# New fictional one-change assurance contract

This new teaching-author contract is not approved customer/product intent, an AI4SDLC pilot/compliance claim or a native Proof record/import. The old page's interface-v3/SW-REQ-241/R17 observations were invented. None is authenticated or reused here; guide05's memory-only retention evidence and SVP's importer evidence are distinct.

One change: make a local reservation handler safe to repeat after its completed result is withheld. Interfaces are baseline_reserve/corrected_reserve(path,key,payload,now_hours) and call_adapter(handler,path,key,payload,now_hours,withhold=False). Exactly one sequential caller and one SQLite writer operate at a time. No HTTP endpoint/header, actual network timeout, replica, stock/inventory service or concurrency scheduler is implemented.

Raw opaque keys are exact nonempty ASCII strings, compared without normalization. Payload has exactly sku:string and quantity:positive integer (exact integer, not bool); values and field semantics remain stable on a replay. json.dumps sorting/separators is this implementation's two-field serialization, not a general canonicalization/parser/security claim. Caller inputs meet these preconditions; no type/key-length/validation boundary is tested.

The NEW authored time domain is a stable monotonic supplied integer now_hours clock, not elapsed real hours. Same-key replay is admitted with 0<=elapsed<24. The selected schedule is first0h,retry1h,conflict2h,new-key2h. Only those values run. There is no expiration/cleanup code, test at24/after24, automatic time validation, clock rollback behavior or claim that selected calls prove retention throughout the full24h window. Exact/beyond24 and retention/cleanup/over-retention decisions remain open; the old invented24h draft is not retroactively approved.

Starting state is a freshly initialized file SQLite DB containing unrelated reservation1(seed-key,seed,1) and ledger(seed-key,{"quantity":1,"sku":"seed"},1,0). Keep every declared column of both tables intact. First key key-A with payload {sku:part-A,quantity:1} creates reservation2 and ledger at0. Handler returns exactly {kind:reserved,reservation_id:2}. The test adapter then deliberately WITHHOLDS/discards that completed result and raises local ResponseLost; the caller receives NO product response. A profiler captures the handler's normal return independently, not a second caller-received response.

Retry same key/payload at1h must return reserved/reservation_id2 and create no additional reservation or ledger write. Same key with quantity2 at2h returns exactly {kind:conflict,code:KEY_PAYLOAD_MISMATCH,reservation_id:null} with no additional writes. New distinct key key-B with the original payload at2h must create reservation3 and a second target ledger entry; this positive operation must not be rejected/collapsed. There are no numeric HTTP statuses. Return dictionary keys/types and full ordered reservation/ledger rows are fixed literally in expected.json BEFORE execution.

Baseline is seeded defective teaching code: on same-payload replay it commits another real reservation but returns original ledger ID2. It rejects conflict before writing. Correction checks prior key/payload before creation, so replay returns original ID without a new write. New reservation+ledger are committed together; no injected mid-transaction storage error/concurrency/atomicity campaign is run. Both variants start independent fresh DBs. This is a local implemented correction, not existing-data cleanup or a deployed upgrade.

Explicit SQL BEGIN IMMEDIATE/COMMIT is used with sqlite3 isolation_level=None and LEGACY_TRANSACTION_CONTROL, journal_mode DELETE,synchronous FULL. Actual live settings/runtime are captured. New observer connections after every attempt read full declared table rows; committed effects at these local observations are supported, not power-loss/fsync/storage-failure/restart guarantees. sqlite_sequence is an internal allocator table outside the declared application-state oracle; no all-database-state or all-side-effect claim.

One deliberately authored GENERATED-STYLE response-only fragment checks received kinds and retry original ID. It assesses the SAME observations, not extra calls or an actual named AI/model-generated test. It can pass baseline while full state fails. New-key success and changed-payload/no-write checks provide useful positive/negative discrimination; no extra control variant is required. Whole state before/after conflict is also compared to distinguish preserved preexisting bad state from a new conflict-side write.

Expected units are four logical adapter attempts/handler calls per variant, but actual counts are derived from real profiles. Handler normal returns, caller normal returns, exception-unwind profiler return=None, withheld results and received replies are different units. The adapter's return=None on exception unwind is NOT a normal returned caller response. Human intent validation/actual approval, model/tool permission, broader testing, Change/release acceptance and accountable real owners remain UNPERFORMED/unassigned. The assurance record keeps these gaps visible rather than inventing signatures or automatic AI4SDLC approval.
