# Compact source native workflow evidence

This package preserves two completed Windows native runs of the default-omission
followup, including the original screenshots, window/accessibility metadata,
process receipts, producers, input fixture, and both saved source bundles.
Run the read-only archive check from any checkout with Python 3:

```text
python harness/experiments/default-native-20260913/verify.py
```

The verifier reads ZIP members in memory. It does not extract files, run the
archived programs, build code, or modify the checkout. `files.json` pins the other
three outer files; `capture-files.json` pins every original payload inside the ZIP.
These hashes establish internal consistency, not cryptographic attestation.

## Source and executable provenance

Both native runs record the same 115,056,640-byte `arcadia-studio.exe`, SHA-256
`4070927256414a770c5fbcedfaab922810cea33d9f07610b81e9ab8c0465da44`.
Their build receipt and the original build stdout/stderr, producer, and complete
613-file executed-input snapshot are included under `build/`. These are copied
unchanged from the
[published consumer package](../default-consumers-20260913/README.md) at commit
`94383107153cccb32f3fdcbd8111b6381f3e15aa`.
The actual build recorded HEAD `df2e1cd65ca5e56f47669e04a9d4bcb426900fcd`
plus the HTTP test subsequently published in `9438310`; the captured source bytes,
rather than HEAD alone, identify the tested input. Production code is unchanged
from `7bb366f66a8c6f8b6f66fc9b04c9e3b39f84b78e`.
No executable is distributed in this evidence package.

Both native process receipts close with exit code zero and empty stdout/stderr.
Trial 2 additionally pins its producer at launch and its immutable baseline ZIP.
Trial 1's runner and the screenshot helper are preserved as original local
producers but were not contemporaneously hashed by that trial's receipt.
Do not mistake a packaging-time hash for stronger historical provenance.
The helper initially selects `trial-1`; the interactive session changed its
output directory to `trial-2` for the continuation. Original producer bytes are
preserved, including their historical machine-specific paths.

## Workflow and evidence boundaries

Trial 1 opened the minimal source fixture, created one activity through Model
Explorer, inspected its compact source, and saved the resulting two-file bundle.
The important frames are `13-created-activity`, `17-compact-new-activity`, and
`27-cleanup-save-submitted`. Intermediate dialog focus and retry captures remain
in the archive; the trial was closed before the cleanup interruption.

Trial 2 reopened that baseline and dragged Work. Its Text view showed x=190,
model Undo showed x=120, and model Redo showed x=190 again. It then used Save As,
cleared the current model, reopened the saved result, and inspected both source
files. Review these frames with their adjacent JSON metadata:

- `03-baseline-loaded`, `04-drag-work`, `05-text-after-move`.
- `06-model-actions`, `07-undo-model`, `08-redo-menu`, `09-redo-model`.
- `17-save-complete`, `19-clear-current-model`, `22-reopened-root`,
  `25-reopened-compact-activity`.

Screenshot interpretation and the intermediate undo/redo observations require
human review. The archive verifier checks their inventory, metadata, timing and
byte hashes; it does not infer visible behavior from frame names or run OCR.
The machine-checkable before/after oracle independently requires exact bundle
membership and bytes: one compact new activity with stable UUID
`1e1da408-da79-42ab-a990-9d00b7dc2631`, one root include, and the sole subsequent
change `position { x 120.0; y 80.0; };` to
`position { x 190.0; y 80.0; };`. The explicit
`port_anchors   [] ; // Keep explicit automatic placement.` survives exactly.

`native/bundle-checks/` preserves both closed CLI checks and their original
`native/check-saved-bundles.py` producer. Both full bundles parsed and validated,
with four expected draft warnings (each activity is unallocated and lacks behavior
context), no errors, no IDs requiring stabilization, and no source changes. These
checks do not compare the entire parsed model. After explicit extraction to a
chosen directory, a rebuilt CLI can check `native/saved-before/project.arcadia`
and `native/saved-after/project.arcadia`; the verification script itself never
extracts or executes them.

On reopen the diagram fits/recenters, so screen coordinates cannot prove the
persisted x value; the reopened Text frame supplies that observation. Save dialog
frames 14–16 retain a filename-selection automation quirk: text first appended,
then was visibly corrected before submission. No incorrect path was submitted.

This is evidence for one ordinary source/graphical workflow on Windows. It is
not an exhaustive usability assessment, Linux result, performance experiment,
packaged application verification, or full release acceptance. See the linked
consumer and core evidence for broader automated coverage.