The v37 run

Kind of page: explanation: the previous run, for comparison.

The aggregator's previous run: gioia-j-sonnet-v37-20260908. It ran contract v37, the version before this repository's v38, and the v38 design work began by reading its handler log.

Not shipped. This repository carries one full run, v38's; earlier runs are described, not included. The v37 instrument's texts are in the history bundle, under aggregator/v37/.

What it was

One analyst, a Claude Code session on J's Mac, worked inside 00_GIVE-THE-AGENT/ and saw nothing else. A handler, a Cowork session with the whole run folder mounted, closed stages, took checkpoints and answered the analyst in files. J opened and closed the analyst's session in Terminal and made the author's judgements. The handover of 2026-09-08 records that "every role is on claude-sonnet-5", instrument v37 with fingerprint 2e1e4686828a3c1e…, eighteen stages in the older numbering (01 to 18), declared fork points at the close of 01, 06 and 09, and per-stage token ceilings at 04, 10 and 14.

Three defects stopped the run from starting and were fixed before stage 01: the fence denied the analyst its own folder, the analyst inherited the account's claude.ai connectors, and its config folder had no login. No fork was taken at 01.

What it reached and learned

The handovers stop at stage 02. The resume notes and the handler's close-out, which sits in the run under CONTROL/handler-log/ and in the Claude project, carry it further: the run closed on 2026-09-21 after all eighteen stages, with eighteen committed checkpoints from start to step-18. Stages 07, 08 and 09 ran in one session and built the model; five model files are byte-identical from the stage 09 snap to the stage 18 snap.

What the close-out carried forward to v38:

• The stage file and the close-check were frozen apart; six validator amendments followed.
• Packets must be sized against the reading tool: packets of 142 to 329 KB failed 21 times at stage 14; rebuilt at about 50 KB, all fourteen returned first time.
• Build, dispatch, check: "a launch that dispatches does nothing else; a launch that thinks dispatches nothing".
• The cheap reader fleet did 87 per cent of the tokens for 12 per cent of the cost on three metered stages.

It also left open questions: which null model is declared first at stage 17, 290 of 2,703 case themes with no placement, and two ordering claims whose snaps were deleted and retaken.

How to resume or fork one of its steps

From RESUME/README.md in the run folder:

1. Run RESUME/open-stage.sh for step-09, the close where "the model is freshest", or open-stage.sh step-10 for another close. It restores the checkpoint into a new dated folder under ~/Jail/interviews/, verifies it file by file, prints the session id and opens the Claude Code CLI. NO_OPEN=1 restores without opening.
2. Paste PREAMBLE.md as the first message. Without it the analyst tries to run the next stage.

What the README says to know first:

• It runs on the subscription only, not through OpenRouter; a session cannot change route once served.
• It works only in the Claude Code CLI; the desktop app, the web and the editor extension keep their own session stores.
• Each restore is a fresh copy, about 165 MB for step-09, and took 43 seconds in the test of 2026-09-22.
• On macOS the restore needs umask 077, which open-stage.sh sets; restore_stage.sh creates two folders before extracting and their modes otherwise fail the check. RESUME/diagnose.sh names the tree that differs.
• Every checkpoint carries session.boundary_verified: false; the restore checks the boundary itself and refuses on a mismatch.
• The restored folder holds the stage files for 10 to 18. The sandbox denies them to the analyst, not to a person at the terminal.

RESUME/NOTE-TO-THE-VISITOR.md is the message to send someone who will talk to the resumed analyst. It warns that a shared session is write access, and that the copy holds the coded Reddit material with real usernames.

A fork, as distinct from a resume, went through v37's fork.sh and restore_stage.sh, which TOOLS-STATUS.md lists as not ported to v38.

Where it lives

The run folder is 5.2 GB in 9,333 files. Its checkpoint tarballs, one per step, are each over 95 MB. It is not in this repository: the repository carries one run, v38's. The v37 instrument's own texts are in the history bundle, under aggregator/v37/.