Quote-batch trial: three memory specialists under the new selection list

The operator commissioned this trial on 2026-09-28 for execution in a new session. Its purpose is to test whether fresh memory specialists, given only the current instruction, adopt the batch mode added to commonplace-quote in commit e984a460 and produce reports with no quotation authoring errors. Only the specialist runs: no coordinator, no exact result, no review, no publication. This document prepares the trial; no worker has been started.

Command for the next session

Paste this instruction into a fresh session at the Commonplace repository root:

Run the quote-batch trial described in
kb/work/agentic-memory-refresh/quote-batch-trial.md. Act as the trial
coordinator: launch one fresh memory specialist per prepared run, audit each
trace for quote-generator use and errors, verify each report, and record the
result in the workshop. Follow the fixed inputs and boundaries in that file.

What changed and what the trial must show

In the 2026-09-27 cleanup regression, all six workers wrapped the quote generator in their own Python loops, one subprocess per quote, each assuming a single-occurrence answer. One wrapper failed on a valid two-occurrence response. The generator now accepts --selections <json-file>: a list of {key, source_path, text} objects resolved in one call, returning a JSON object keyed by selection with a citation, a candidate list, or an error per key, and exit status 2 when any key needs attention. The command reference is the only place the flag's semantics are documented; the memory instruction names the flag in one clause. The trial must show whether a specialist finds and uses it unprompted, and whether the quotation errors of the earlier runs recur.

Prepared inputs

Three run directories exist under kb/reports/state/agentic-system-analysis/, each with a validated running run state that carries the frozen source of the 2026-09-27 run it copies, and a memory-input.md copied from that run with only the run ID replaced. They are local, git-ignored state.

Run	Copied from	Source commit	Baseline specialist report	Baseline configuration
AAS-2026-09-28-dynamic-cheatsheet-quote-trial-01	AAS-2026-09-27-dynamic-cheatsheet-04	5cfe3c37e8e52b1d858d0f3df46e7f17c50991b9	AAS-2026-09-27-dynamic-cheatsheet-04/memory-report.md, 20 quotes	gpt-6-astra, medium
AAS-2026-09-28-mem0-quote-trial-01	AAS-2026-09-27-mem0-04	94c3fe9f238f3dbf29c9ce98643bd71eb13077cd	AAS-2026-09-27-mem0-04/memory-report.md, 23 quotes	gpt-6-astra, medium
AAS-2026-09-28-napkin-quote-trial-01	AAS-2026-09-27-napkin-05	7582d6a46f5a11995956e60a59c41a5b242109f1	AAS-2026-09-27-napkin-05/memory-report.md, 35 quotes	gpt-6-astra, high

At startup record HEAD, the SHA-256 of kb/instructions/analyse-agentic-system/jobs/memory.md, of each run's memory-input.md, and of the installed commonplace-quote source (src/commonplace/cli/quote.py, src/commonplace/lib/quote_generation.py). Confirm commonplace-quote --help shows --selections; if it does not, the editable install is stale: stop and report. Verify each checkout's origin and that the commit object exists. Do not change source worktrees.

Execution and isolation

Launch one fresh specialist per run, one at a time, with the baseline model and effort. Create each with fresh context (for the collaboration tool, fork_turns="none"). Supply only: the instruction path kb/instructions/analyse-agentic-system/jobs/memory.md, the run ID, the input path, the report destination <run-dir>/memory-report.md, the permitted source access root from the run state, and the statement that the parent has commissioned a fresh source-only memory analysis. Explicitly supply repository doctrine if the runtime does not load it. Do not mention the batch flag, the earlier runs, their reports, this trial, or any quotation failure. The instruction must be the only route to the new flag.

Workers own their report only. The coordinator owns scheduling, the audit and this record. Use completion events and the owned output path, not agent-status listings. A worker that reports prior-analysis exposure stops; record it and launch a replacement with a new run directory copied the same way. Do not edit the instruction, the generator, or the run states during the trial.

Evidence and acceptance

After each worker ends, retain its trace under kb/reports/cache/agentic-memory-refresh/quote-batch-<date>/ with path, hash and model identity. Audit the full trace, including nested tool results, exit statuses and stderr, and record:

• Adoption. Every commonplace-quote invocation: whether it used --selections, how many selections and distinct source files each call carried, and how many single-selection calls remained. Whether the worker read the command reference or --help before its first call. Whether any hand-written loop or wrapper around the generator remains, and what assumption it makes about the output shape.
• Outcomes. Each exit status. For exit 2, which keys were candidates or errors and how the worker resolved them: chose a candidate's citation unchanged, lengthened the selection, or dropped the quote. Any exit 1 and its cause.
• Authoring errors. Any citation edited after generation, any quote block written without the generator, any YAML quoting error such as a bare yes key, and any validation failure on the first commonplace-validate --full of the report, with the diagnostic.
• Verification. Run commonplace-validate --full <report> and uv run python scripts/verify_report_quotes.py <run-state> <report>, which resolves every citation against the frozen commit as completion verification would. Record the counts and every failure.
• Comparison. Quote count and word count against the baseline report; the fourteen memory-comparison assessments and values beside the baseline's. Differences are context, not the acceptance question.

Write quote-batch-trial-<date>.md in this workshop and link it from its README. Answer separately for each specialist: did it use --selections, did every generated citation enter the report unchanged, did any quotation or YAML error reach validation, and did every citation resolve against the frozen source. Three stochastic runs do not establish a general rate; claim zero issues only within the inspected evidence.

This commission covers three specialist runs, their audit and the workshop record. It does not complete or publish any run, produce an exact result, alter the instruction or the generator, or authorize Git commits. The three run directories may be deleted after the record is written. Finish with the evidence-backed answers or concrete blockers.