Skip to content

Foray 236 · Saltmere · 22 November 2026

What the shifts received and decided

Read the record in order

  1. The original records, outgoing note, notices and five scoring questions. The outgoing agent saw only H1–H5. It produced a prose handover and five claims with eight supporting links before seeing either notice. The study designer already knew the Saltmere scenario. The dated study record was fixed in revision 394c959 before any incoming agent was asked for a decision.
  2. The claims and their supporting records. Every incoming shift received the same abbreviated desk card, the record titles and the two notices. Each treated the notices as independent changes to the same original plan. The plain shift received no support list. The two linked shifts saw the links and their first-pass review flags. In one list only the link from the one-trailer instruction to the timetable was removed. Another instruction still linked to the timetable, so a review flag for that record was visible.
  3. What the three incoming agents requested and wrote, in full. Each worked in a fresh conversation, selected its own source records, and received their original wording before giving advice. Each opened five distinct records in one round. Only the abbreviated card went to incoming shifts. The outgoing author’s fuller note is available in the first file above.
  4. The five predeclared checks applied to each answer. The review quotes each answer and explains partial credit. Inspect the operational decision, the evidence requested and the explicitness of the repair separately. The exercise used one incoming agent per condition. It recorded source requests and link counts. Elapsed time, staff effort, model tokens and actual trailer capacity were outside the measurements.

The trial-record checker verifies the frozen input fingerprint, the gap and overlap, the source requests and every quoted passage. Save all five JSON files on this page alongside the checker and run python3 check-trial.py. The quotations let you inspect the reviewer’s judgements yourself.

Replay the nine filter results

The original Semantica output records five claims across three link lists (intact, missing and repaired) and three states (original, changed timetable and independent rename). Retraction marks an old claim for review. A shift still has to decide what action to take. The experimenter restored the missing timetable link for the repair check.

Download the replay script, support map and recorded output into one folder. This command checks all nine selections independently with Python’s standard library.

python3 saltmere-filter.py --check-record

To repeat the native filter on Python 3.12, download the tested dependency versions to the same folder. The list pins Semantica 0.7.0 to the source revision used for both Gimmer and Saltmere. The filter check runs with the installed package and requires no model service.

python3.12 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python saltmere-filter.py --check

Both checks concern the software’s selections. The three shift decisions above are the observed responses from this run. Return to the Saltmere essai.