# Documented claim: structural linkage and an actual witness

This kit has two distinct evidence paths. The portable command reruns the
selected **Python renderer behavior and witnesses only**. Captured native
Proof results show what the current-source structural checker actually reported
for three separate project snapshots. Neither is customer approval, permission
to release, a complete documentation inventory or project-wide fulfillment of
documented_contract_tested.

## Portable behavior replay

Extract into a new directory. With Python 3.10 or later and its standard library,
run from the extracted root, selecting an output directory that does not exist:

```sh
python3 -B reproduce.py --output /tmp/documented-claim-behavior-rerun
```

No network, Proof, credentials, pytest, unittest or third-party library is needed.
The primary behavior run used CPython 3.14.7, optimization disabled. Runtime
timestamps and absolute command paths naturally differ when reproduced.

The command makes six actual render calls: baseline, corrected and deliberately
seeded always-cancel control, each with cancellation-after-first and normal
completion. Every case starts target bytes KEEP and unrelated bytes OTHER.
The independently fixed expectations are target KEEP on cancellation and AB
on normal completion, plus exact typed replies and unchanged OTHER. The baseline
leaves A on the cancelled path. The correction stages bytes in memory before
ordinary final publication. The no-write control fails required normal work.

| Variant | Strong selected cases | Child exit | Weak cancellation-status check |
|---|---|---|---|
| Baseline | 1 of 2 | 1 | Passes on the same cancellation observation |
| Corrected | 2 of 2 | 0 | Passes on the same cancellation observation |
| Always-cancel control | 1 of 2 | 1 | Passes on the same cancellation observation |

The wrapper exits 0 only when those expected contrasts occur. This does not
accept the baseline or control. Witness functions consume captured observations;
run_behavior.py drives and profiles the actual implementation calls. There are
no additional render calls for the weak comparison.

## Retained native structural run

The exact templates are in native-projects/S0, S1 and S2. Each declares one
eligible draft cancellation cell. The normal-completion sentence/function is
a behavioral positive control, not a second native claim cell. Source mappings
bind each snapshot's renderer and witness to the executed public variants.

| Snapshot | Structural condition | Actual row/severity | Explicit warn-threshold exit |
|---|---|---|---|
| S0 | witness.py absent | WARN / warning | 2 |
| S1 | Existing weak witness and baseline | PASS / warning | 0 |
| S2 | Existing strong witness and corrected source | PASS / warning | 0 |

S1 PASS is expected checker behavior. It establishes the declared presence
floor, not execution or adequacy of the witness, and is not a Proof defect.
Each raw JSONL contains one selected check and the persisted native result
records its severity. S0's work_units.requirements=2 is telemetry, not two claim
cells. S1/S2 content-addressed run IDs happen to match; exact source inputs and
separate invocation receipts distinguish those executions. No behavioral or
unique revision identity is inferred from that matching ID.

The actual native command, run on fresh copies of each template, was:

```sh
PATH_TO_MATCHING_PROOF audit --check documented_claim_verified --stage spec --no-cache --format json --fail-level warn
```

This is a captured-command/prerequisite description, not part of reproduce.py.
The native campaign used one fresh offline current-source build with Go 1.26.3
on darwin/arm64. Exact binary SHA, build argv/environment/buildinfo, selected
source identities and snapshot hashes are retained under captured/build and
captured/native. The 165 MB binary, full current dirty checkout, full internal
4,033-file compilation identity ledger and temporary index databases are not
bundled. Reproducing the exact native build requires the matching source/context,
Go/dependencies and binary identity; installed Proof or a Git revision alone is
not presumed equivalent. A different available verifier may report differently.

The CLI refreshed trace/index metadata and retained run/cache artifacts;
--no-cache bypassed lookup, not every output write. All authored snapshot input
bytes were verified unchanged after each invocation. The selected checker did
not run the Python witnesses. Raw native stdout/stderr, exact command receipts
and persisted result JSON are supplied for inspection. No extra native cases
or native replay is hidden in the portable command.

## Inspect and bound the claim

Read CONTRACT.md, cases.json, the three implementation files and weak/strong
witnesses. Primary raw behavior results are under captured/behavior; native raw
results under captured/native. provenance.json and source mappings distinguish
new teaching intent, selected runtime observations and structural source rules.
native-source/current-source-regions.json embeds the exact inspected checker,
model, selection/configuration/help, threshold and contextual source regions.
It includes the current help's unsupported dedicated enabled:false recipe;
the supported generic disabled list is source-inspected, not exercised here.

The methods note uses official Python 3.14-family documentation currently
labeled 3.14.10; the actual primary runtime remains 3.14.7. BytesIO/getvalue and
Path.write_bytes do not supply this teaching cancellation policy, atomicity,
failed-write rollback, fsync, asynchronous safety or concurrent isolation.

Only two selected synchronous input conditions across three variants and two
named files are covered. No general
file-tree, arbitrary input, other cancellation position, I/O fault, concurrency,
power-loss, restart or durability claim. Native manual/signal kinds, generic
disable policy, nil/no-cells/load-error paths and all suffix/path branches remain
unexecuted. A structurally complete cell does not discover every shipped claim,
resolve a symbol/test function, validate a signal or authenticate manual review.

manifest.json lists content members except itself and SHA256SUMS. SHA256SUMS
also includes the manifest. Where shasum is available, verify supplied bytes
with `shasum -a 256 -c SHA256SUMS`. The separate ZIP checksum identifies the
download. Hash identity is not authenticity, current deployment or approval.
