graded rows 100% (2 pass / 0 fail) 20
Compound dishes
fanout-compound-dishes-A-fab27-20260718-grade.json
7%
1 pass / 0 fail
no graded rows 100% (1 pass / 0 fail) 0
Food A
food-corpus-a-324f4ab3-c63-20260802-grade.json
74%
148 pass / 48 fail
no graded rows 86.4% (102 pass / 16 fail) 81
Food B
food-corpus-b-FROZEN-20260712-324f4ab3-c63-20260802-grade.json
20%
59 pass / 27 fail
no graded rows 90.8% (59 pass / 6 fail) 21
Food commands by type (log/update/delete/query)
fanout-foodtyped-w8-20260718-grade.json
10%
10 pass / 3 fail
no graded rows 90% (9 pass / 1 fail) 3
Food mixed C (everyday)
food-corpus-mixed-C-324f4ab3-c63-20260802-grade.json
60%
121 pass / 48 fail
0% (0 pass / 1 fail) 88.9% (88 pass / 11 fail) 97
Food mixed D (everyday)
food-corpus-mixed-D-324f4ab3-c63-20260802-grade.json
56%
113 pass / 56 fail
0% (0 pass / 1 fail) 82.8% (77 pass / 16 fail) 104
Lists & settings (new-types)
fanout-newtypes-w8-20260718-grade.json
10%
9 pass / 0 fail
no graded rows no graded rows 9
Multi-item meals
multifood-corpus-a-324f4ab3-c63-20260802-grade.json
63%
38 pass / 14 fail
no graded rows 66.7% (18 pass / 9 fail) 26
Notes/todos/lists
fanout-notes-w8-20260718-grade.json
9%
13 pass / 4 fail
no graded rows no graded rows 17
Query frozen (exact keys)
query-frozen-A-324f4ab3-c63-20260802-grade.json
92%
72 pass / 6 fail
no graded rows no graded rows 78
Utility
fanout-util-w8-20260718-grade.json
9%
4 pass / 0 fail
no graded rows no graded rows 4
Workout
fanout-workout-w8-20260718-grade.json
9%
31 pass / 2 fail
no graded rows no graded rows 34
Workout commands by type
fanout-workouttyped-w8-20260718-grade.json
3%
3 pass / 0 fail
no graded rows no graded rows 3

1c. Instruction checks still green

7 of 7
Standing-instruction checks currently green

Recomputed from source at page build time. A red check that stays red past one cycle means the machinery is being overridden.

StatusCheckHoward's instruction (as written)
green — check passed Keep working on the new code first
new-code-first
Keep working on the new code, make it better, the lines will cross. Old code is frozen reference only.
green — check passed Legacy main app file stays frozen
apptsx-frozen
App.tsx is radioactive: nobody touches it ever again, for any reason, except its eventual deletion.
green — check passed Only Howard dismisses defects
only-howard-dismisses
Only I dismiss defects.
green — check passed Cardio stays in the backlog
cardio-backlogged
I don't want to focus on cardio logging right now; it belongs in the backlog.
green — check passed Defaults and history always on
defaults-history-always-on
We definitely always want to use defaults and history — can't imagine when we want that off.
green — check passed Direction changes must be written down
direction-changes-surface
Nothing gets decided against my instructions without telling me. (2026-07-21)
green — check passed Unlogged direction changes are violations
direction-change-probe
An unlogged reverse/deviate/reinterpret of a standing instruction is itself a violation — Layer 3 must detect measurable deviations and require a matching DIRECTION_CHANGES authorization (2026-07-24 charter / DIRECTION-CHANGE-PROBE).

Module scores (continuous testing)

Continuous testing — each module on its own. We never blend food + workouts + notes into one fleet number. Every completed corpus run also appears in the corpus-runs list as when it ran / which build / what score. (Legacy nightly history rows, if any, are readback only.)

Food: n/a · Workouts: n/a · Notes: n/a · Query: n/a (build e0a1efe8, 0 of 17 scripts complete)

Source: continuous module snapshot + legacy history (same as the machine DETERMINISTIC box). Full when/build/score list: corpus-runs.html.

2. Decision log

Headlines from the daily session book (today and yesterday). Each line is a real decision or discovery recorded the same hour it happened.

2026-08-01 from docs/reset/For Other Codex/SESSION_BOOK_20260801.md

  1. HOWARD-REPLY pickup
  2. ~6:35 PM ET — Howard ruling: old unclustered rows are NOISE — purged
  3. ~7:20 PM ET — Model routing decision (Howard)
  4. 08-02 ~1:15 AM ET — Howard live sighting: clarify-by-macro-cluster doctrine

Also checked:

3. Direction changes

Any decision that reverses or reinterprets a Howard instruction must be written here the same day it is made.

