Windows native observations from closed Runs 7/8 on 2026-09-22 exercise existing-record service editing on source b10b28cefd1d0794073ba241a0c8de9a83e85f68. Root operated the frozen GUI; this archive was prepared by read-only verification of closed records. All 20 typed comparisons pass exact complete admitted saved-Project equality to expectations published before these native edits. Native interaction observations and saved-source comparison results are distinct evidence.

Observed workflow and acceptance

The operator opened service-details.arcadia, verified Valve service Details trial, navigated Behavior & scenarios and explicitly selected Saved Service calls - Details trial. Existing-record Properties was reached through selected-message Details; no creation form or direct source rewrite performed these edits. Run 7 operator notes record selections, draft/Apply/Save, model-focus Undo/Redo and exact checkpoint mapping.

Accepted edit	Independent expected checkpoint	Saved edit / Undo / Redo
First Call Scope expression: quoted `primary` to quoted `secondary`	`1`	02 / 03 / 04
First Reply: remove Status/Return annotation	`3a`	08 / 09 / 10
First Call: choose SetLimit allocation, replace incompatible Scope using Pressure limit/In picker, expression `18.75`; one Apply	`3b`	11 / 12 / 13
First Reply: add compatible Accepted/Return annotation, expression `true`	`3c`	14 / 15 / 16
First Call Pressure expression: `18.75` to `19.25`	`3d`	17 / 18 / 19

Comparison 01 establishes the independent baseline. Each accepted edit was saved, followed by one model Undo/Save and one Redo/Save. Exact comparisons cover the whole Project, including control pair, semantic Steps, IDs, names, remaining records and saved layout. They do not merely compare the changed expression. The first Reply inspector derived its operation from @first_call; the operator observed no independent Reply operation selector.

Between edits 1 and 3a, the operator deliberately chose SetLimit while retaining incompatible Scope/quoted secondary and the existing Reply Status annotation. Apply rejected with message_value_field: The qualified field must be carried by the exchange, following Value record fields only. Scrolling back exposed retained raw expression and incompatible address. Explicit Revert then Save produced checkpoint 05 equal to accepted 1; one model Undo/Save at 06 returned baseline and Redo/Save at 07 returned 1. Together the native observations and exact saved comparisons support rejection recovery with no intervening accepted model-history step. They do not inspect the internal history stack directly. The invalid draft was not applied by Save.

Fresh-process durability and preceding context

Run 7 PID 4100 started at 21:44:44 UTC. Root issued Alt+F4 at 22:03:39; the bound observer recorded exit 0 at 22:03:53. Run 8 PID 2016 started at 22:04:47, reopened the exact recent service file and explicitly chose the saved scenario. First Call SetLimit/Pressure 19.25 and first Reply Accepted/Return true remained visible; four messages and two lifelines rendered. Save all and the twentieth typed comparison preserve final 3d. Root issued Alt+F4 at 22:07:33; the observer recorded exit 0 at 22:07:40. Run 8 notes and both normal-close.json records distinguish action times from observed process-exit times. Both runs have empty native stdout/stderr and unchanged bound immutable inputs.

Run 7 final and Run 8 reopened source are exactly 6,278 bytes, SHA-256 92723225deb64884d7cd01e3bf50f24620fbdd583e8bab7bd3f391fb22a46f5d. The incident working file remains the preceding spacing result, SHA-256 b746965f034a4befb449d17e568dd5ae0b2ee81b3c34f82883e48e0c54833b25. No last-view or camera-persistence claim follows: Run 8 opened Automatic before explicit saved-view selection.

All six small Run 6 files are retained as context. Its failed-opening record and resume observation establish no successful service edit. The observer closed with exit 0 on September 14, but root did not observe its close cause; normal-close.json alone does not prove Alt+F4. Run 7's launch binds that exact prior close. Run 6 has no images and is excluded from service-edit acceptance.

