# One-change assurance: replies and committed effects

This is NEW fictional author-created teaching evidence, not an authenticated customer incident, named-model generated test run, AI4SDLC pilot/compliance claim or native Proof record. Read CONTRACT.md, inputs.json, expected.json and assurance-record.json. Human validation/real decision ownership remain unperformed/unassigned.

## Reproduce

Use CPython3.12+ with stdlib sqlite3 and a writable local directory. No Proof, solver, network, pip, account or credentials. The primary used CPython3.14.7/SQLite3.53.4. From this extracted directory:

```sh
python3 -B reproduce.py --output replay-output
```

Use a NEW output directory. Expected wrapper exit0 means the deliberately faulty baseline is rejected by the strong gate (child exit1), and corrected implementation accepted for this selected contract (child exit0). It does NOT accept the baseline. Both weak gates pass. Source/oracle/inputs remain fixed during execution.

## What actually ran

Two fresh file databases, four local adapter attempts each: first key-A at supplied0h, same payload retry1h, changed payload conflict2h, distinct key-B2h. Each handler commits/returns before the adapter deliberately withholds the first result and raises local ResponseLost. The caller receives NO response for that attempt; the profiler separately observes the actual handler return. There is no real network/HTTP timeout. Adapter profiler return=None on exception unwind is not a normal returned product response.

Baseline returns original reservation2 on retry but commits extra reservation3. The response-only fragment checks received kinds/original retry ID and passes. Complete independent literal typed rows and response checks reject baseline11/16; corrected passes16/16. The final baseline new-key ID4 rather than literal3 follows this fresh seeded allocator/state; it is not a universal promise of sequential IDs. Both variants reject conflict before another write; the baseline's mismatch after conflict is carried-forward retry duplication, not a new conflict-side write. Distinct-key success ensures useful new work remains possible.

Actual units across variants:8logical attempts,8handler+8adapter entries=16selected public entries,8normal handler returns,2withheld results,6received caller replies,10new-connection full-state observations.32strong checks include5baseline failures; weak2checks per variant assess the SAME observations, not extra calls. Complete ordered reservations and ledger rows include unrelated seed state. sqlite_sequence is internal allocator state outside this declared application-state oracle. The exact source/function/argument/return profile and SQL trace are retained; no branch coverage/native Proof grounding is claimed.

Raw keys/exact two-field payload/stable monotonic supplied clock are preconditions, not an executed validator. Only0/1/2hour inputs run; no actual waiting, full24h retention evidence, expiry/at24/after24/cleanup policy or real HTTP/header/stock/replica/concurrency/storage-failure case. New observer connections show selected committed state, not power-loss/restart durability. Fresh corrected DB is not cleanup/migration of baseline duplicates. BEGIN/COMMIT grouping exists without an injected SQL-failure campaign.

The deliberately authored generated-style check is not an actual AI/model output. EVIDENCE-HANDOFF.json links actual results while preserving human validation/approval/model permission and Change/release gaps. The old article's interface-v3/SW-REQ-241/R17 observations were invented; guide05's actual memory-store evidence is separate and not reused.

MANIFEST.json covers payload files excluding itself/SHA256SUMS; SHA256SUMS additionally covers MANIFEST.json, excluding itself. ZIP checksum lives outside archive. PROVENANCE.json retains versions/commands/source identities/units/limits. A separately identified replay can have different timestamps/runtime/paths/SQLite binary bytes; never rewrite retained primary captures to match it.
