# Frozen e34 Windows gate evidence

The portable verifier passed for source
`e34ba27e26ede3ab55f276acc250d92b2e0c3204`, Git tree
`29f769926e2e907f0b40c3b33ec36658ee6601b7`. Its finite semantic-control run
passed all 13 cases: one positive accepted and 12 deliberately mutated packages
rejected. This verifies recorded evidence and exact Git inputs; it does not rerun
Cargo, tests, or native workflows.

The 1,073,739-byte archive is unchanged from the original capture package:
SHA256 `08c9a6e72e74e3368a1b167b358bf76923db3986d53457d5ead707926d6b3b5b`.
The reviewed verifier is 24,051 bytes:
SHA256 `2e2186141957d9490fecd8ada6956aa1419510602735007b287b36235ae477a8`.
The top-level archive, manifest, verifier, and `.gitattributes` are exact copies
of the tested v2 package. The manifest's status field records its packaging-time
state; the completed verification is established by the execution receipts below.

## Recorded gates and source bindings

- Literal workspace inventory: 2,013 names, 60 executables, three doctest sections.
- Literal unfiltered workspace run: 2,005 passed, zero failed, eight ignored.
- Two direct repaired desktop groups: six and one passed.
- Strict workspace/all-target Clippy, three native artifacts, and one external
  transition-oracle build passed their recorded admission checks.
- Every e34 gate binds the same 760 committed inputs. Verification checked 1,520
  Git blobs across the separate e34 and historical f82 subjects.
- The workspace run rebuilt three executable hashes. Three fresh direct listings
  bind their 0 / 12 / 1,347 members; admission uses those three replacements and
  the other 57 original artifact identities. Both the initial missing-DLL failure
  and successful supplementary-listing retry remain in the unchanged archive.
- The external transition build binds its three prepared files and preparation
  receipt, distinguishing the original f82 preparation from its actual e34 build.

Historical f82 `combined-desktop-diagnostic-1` remains a failed run: 1,337 passed,
three failed, seven ignored. The verifier independently pins its original raw log,
failed identities, and ignored names/reasons. The repaired e34 source is a distinct
subject. The preceding published 13-gate checkpoint is at
`2e7c39965ccf91449054d247e82a9882da80b9b3:harness/experiments/peer-union-windows-20260923`.

## Verification and controls

Run from this directory with Python and Git, against a repository containing both
the e34 and f82 commits:

```text
python verify.py --repo /path/to/ARCADIA-Studio
```

The verifier parses raw logs itself, checks exact commands and artifact bindings,
and uses read-only Git commands. It never imports or executes archived launchers.
The standalone positive run exited zero with empty stderr in 9.688 monotonic
seconds. The separate 13-case driver run exited zero in 69.375 monotonic seconds.
These are evidence-verification timings, not application or test-suite timings.

| Semantic mutation | Rejection boundary in the reviewed verifier |
| --- | --- |
| Historical failed/ok identity swap; removed panic details | Line 293: fixed historical log hash |
| Ignored/ok identity swap; changed ignore reason, including manifest updates | Line 193: independent eight-test ignore policy |
| Supplementary leaf receipt reports failure while aggregate reports success | Line 166: aggregate/leaf equality |
| Supplementary before snapshot changed | Line 158: before-snapshot equality |
| Rebuilt artifact rolled back to its original hash | Line 181: expected artifact identities |
| Outcome replaced with an unlisted name | Line 86: exact outcome membership |
| Duplicate outcome replacing another | Line 85: unique outcomes |
| Filter added to workspace command | Line 134: exact command |
| Source-input digest changed consistently across records | Line 321: actual Git blob comparison for e34 `Cargo.lock` |
| Compiler error appended to native build log | Line 225: compiler-error refusal |

Ordinary archive, payload, reference, and verifier hashes were valid in every
control; rejection did not depend on stale wrapper hashes. The source-input
mutation reached Git validation. There were zero unexpected results.

