# Reharm architecture: the fresh local check

Measured on **Jacks-iMac**, 2026-10-17, against the September 10 Claude Code audit.
Full remediation prompt and plan: `~/Projects/_outputs/reharm-architecture-remediation/REHARM-REMEDIATION-PROMPT.md`

## The four verdicts

| Verdict | Claim | Measured |
|---|---|---|
| **Held** | per-turn context cost ~556,000 | **625,521** cache-read per assistant turn |
| **Wrong target** | edit `run_gates.py`, `completion_stop_hook.py`, `no_stop_gate_hook.py` | none of the three exists; real chain is **16 Stop hooks** (14 Python gates plus 2 shell one-liners), and **all 14 Python gates can block** |
| **Worse than stated** | CLI imports and inspects, 3 commands missing | **zero bytes of executable Python**; `SKILL.md` line 41 already marks it not implemented |
| **Unverified** | 12 sessions, 1,610 turns, 496,941,520 tokens | Sept 10 here: **1 session, 70 turns, 43,786,490 tokens**; MacBookPro unscanned |

## Context baseline

| Item | Audit | Measured |
|---|---|---|
| Skill descriptions | 23,134 | **41,486** (169 enabled, 150 described) |
| Project CLAUDE.md | 4,125 | 3,535 |
| MEMORY.md | 10,438 | 4,968 |
| Global `~/.claude/CLAUDE.md` | 2,704 | **absent** |
| **Auto-loaded total** | **40,401** | **49,989** |
| POINTERS.md | not counted | **53,281**, retrieval only |

`POINTERS.md` is larger than the whole audited baseline and is correctly kept out of it, reached by grep. It is the pattern the skill roster should copy.

## Counts are a snapshot

The first read at 13:0x returned 17 Stop entries; a re-read at 13:41 returned 16. `settings.json` was written at 13:28:06 by one of the two other live sessions, leaving `settings.json.bak-before-guitar-evaluator-off-2026-10-17`. The removed entry was **an LLM prompt-judge**, a `{"type": "prompt"}` hook rather than a script, turned off at Brandon's request; verbatim copy at `hooks/DISABLED-guitar-completion-evaluator-2026-10-17.json`. So Stop held **17 executable hooks before and 16 after**. An earlier pass of this file called it a blank placeholder, which was wrong: the extraction read each hook's `command` key, and a prompt-judge hook has none.

## The 17th hook was the audit's own thesis

That prompt-judge instructs that if Claude "says it is stopping, pausing, queuing work, leaving an item for later, asking whether it should continue" it must return INCOMPLETE and direct Claude to continue, with its reason fed back as Claude's next instruction. A `type: "prompt"` hook also spends a full model call on every stop, which no script in the chain does. The session log records it looping about 12 turns on a blocked external dependency before being switched off, which is that session's report rather than a measurement here. The audit asked for recursive stopping to be removed from two scripts that do not exist; the real instance was this, and Brandon had already removed it.

The 14-gate blocking census is unaffected and all 14 Python gates are wired to Stop.

## My own census was wrong, in the direction that understated the problem

An earlier pass said 11 of 14 block. That grepped for a `"decision"` field and a literal `exit(2)`, missing the `return 2` plus `sys.exit(main())` form. Three gates block that way and were miscounted as silent: `completion_stop_gate.py`, `chat_link_gate.py`, `chat_donelink_gate.py`. Five of the fourteen were then observed blocking during this review, including two of those three.

## The chain does not short-circuit

On one Stop event this session, the gates at positions 5, 6, 8 and 13 all blocked and all reported, so execution ran past the block at 5 and continued to 13. One unmet stop can collect refusals from up to fourteen gates, each costing a fresh turn carrying the whole conversation. Bounding only the first blocker would achieve nothing.

## completion_stop_gate is already solved

It blocks with `return 2` through `sys.exit(main())` and writes its reason to stderr at line 302. The line-301 comment documents a bug **already fixed** on 2026-09-14, with `gate_repair_regression_2026_09_14.py` passing today at exit 2 with 596 characters of stderr. It is already bounded at `HARD_CAP = 25` with a `COMPLETION_GATE_CAP` override, and it carries the fix for the trap the others will hit: every counter once read 1, because refusal feedback was itself counted as a fresh user record. The real target is the **13 gates with no bound**, copying this one.

## The engine

Eight empty `__init__.py` files. Empty `bin/`, `scripts/`, `tests/{unit,property,integration,golden}`, `examples/reports`, `fixtures/{incomplete_evidence,midi_only,gp_full}`. Real and usable: `reharm-ir.schema.json` at 16,916 bytes plus three reference documents. The build is greenfield against a settled schema, which is cheaper than the repair the audit assumed.

## Three-phase plan

1. **Bound the supervisor, shrink the baseline.** ONE session-wide retry budget shared by every gate, copying `completion_stop_gate`'s `HARD_CAP` mechanics. Thirteen separate counters at two retries each would admit up to twenty-six extra turns, and the chain does not short-circuit, so all thirteen fire on the same event and per-gate ceilings multiply where they should cap. Each gate records its reason against the shared budget; when it is spent, the chain reports every unresolved condition and lets the session end, with failed musical verification staying marked failed. Move trigger lists out of front-matter. Exit: alternating-reason test still ends the session within budget, baseline under 25,000.
2. **Build the engine, tests first.** `doctor import map generate write-midi verify`; `doctor` exits 0 only when the five run on the fixture. Tests fail on their recorded bad case before being trusted. Exit: `bin/reharm doctor` exit 0.
3. **One song, then the MacBookPro scan.** One Shiner song, three distinct candidates, melody and bass preserved, exported, reopened, tested by bar and pitch. Exit: live page at http 200, fraud check green.

## Accuracy boundaries

File token counts are chars/4 estimates, within roughly 10%. Transcript token counts are exact from each record's `usage` field. Retention is not a factor: `cleanupPeriodDays` is 36500, 98 dates present from 2026-05-05 to 2026-10-17, all thirty September days included.

Unverified and labelled as such: the twelve-session count, the 1,610-turn count, the 496,941,520-token total, the 251 blocked stops, and the 204-of-251 attribution.