# Privacy decision to observed live-store outcome

This is a synthetic engineering teaching kit, not a GDPR assessment. The
unchanged `policy.json` assumes a fictional decision was approved only inside
the scenario. Actual identity, eligibility, exceptions, retention, legal
authority and destination inventory are decisions outside this execution.

## Run without Proof or a network

Extract the complete evidence ZIP into a new directory. Use CPython 3.12 or
later with its standard-library `sqlite3` module, from the extracted kit root:

```sh
python3 -B reproduce.py --output /tmp/privacy-outcome-new-run
```

Choose an output directory that does not already exist. The wrapper refuses
to overwrite observations. It needs no Proof install, account, signup,
credentials, network access, Git checkout or third-party Python package.
Do not use optimized Python (`-O`), because this teaching harness retains
assertions. The captured primary run used normal CPython 3.14.7, SQLite
3.53.4, Darwin/arm64. A different runtime is a new execution identity; its
versions and results must be retained separately.

`summary.json` and raw stdout/stderr are written to the chosen output. The
wrapper succeeds only when these four child outcomes occur:

| Child | Expected exit | Meaning |
| --- | --- | --- |
| Original runner replay | 0 | Its nine checks detect the buggy contrast and accept its corrected sequence. |
| Baseline literal gate | 1 | The real buggy handler falsely reports complete after the analytics deletion fails. |
| Corrected literal gate | 0 | Five real Request-method observations match the literal snapshots and full rows. |
| Swapped-report control | 1 | A deliberately wrong report adapter is rejected despite the unresolved aggregate being correct. |

The wrapper's exit zero does **not** accept the buggy handler or swapped
report. Inspect each child's expected and actual exit code and raw result.
`baseline.stdout.json` and `swapped-report.stdout.json` contain intentional
failures; they are not setup failures. Every retained stderr is empty.

## Inputs, independent expectations and exact calls

`deletion.py` and `policy.json` are unchanged copies from the original public
sample. `historical/` preserves all nine original public files, including its
original ZIP and the October 3 observation. `expected.json` is a separately
authored literal oracle derived from that unchanged fictional policy and the
declared synthetic inputs. It checks complete ordered subject/value rows,
exact result keys, recursive types and each destination's own result. It does
not derive expected rows from `Request.state` or `Store.contains()`.

The new harness calls the actual `buggy_delete`, `Request.attempt` and
`Request.reconcile` functions. Captures retain their source identity, source
line, function-source hash, returned result and per-store SQL trace. SQL
observations query each named live connection directly. The only two live
destinations are `profiles` and `analytics`, initially containing these rows:

```text
synthetic-subject-001 | synthetic target
synthetic-subject-002 | synthetic control
```

The failure flag raises `DestinationUnavailable` before the analytics DELETE.
This is not a measured network outage, post-write SQL failure or commit
failure. The connection remains readable by the independent observer.

The corrected five observations are partial failure, recovery retry, repeated
completion, explicit reconciliation after actual analytics backup restoration,
and deletion of the restored live target. On the repeat, the live rows and
complete/absent result semantics stay stable; observable attempt metadata
intentionally changes from 2 to 3. Whole snapshots are not identical.

The seeded report control swaps only the two destination status values in a
copy of the result after a real corrected partial-failure call. Its aggregate
still says unresolved and its actual stored rows remain correct. The exact
named snapshot comparison fails. This designed report-mapping error is not a
newly discovered product defect or an additional outage.

## Distinct history and counting units

The archived observation is dated October 3, 2026 and reports nine passing
checks across six scenario observations. Those nine checks include detecting
the buggy false completion. Its original stored-effect observer checks row
counts on live connections, not complete payload values or reopened files.

The fresh October 28 primary campaign separately contains:

- One unchanged original-runner replay: nine checks, six observations.
- New literal primary evidence: one baseline and five corrected observations.
- One seeded report-adapter control observation.
- Four child gates with expected exits 0, 1, 0, 1.
- Seven direct top-level handler calls recorded by the **new harness only**.

There are 13 observation instances across these separate child runs. That is
not 13 distinct scenarios, requirements, classes or compliance controls. Do
not add the old archived run into this fresh-run denominator. `PROVENANCE.json`
and `observed/primary/summary.json` preserve the identities and units.

## Restore and observation limits

The harness takes a real SQLite backup before deletion and restores it into
analytics. Immediately after restoration, cached Request state is still
complete although the row has returned. Only the explicit `reconcile()` call
reopens the request. This demonstrates caller-triggered reconciliation, not
automatic restore detection or a durable recovery service.

The backup still contains both original rows, including the target, before
temporary cleanup. Backup erasure is not performed. Temporary-directory
cleanup is not evidence of physical sanitization or erasure of every copy.

All observations use the same live local connections. The captures record the
runtime's actual legacy transaction defaults (`autocommit=-1`,
`isolation_level=""` in the primary run), not newly pinned settings. No
reconnection, crash/power-loss durability, concurrent writer, restart, durable
request queue, recipient notification, deadline handling, complete downstream
inventory, access-control adapter or physical erasure was tested. Subject
identifiers here are synthetic; hashing a real identifier would not by itself
establish anonymisation or authority to retain it.

## Integrity and new observations

`manifest.json` hashes every substantive member before the two top-level
integrity indexes. `SHA256SUMS` also covers `manifest.json`; the outer ZIP's
separate `.sha256` file covers the complete archive. None of these files can
meaningfully include its own hash. The outer public evidence manifest lists
every actual downloadable byte, including these indexes and the ZIP.

Retain new run output beside, not over, `observed/`. Any policy, destination,
implementation, oracle, harness, schema, runtime or restore-process change
requires an accountable freshness decision and appropriate rerun. A passing
finite local teaching campaign does not establish production correctness,
all-copy deletion, accepted legal advice or compliance.