DateWhat changedWhich instruction it affects Who decidedTold Howard?
2026-07-21 Fleet builds force food module LIVE; App.tsx frozen at dispatch new-code-first (restores it) conductor, executing Howard's standing order this ledger + chat, same day
2026-07-22 App.tsx declared RADIOACTIVE — delete-only, no line-count ratchet, reading-as-reference allowed; enforced at spec-lint + git pre-commit hook + register probe apptsx-frozen (strengthened) HOWARD this ledger + chat, same wake

Source: docs/reset/For Other Codex/DIRECTION_CHANGES.md · showing 2 of 2 row(s)

4. Legacy-file exception permits

The old main app file is frozen. A one-time exception can only land with a written permit. Status words: granted (approved, not locked yet) · armed (exact change locked, ready once) · consumed (already used) · expired (timed out).

P1-spine-adapter Expired

This one-time exception ran out of time without being used.

Granted by: Howard, chat 2026-07-23 ~9:15am ET (AskUserQuestion: 'Yes — grant it')

Scope: REVISED by Howard 2026-07-23 ~11:35am ET (supersedes shadow design): single DEFLECTION change — utterances are handed to the new command code and the old ladder does NOT process them; one env-driven setting (same pattern as the food-module flag) whose off-position restores legacy behavior as an emergency revert only; no side-by-side comparison mode. Howard verbatim: 'the one that deflects the utterances so the old code doesn't process them, and the new code is what processes them. New code. That is what we should test. If it's broken, we'll fix it.'

Expires: 2026-07-26T13:15:00Z

Used in commit: not yet

Internal status field: awaiting_diff

Source: scripts/lane-h/APPTSX_PERMITS.json

5. What changed since yesterday

One line per commit since 2026-08-01 (Eastern midnight), newest first.