`verification/positive.result.json` and `verification/controls.result.json` retain
exact argv, working directory, exit code, UTC observations, monotonic duration,
and raw-output pins. Their adjacent stdout/stderr files are byte-preserved.
`verification/controls/report.json` and each control's manifest/stdout/stderr
preserve individual results. `verification/boundary-audit.json` adds inspected
traceback boundaries and hashes for each case. Original machine paths in these
receipts are provenance, not portable execution instructions.

`verification/before.json` and `verification/after.json` are byte-identical,
SHA256 `14511bf9c9548259345a29920fc6b58d92934043c9d414aa1739112005da2392`:
all 760 current source inputs, clean tracked state, source HEAD/tree, both original
and v2 packages, and watched helpers remained unchanged during execution.
`verification/copied-file-pins.json` maps every copied file to its original local
path and byte pin. The boundary audit is 21,458 bytes, SHA256
`4fe12bd061756617727b4dccfe2879616caf53851ef4ab420a4db91f9c909b10`.

Duplicated and mutated control ZIPs are intentionally omitted here. Their original
copies remain local-only under
`.local/scunion/.local/e34-evidence-v2-verification-run-20260923-1/controls`.
Each omitted ZIP's identity is retained in its copied case manifest and boundary
audit. Per-case verifier/README/attributes copies are also omitted. This directory
contains the original shared evidence ZIP and reviewed verifier needed to recreate
the finite controls.

To reproduce controls without changing this evidence directory, run the following
Python setup from the repository root, then run the printed command. It creates a
fresh temporary directory and places the exact driver in its expected layout:

```python
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile

repo = Path.cwd()
evidence = repo / 'harness/experiments/peer-union-e34-windows-20260923'
scratch = Path(tempfile.mkdtemp(prefix='arcadia-e34-controls-'))
package = scratch / 'peer-union-e34-evidence-proposal-v2-20260923'
package.mkdir()
for name in ('evidence.zip', 'manifest.json', 'verify.py', 'README.md', '.gitattributes'):
    shutil.copyfile(evidence / name, package / name)
driver = scratch / 'e34-evidence-v2-controls.py'
shutil.copyfile(evidence / 'helpers' / driver.name, driver)
command = [sys.executable, str(driver), '--repo', str(repo), '--out', str(scratch / 'results')]
print(subprocess.list2cmdline(command))
```

The driver retains all generated cases and raw outputs. The exact original v2
producer, external execution recorder, and review diff are in `helpers/` for
provenance; those two original-workspace helpers expect their original sibling
directory layout. The control driver and the portable verifier are the supported
reproduction entrypoints described above. Static review added independent
historical/ignore policy checks, supplementary receipt consistency including the
retained failed attempt's source/baseline bindings, and native error refusal.

## Evidence limits

- Verification checks recorded evidence; it does not rerun gates or reconstruct
  executable bytes. No executable binaries or DLLs are included.
- Eight named tests were ignored, not executed or passed. Their exact names,
  sections, and reasons are in the manifest, positive stdout, and independent
  verifier policy. The unwired Scenario label experiment is outside the inventory.
- Native workflow trials, native usability, public source8, and release acceptance
  are outside this package's claims.
- The full gate launch rejected lifecycle-phase and stack overrides, but its fb
  receipt omits those environment fields. Launch preconditions are transcript
  evidence, not a captured receipt snapshot. Direct focused and supplementary
  listings capture both as null.
- Host UTC clock jumps occurred. The main fb recorder lacks a monotonic duration;
  no original gate elapsed-time claim is derived from UTC subtraction.
- Historical f82 failures and the initial missing-DLL supplementary attempt remain
  failures in the retained evidence.

## Adding this evidence to Git

The archive and some evidence names can match repository ignore rules. The parent
integrator must explicitly include them; do not assume an ordinary add captured
every file. From the repository root:

```text
git add -f -- harness/experiments/peer-union-e34-windows-20260923
```

Top-level `.gitattributes` is byte-preserved from the tested package. Nested
attributes disable text conversion for copied helpers and raw receipts. Inspect
the staged files and run the staged package verifier before committing. No source
change, publication, or release acceptance is implied by this directory.