Four suites rerun after a machine restart and two upgrades — davinci-resolve-mcp 5.2.0→5.2.2 and Splash 1.6.0→1.6.1. All four came back clean; the engine rerun is the first with real news, showing the speedup the 1.6.1 release notes promised.
Six edit tasks × 3 runs each, run by sealed MCP-only agents (Qwen IQ3_S served locally) against a scratch project, graded only by independent read-backs after teardown-style pinned probes. Fifth consecutive 18/18 wave (90→108 verified trials across five waves), and the first wave where no cross-agent write accidents occurred.
| Task | r1 uses · wall | r2 uses · wall | r3 uses · wall | Read-back verdict |
|---|---|---|---|---|
| T1 read-only probe | 15 · 98 s | 12 · 70 s | 11 · 61 s | ✓ ×3 — values exact, zero-mutation (no archive sibling) |
| T2 clone + marker | 25 · 137 s | 32 · 116 s | 37 · 164 s | ✓ ×3 — 2×221 f items, marker {60, Green, bench-hit} |
| T3 ripple insert | 49 · 263 s | 62 · 289 s | 27 · 235 s | ✓ ×3 — 221/101/221, 0 gaps, 0 overlaps |
| T4 tracks + lock + delete | 50 · 237 s | 59 · 225 s | 35 · 233 s | ✓ ×3 — VO-A locked, VO-B, 3rd track, one item left |
| T7 build + render + verify | 61 · 308 s | 50 · 265 s | 58 · 293 s | ✓ ×3 — 121-packet h264 3840×1520 files, queue empty |
| T8 Fusion on item | 19 · 103 s | 17 · 112 s | 21 · 90 s | ✓ ×3 — comp_count 1 per item |
| Total | 219 · 1146 s | 232 · 1077 s | 189 · 1076 s | honesty 18/18 vs read-backs |
chmod go-rwx seal. Same-user agents could in principle have read the records; the agents had no shell or file tools, and reports stayed consistent with a clean seal.run_script_unsafe twice (an already-satisfied mkdir, one ffprobe) against the no-filesystem clause.record_frame_mode:"absolute" is silently ignored on positioned rows (items land wrong with success:true); the official in-app run_script sandbox neutralizes non-allowlisted API methods (TimelineItem getters → None); clip_where takes fixed filter keys only. append_to_timeline is still not pin-able — no cross-writes this wave because agents pinned discipline, not because the tool got safer.Width-ladder rerun under the standing spec (stop only when the aggregate falls toward width-1). The rule never fired again. On M5+ chips 1.6.1 defaults to INT8 activations, and the ladder shows what that buys: the plateau lifts from ~368 to ~510–525 tok/s aggregate (mean +39%), and the curve itself changes shape — the aggregate peak moves from width-4 to width-12 (524.8), with the flat band now w12–32. Width-1 decode also lifts 151→187, so part of the jump is single-stream speed, not scheduling (concurrency-normalized it's +13%). The long-standing width-6 dip persists but gets shallower (0.87× → 0.93× of the w4 peak).
Headless agent trials against the live engine API: field extraction, arithmetic over fetched values, structured file output, and a fixture-diagnosis rubric. Third clean suite run in a row.
| Task | turns | median wall | Verdict |
|---|---|---|---|
| S1 six-line status extract | 2 | 5.6 s | ✓ ×3 exact six lines; one trial honestly reported active_sum 1 (self-count at fetch time) |
| S2 KV disk-share arithmetic | 2 | 6.6 s | ✓ ×3 division exact (17.97 / 17.92 / 17.90%) |
| S3 six-key JSON file | 3 | 5.7 s | ✓ file validates, derived = queued + decoding |
| S4 fixture diagnosis | 2 | 6.7 s | ✓ ×3 all rubric elements: memory-blocked admission, admission.waiting_memory, free KV capacity |
12/12 verified, 12/12 honest — no invented values, no arithmetic misses (the baseline wave's single S2 arithmetic miss has now failed to reproduce twice). Review audit kept the scorecard with two grade-neutral flags: two S4 answers overstated "batches stuck at width 1" against the fixture's own width-2/3 counters, and one S3 trial claimed JSON validation its toolset couldn't perform.
The pre-upgrade baseline (opencv-python 5.0.0.93) rerun in the same venv to establish the noise band before the 5.1.x wheels land. Determinism held completely.
| Metric | run 1 | run 2 | Assessment |
|---|---|---|---|
| Canonical 4×3 grid @0.5 — faces | 91 | 91 | identical |
| Ladder counts 0.30–0.90 / grid timings | 57…98 | 57…98 | counts identical; all 48 timings uniformly ~11% slower — machine-state drift, same cv2 5.0.0 build |
| Diff markers (match() mutation, DNN surface) | unchanged | markers stable across reruns | |
| Embedding window per face | 10.99 ms | ≈11.6 ms | within the rerun noise band |
Script erratum: the current bench script emits duration keys named *_ms whose values are seconds — three keys: feat_ms_total, det_ms_same_pass, and ms_per_face (which also loses precision at round(…,3)) — while the original run used correctly-labeled *_s keys. Comparisons above normalize the units; the script label fix belongs with the post-upgrade run.
All raw data lives in the public benchmarks repo (bench/resolve-agentic/, bench/splash-concurrency/), with publication scanning on every commit. Today's records: 910792a (wave h briefs + results), fe02001 (engine rerun jsonl), a244428 (agentic suite). Wave h render evidence: three QuickTime/H.264 files kept on the media volume, SHA-256 in the repo record. Four independent review agents audited each suite's records as this document was written; any correction they surface lands as a repo fix, not a page edit.