MOS small project demo
Version: 2026-10-11.1
Purpose: test whether a fresh AI session follows current project decisions and handles deliberate changes correctly.
All project data is fictional. This package contains no private MOS records, account identifiers or personal facts.

QUICK TRY
Open a fresh session. Provide 04_MOS_BOOTSTRAP.txt and 01_FACTS.txt, then ask C1 from 05_TEST_PROMPTS.txt. Compare the answer with the tester-only key afterward. The model must not receive the key or this result discussion.

BOUNDED COMPARISON
Run the same four cases under two conditions: baseline uses 03_BASELINE_INSTRUCTIONS.txt; MOS uses 04_MOS_BOOTSTRAP.txt. Both get byte-identical project facts for each case. C4 uses only 02_MISSING_SOURCE_FACTS.txt. Keep model/version, tools, account memory and custom instruction settings matched; record any settings you cannot control. A new chat may retain account memory, so label isolation honestly.
Use eight independent sessions for one exploratory pass (four cases times two conditions), or run only C1 as a clearly labeled two-session quick check. Randomize which condition goes first. Stop after that declared pass; do not rerun until the desired answer appears. Do not use private live project data.

WHAT TO RECORD
Record model/version, date, settings, packet version, start and finish, complete answer, first-answer score, human correction count, human correction minutes, initial setup minutes, and ongoing maintenance minutes. Capture tool calls and tokens only if exposed. Unknown costs remain UNKNOWN. Separate waiting time from human time and one-time setup from recurring upkeep. Never set unmeasured upkeep to zero.
Use 08_RESULT_TEMPLATE.json. Compare accuracy first, then correction effort and observed time. A smaller prompt alone is not evidence of better results. One small pass is exploratory, not a reliability benchmark. Sustained upkeep needs observation during ordinary work; this fixture alone cannot establish it.

OPTIONAL WRITE CHECK
The attached receipt records an ordered stale-snapshot check already executed on a disposable Drive Markdown file. Two candidate writers began with the same snapshot; one update was saved and verified, then the second candidate's prewrite check found changed content and withheld its write. A single orchestrator simulated both roles. No simultaneous independent AI writers were tested. The disposable file was deleted after the receipt was saved.
Fresh reads are not atomic locks. A read/write race still exists on hosts without conditional writes. Do not run conflict experiments on real canonical records.

FILES
01 and 02: alternative fact packets, one per case.
03 and 04: baseline and MOS instruction conditions.
05: prompts. 06: tester-only scoring key.
07: plain-language public introduction.
08: blank result template. 09: actual bounded write-check receipt.
Fresh-session answers and maintenance measurements are NOT_RUN / UNMEASURED in this release.