Integrity and reproducibility

archive-manifest.json binds 335 original files, 11,342,114 bytes, including every raw Run 6/7/8 file, 24 JPEGs and metadata (21 + 3), all comparison logs and artifacts, checkpoint sources, recipe and baseline. All 24 JPEG hashes are distinct and match JPEG signatures/MIME and capture metadata. This archiver did not independently repeat visual inspection; native image descriptions are the operator's observations. Returned window-title metadata can lag the actual image, as the original notes disclose.

Every original was copied as raw bytes and compared by length and SHA-256; JSON containing 64-bit fingerprints was never parsed and rewritten. All 200 receipt-listed output artifacts, before/after bound-input equality, request checkpoint/hash, exact actual/expected canonical bytes and saved-checkpoint bindings were verified. The initial baseline comparison historically bound the mutable working path: its preserved input-1.arcadia and immutable baseline establish the original bytes, rather than incorrectly comparing that old binding to today's final working file. Later comparisons use immutable checkpoint files. The manifest also records checked external immutable comparator inputs and prerequisite Git blobs. EXE/DLL payloads, private profile, current mutable working files and duplicated prior build/source archives are excluded. Intentional raw .log files need forced Git addition.

Peer machines must retrieve published Git history; original absolute paths are provenance, not shared filesystem locations. Use these pinned prerequisites:

Current b10 comparator and GUI build evidence: successful builds, source/dependency bindings, comparator self-checks, GUI recipe and transitive gate prerequisites. Its copied preparation README's deferred-build wording is historical; the enclosing archive records the completed builds.
Independent core expected checkpoints: six expected Project files, separately byte-checked against the local catalog and every typed comparison. Native actual output never supplied these expectations.
Preceding spacing archive: incident result and prior close chain, with transitive staging/validation links.

The bound service comparator SHA-256 is 996803f938b7503db8508adf9b7fc2dd379852b5b603463edf6987dd1e5270a6; build receipt is 5c700a6b63ce1060c0091d6370d6e07b824fd871adb6bf89fd8c4a155b1c1c18. GUI is 7ece85c6012ae0ae75fc56ed3c3aadb8bc116448ad1782be422ef34a629dc3ce; runtime is aa282fa47bb2c4e210cea5e470cdbdca58568f0be0bc73e404d86ab9cd95ef6a.

For read-only audit, hash the files in this manifest, compare each request's expected SHA against the published six-file catalog, and compare its actual.project.json and expected.project.json bytes. To repeat execution on another Windows machine, check out the exact product and pinned evaluator sources, rebuild following the referenced admitted toolchain/gate recipe with machine-local paths, and record new binary/build hashes. Use details/check.ps1 -Mode typed -ExpectedProject <published expected JSON> -Checkpoint <catalog key> -Actual <archived immutable checkpoint.arcadia> with that newly reviewed comparator/build receipt and a fresh output directory. Native repetition additionally requires a newly staged private package/profile and the archived plan, stage, launch and verification recipe; do not reuse prior run directories or assume this machine's absolute paths or compiled hashes transfer. No comparator, compiler or GUI was executed during archive preparation.

Independent review found no blocking discrepancy in this bounded result and independently checked all 20 declared comparisons plus selected images. It notes that the rejected draft combines incompatible Call and Reply fields, so it does not isolate Reply-only invalidity. One inspected draft frame retains an All source files saved footer while the title/header are dirty; that footer is stale last-action feedback, not proof the new draft was saved.

Friction and remaining scope

The operator reported substantial inspector scrolling between values and Apply/Revert; wheel input over the multiline Description editor did not scroll the outer inspector, while the margin worked. The incompatible-field error appeared in a small global footer, distant from the offending input, with no observed inline jump. Long operation labels and scenario captions truncated. Recoverable raw drafts and passing data-integrity checks do not remove this interaction friction.

Expression strings were edited; numeric/type/unit evaluation was not established. Arbitrary operation migration, deletion, additional lifecycle cases, definition navigation, remaining incident navigation, large models, Linux, broad accessibility and continuous held-motion remain outside these runs. No full issue, native-plan or release acceptance follows from this archive.