TIM single provisional proposal and reviewed dispatch, 2026-11-15

The experimental path did not qualify. No new original was started. Three fresh native operator fixtures could not be personally reviewed as complete current scenes. Each proposal was discarded before intent creation, with zero commands, native input calls and callbacks. There is no new latency admission, visible input consequence, Puzzle 2 construction/Run/clear or native Quit claim.

Evidence is in scratch/runs/20261115-tim-single-reviewed-dispatch-puzzle2/. The preceding published helper/report remain edbdea8c8 / 430a52c23: Puzzle 1 constructed/Run/clear/618/earned Advance; Puzzle 2 motor placed and flipped, no belt/Run/clear; native Quit YES exit zero at 606.221 seconds. That original's owner-changed UNKNOWN is permanent. Historical password image 2791 remains ambiguous and unused; nothing was guessed or typed.

The previous whole-scene operator admission was 2.994704 seconds. Its actual image-return to post-review request was 2,851 ms; the original's median was 3,193 ms. Python entry, schema, hashing and writes are small costs and cannot establish a reduction in image/tool/model latency. Before execution this task required capture-to-admission at most 2.5 seconds and image-return to post-review request at most 2,400 ms, plus personal whole-scene review, visible native effect and independent UP. No current sample reached confirmation, so both admission measurements and improvement remain unestablished.

The NEW focused helper tests one manually chosen provisional gesture before capture. A proposal carries no intent ID or nonce. Capture binds one entirely new frame and cannot confirm or dispatch. Only a separate explicit personal whole-scene confirmation delegates to the existing reviewed helper, creating a fresh independent intent after review. No future gestures or automatic interpretation, target discovery, OCR, solver or player code are present.

Immutable proposal/capture/consumption records bind source hashes, native owner, frame, PNG hash and native capture timestamp. Mismatch, replay, held or pending written input and UNKNOWN refuse with no later requests. Positive native-attested UNSENT alone permits a genuinely new proposal, capture, review and intent. Five-second game/thirty-second menu freshness, full one-second dispatch reserve, finite gesture limits, independent UP, native deadlines and original cutoff remain delegated unchanged. No existing controller, source, module or build implementation was edited.

--orchestration prints versioned conversation-local P, C, Y expressions. P(scene,target,action,consequence) records ONE provisional candidate. C() requests and returns the entire new scene. After personal review, the separate short call await eval(load("Y"))("confirm") explicitly confirms that candidate. The shortened post-review call is experimental, not a measured speedup. If the scene differs or cannot be reviewed, --discard RUN PROPOSAL_ID REASON closes the workflow without intent or input. Capture never invokes confirmation.

After an actual post-UP receipt, the helper can return the already-rendered whole consequence image without requesting another capture. It requires an unheld changed frame and PNG integrity, and suppresses the handoff after UNKNOWN. This route has pure coverage but no successful current native operator sample. A subsequent candidate would still require its own new capture and review.

The capture/image boundary was investigated directly. view_image returned PNG bytes under a data:application/octet-stream URL. Explicit image/png ImageContent, two byte-identical full PNG representations, and original-detail metadata were tested without cropping, changing pixels, making another capture or moving the capture clock. Offline the same bytes sometimes showed the full platforms, objects, ramp, pipe and tiles, and sometimes appeared flat. The three fresh native captures appeared flat to this operator, including both duplicate representations in the second and third trials. None was confirmed. A different operator's offline full view cannot substitute for this operator's fresh review.

All three source PNGs have the exact same SHA-256 7e0f193b54659d5991e150e1ea21d57e0cf68099d7aa42989beb96db84986c22, 1,369 bytes, 320×200 dimensions and 15 decoded RGBA colors. Their native metadata hashes match their stored bytes. This disproves neither presentation failure nor a particular cause; it does not establish flat native rendering. The final capture returned to Python in about 22 ms and reached image-tool return 84 ms after native capture. Fast file delivery alone did not make personal review possible. No cognitive latency is invented.

Qualification ran only after personally verifying both explicit predecessor releases: Q3 ccb19567… (67 actual waits/both EOF, 147 absent PIDs/groups) and Sims ff2c64fa… (468 actual waits/both EOF, 838 absent PIDs/groups, including its sealed nested direct closure). Both shared locks were free. No workers or boxes were created; all foreign WIP was preserved.

The full locked build passed initially and after each presentation variant. The initial suite passed 75 pure checks: 26 policy, 12 exact schema, 10 margin, three prior reviewed-action and 24 new focused checks. Native context, held keys, exact schema, both dispatch-margin cases, 13 controller cases, two scene/owner opposition cases and seven new candidate opposition cases passed. The final code passed another full locked build, 24 focused checks and seven fresh native candidate cases. The new cases cover frame, owner, source and candidate changes, replay, held and UNKNOWN; each proves zero native input and zero later refresh requests. These adversarial checks are not automated successful personal reviews.

One intermediate repeat exposed an existing supervisor cleanup race: native exit raced killpg(SIGTERM), raising ProcessLookupError before the supervisor reaped its watchdog. Native and supervisor-child actual waits/both EOF exist; the watchdog PID/group is absent, but its actual wait receipt is missing and remains permanently UNKNOWN. No input or intent existed. That diagnostic was closed and never replayed. Only the NEW candidate test's cleanup changed to terminal diagnostic UNKNOWN, allowing normal native/watchdog reaping; the runtime supervisor and all guards were preserved. Later fresh candidate cases and all three operator fixtures have direct native, watchdog, supervisor-child and launcher actual waits with both EOFs. Diagnostic closure is not native Quit.

At native closure, 209 recorded actual waits have both EOFs and all 209 recorded PIDs/groups are absent. This count does not reconstruct the missing historical watchdog wait. final-native-closure.json explicitly records incomplete expected historical waits and failed-watchdog-wait-UNKNOWN.json preserves that failure. The final release seals direct closure and verifies current PID/group absence; no native work continues after release.

Only one selected result-named native diagnostic PNG is retained, with its hash and original frame provenance. It is an authored transport fixture, never a TIM gameplay screenshot. Other captures are pruned after hash verification. Git SHA, relevant binary diffs (including owned new files), media/source pins, compact logs, waits and publication records provide recovery without source-tree, media or WASM archives. Box ID/expiry are null, state NOT_CREATED, stop receipt NOT_APPLICABLE.

The concrete unresolved condition is reliable fresh complete-scene presentation to this operator. The published helper remains experimental and ineligible to authorize an original until that handoff and actual measured native consequence qualify under the unchanged guards. This task consumed no conditional original.