THE BATCH BOUNDARY CHANGED THE AUDIT VERDICT

Recorded synthetic local investigation. No customer data or real payments.
The original local experiment used no model calls or production API calls.
This package records implementation behavior, not model accuracy, audit
assurance or a measured financial loss.

Recorded engine revision:
6205721f91352039de5ce861c1c94d15d9903c7d

INSPECT THE RECORDED EVIDENCE
Download evidence.zip, then run:

  unzip evidence.zip -d batch-scope
  python3 batch-scope/check_results.py --evidence ./batch-scope

The checker uses only Python's standard library. It reads the downloaded
JSON files and prints a report; it writes nothing, performs no network calls,
imports no engine modules and contains no engine source. It checks the
recorded before-and-after outcomes, input consistency, arithmetic, request
hashes and agreement between focused summaries and full responses.

This checks the published evidence; rerunning the engine requires access to
the pinned source revision. Internal consistency does not independently
authenticate the original experiment. Engine source and the original
execution harness are not included in the public package.

RECORDED LOCAL RUN
33 scenarios, 40 focused item results, 297 local checks passed, zero failed
local checks and zero model calls. These are the original local run counts;
the independent public checker performs a different set of evidence checks.

FILES
check_results.py           Standalone recorded-evidence checker.
inputs.json                All 33 exact requests, indexed by scenario name.
results.json               Forty focused item outcomes across those scenarios.
manifest.json              Recorded revision, run provenance and checker hash.
checks.json                The original local run's recorded check outcomes.
synthetic-population.json   The shortest invoice pair and its fictional vendor.
asset-manifest.json         Hashes and sizes of published assets except itself.
evidence.zip               Core research files plus every exact request and
                           full recorded response under inputs/ and results/.

After extraction, inputs/ and results/ each contain 33 JSON files. The
checker also works in the published asset directory when evidence.zip is
present, by reading those full responses directly from the archive.

KEY CASES IN INPUTS.JSON AND THE ARCHIVE
payments-joint: PAY-B fails matched; open balance 400; overall false.
payments-split-omitted-PAY-B: matched holds; open balance 1000; overall true.
payments-split-full-PAY-B: matched fails; open balance 400; overall false.
joint-omitted: both invoice rows fail not_duplicate; overall false.
split-audit-omitted-ERP-ROW-A: not_duplicate holds; overall uncertain.
single-verify-omitted-ERP-ROW-A: not_duplicate insufficient_evidence.
single-verify-empty-ERP-ROW-A: not_duplicate holds.

Invoice overall uncertainty is intentional: goods_specific was not asked
because questions=none. It is not an overall true result. Payment assertions
in these scenarios are deterministic and can return true without a model.

Full history is complete only for the defined synthetic snapshot. These
records cannot establish that an operational export is complete, authenticate
business documents, infer real settlement or resolve general credit allocation.