  1. ce267c9dce NIGHT_LOOP: restart long-running units after merging machinery fixes — the scheduler ran 4-day-old code all day and silently negated the lane fix
  2. b8590ecc6a proof launcher: keep the proof lane full without the conductor (the 08-02 idle gap proved every wake mechanism inherits the conductor's failure mode)
  3. d23255c72a restore exec bit on proof driver
  4. 811ad9571f parallelism bug: a proof's startup sweep destroyed sibling runs' simulators (same branch = same slug); now it keeps any sim claimed by a live run
  5. 8150917b14 NIGHT_LOOP: never edit a running shell script — zsh reads incrementally and my mid-run edit executed garbage in two live proofs
  6. da294ebe52 rebuild PENDING_MERGES board from ground truth (stale since 07-31 — caught by the verdict triage)
  7. a4af533bfc CLUSTERING_SOP rule 11: assert the trace-join matched before trusting a diff (a zero-match join looks identical to a clean flat verdict)
  8. ac6d49a29e measurement bug: proof artifacts now keyed by run LABEL — same-branch+corpus reruns were silently overwriting each other's traces, producing false 'flat 0/0' verdicts
  9. 180f1c6310 fix(lane-runner): finish corpora through row timeouts
  10. 7da8b6fe98 untrack CURRENT_BUILD.json: it holds a machine-specific absolute appPath and blocked both new hosts (each host stamps its own)
  11. f02b70f990 make the portability fix actually work: DEFAULT_STACK_DIR honors VOICECOACH_APP_DIR and the driver passes --stack-dir explicitly (both new hosts were blocked because the export was inert)
  12. 8648605ed9 stop the 19-day bus flood: disable the retired blocked-goal guard inside message-delivery (65,121 alarms since 07-14, ~315/hour, from a keeper unit that still carried a piece of the shut-down machine)
  13. 85a7bc991f host-portable proof driver (VOICECOACH_MACHINERY_DIR / VOICECOACH_APP_DIR) — the laptop and Mini could not run proofs because Studio paths were hardcoded
  14. a67d356938 restore agent-comms-sender-policy.mjs (silently missing; broke ALL fleet messaging to the new hosts)
  15. e8c8a9bbac NIGHT_LOOP: proof-launch preflight (pool accounts only, verify corpus path, require a complete baseline) after three launches burned on these
  16. 0f1bfe6af0 defect states are DERIVED, not hand-maintained (Howard 08-02): reconcile-defect-states.mjs computes state from git + running processes, runs automatically at the end of every proof and at every conductor wake; never un-merges
  17. c44cec4c4f NIGHT_LOOP: document the three wake mechanisms and which one actually fires (scheduled-tasks MCP silently dead 23h; in-session cron + background watcher are real)
  18. 6fea2193ed SESSION_BOOK + memory: Howard's clarify-by-macro-cluster doctrine (08-02 pineapple juice sighting)
  19. fbb3c7ab22 every worker prompt now opens with the MISSION: the rows are a sample of a class, not the goal — cheating and hard-coding are incoherent with the actual objective (Howard 08-02)
  20. 94c877d8a7 brief-writer rule: word grading units as judge problems, never 'make N rows pass'; state behavior-neutral invariant for refactors
  21. 1226b7a4d7 integrity block: grading-unit rules — objective is a correct judge not passing rows, expect MORE failures, blast radius required, no assertion edits, no invented thresholds
  22. ccc0a3e0db NIGHT_LOOP: safe sim-sweep method (UDIDs from running proof logs, never name matching) after a sweep killed two live proofs
  23. bfd95d1fc8 NIGHT_LOOP.md: the ordered wake checklist (refill first, verdict method, frozen-key law, known noise, full-run reporting format)
  24. 7c041e261a worker integrity block in every dispatch prompt (Howard 08-01): review stated up front, rejected patterns named, honorable exit granted, self-authorized key/assertion changes forbidden
  25. 58d9a9b828 evidence: record sequential four-row corpus proof
  26. ef86d796ae corpus: add decline-estimate and filtered list coverage
  27. a5a9590e9c operating rules (Howard 08-01): verification gate by change type (machinery-only = regrade diff, app = full-corpus sim proof) + proof parallelism with max-4 limit
  28. d1a159a6ac grader: mechanically score OTHER response subtypes
  29. 56fda8da7c grade ambient speech by durable behavior
  30. b67373554b SESSION_BOOK: conductor to Opus 5, Fable as authorized loud escalation tier (Howard 08-01)
  31. a22cb9ece0 grader: canonicalize saved food identities
  32. a9ef2f4466 packet launcher experiment (Howard-approved 08-01): launch-only 2-slot deterministic lane, one-shot packets, area-skip, heartbeat + kill file; briefs 03-08 approved+tagged
  33. 958e2fe5ed SESSION_BOOK: Howard ruling — stale unclustered rows purged; current-cycle failures are the only reportable failures
  34. 5da7d0cb86 CLUSTERING_SOP rule 9: proof baselines must share the branch parent — stale baselines attribute trunk changes to branches
  35. d0a7de348a story: weave Howard's factory/human-management analogy through all five acts (one worker one car, retooling the floor, plant-manager SPOF, bureaucracy, lights-out factory, compliant lying, foreman shop, three AI differences)
  36. a5d042010f the story version: All I Ever Wanted Was for It to Work — Fable-narrated five-act account from mined transcripts
  37. 8261d8cbaf refill-before-review law + dispatch-ready brief queue (Howard 08-01)
  38. dbe47dd024 CLUSTERING_SOP rule 8: stability windows lie for general-fallback rows — check the answering path
  39. 3721a7445a proof driver: recording tripwire — food runs with zero resolver events refuse to grade (watched refusing on dark baseline, passing on restored-telemetry proof)
  40. f8448b1f46 AI fleet docs: technical rebuild manual + article (Howard's two-document request; article renamed from lessons doc)
  41. f297c772f0 Six weeks of AI fleet lessons — Howard-requested retrospective + operating recipe (for humans and their AIs)
  42. 1e8518e10e fix run page completion gate
  43. 1d6212879e submit-note confirmation appears at the button (Howard 08-01), current-cycle pages regenerated
  44. 43c8f75fce report lockdown: violations alert-and-continue, never block new publishes (Howard 08-01)
  45. 4b83571936 publish: migrated permanent run archive live on Cloudflare, guard drill passed
  46. d86f00e465 snapshot: live lane generated outputs before publisher-rebuild merge
  47. c29c54fe9d reports: make run pages permanent and paginated
  48. 8d3acbd4c9 port Howard's round-2 corpus verdicts APP->MACHINERY (byte-identical sync of 7b131d2eb)
  49. 6727b837f8 run-history page (Howard 08-01: historical runs invisible — the testing report keeps a rolling 12-run window and fast cycles pushed last night off within hours): every archived graded run, newest first, filterable; wired into the publisher
  50. 4b67a4a66b hourly floor check (Howard 08-01 after a 6h45m overnight stall): the four questions he asks — workers busy? anything finished unprocessed (incl. reports orphaned in worktrees)? anything to cluster? fix units unassigned? — as a scheduled hourly wake, no daemons
  51. a2fe438a60 land the wave-4 analysis report (was written in its worktree and never copied into the repo — its 7 families were therefore invisible to the queue)