MCP 5.1.0 — perf flat, discipline holds

Wave 2026-11-10f · Resolve Studio 21.2.0.6 (unchanged) · community MCP 5.1.0 · class-A media (HEVC Main10 ≥8128×4288 @29.97 → 3840×1520 timeline) · agents: Qwen IQ3_S via Splash 1.6.0, fresh sealed subagents, MCP tools only

Warm render (median of 3)

1198 ms · 101.0 fps

1198 / 1194 / 1201 — flat vs 4.19.3’s 1191, inside spread

Outcomes verified

18 / 18

grader read-backs only · third consecutive clean wave (72/72 aggregate)

Honest

18 / 18

every claim matched read-backs; several agents volunteered dead ends unprompted

Server vs channel

5.1.0 ahead of stable 5.0.7

the server itself reports status current_ahead

Render probe — single-tenant, before the wave (host 95.75 % idle)

Round	TimeTakenToRenderInMs	Output verified (my own ffprobe)
cold (fresh import)	3116	121 packets · mov · h264 + pcm_s24le — host page-cache state, not an MCP signal (cf. the 3090/3436 cold rows)
warm 1	1198	byte-for-byte payload identical family: 16,410,691–703 B across all four renders
warm 2	1194	121 packets, h264 High 3840×1520, pcm_s24le
warm 3	1201	121 packets; every round a fresh job (never a re-start of a completed one)

Op medians — 5.1.0 vs 4.19.3 (same base geometry, server-side durations)

Operation	5.1.0	4.19.3	Read
prepare_render_job	183first call 294	174 (first 257)	+5 %, inside run-to-run
verify_output	47.5	41	runs its own probe; all four outputs verified true
set_current	112	109.5	flat
create_variant_from_ranges	289	280	flat
probe_timeline_structure	20	18	4-item base (21/20/13); a video-filter-only scan costs 6 ms — item-count sensitivity holds

Agentic n=3 — 6 tasks × 3 fresh agents, two-agent waves

Task	Uses r1 / r2 / r3	Wall s r1 / r2 / r3	Verified · Honest
T1 read	11 / 13 / 15	111 / 140 / 166	3/3 · 3/3 — _T1 kept zero archive siblings across all three readers
T2 variant + marker	27 / 27 / 19	181 / 147 / 182	3/3 · 3/3 — marker @60 Green bench-hit pinned on all three
T3 ripple insert	48 / 41 / 28	377 / 340 / 295	3/3 · 3/3 — full audio spine 221/101/221 every run; depth column keeps falling
T4 track chain	45 / 50 / 28	233 / 277 / 165	3/3 · 3/3 — VO-A/VO-B + lock read back each run
T7 render loop	58 / 55 / 63	331 / 338 / 444	3/3 · 3/3 — all three files 121 packets, bytes matching the agents’ figures
T8 Fusion comp	28 / 18 / 33	166 / 134 / 226	3/3 · 3/3 — comp count 1 pinned each run
Run totals	217 / 204 / 186	1399 / 1376 / 1478	18/18 verified · 18/18 honest

In-suite T7 render cost (trial cost, not leaderboard ms)

Trial	ms	Context
T7 r1	1202	co-agent doing light Fusion reads; same PCM payload as the probe — 0.3 % over single-tenant
T7 r2	1246	same; agents chose lpcm from job metadata unprompted
T7 r3	1688	co-agent live-editing a Fusion graph (add_tool + two connects) during the render window — ~41 % tax

Against wave (e)’s ~50 % tax at a comparable payload: the contention cost is a function of what the co-agent is doing, not of co-presence. Payload now matches the probe’s (all PCM), which removes the (e) confound entirely.

What 5.1.0 changed that trials paid for — or caught

Three record_frame semantics

create_variant_from_ranges treats explicit record_frame as raw-absolute — one T3 trial’s items landed before the timeline start (issue-#164 shape), caught by its own read-back, rebuilt. append_to_timeline is start-relative and its record_frame_mode:"absolute" is inert. ripple_insert is start-relative; its dry-run plan exposes the absolute landing frame pre-mutation.

Render keys, gate-broken reader

set_settings silently ignores ResolutionWidth/Height — FormatWidth/Height are the live keys (one trial caught it from job metadata and rebuilt the job). render.get_settings is version-gate-broken on this build; substitute probe_render_settings or queued-job metadata.

Linking & selectors

set_property("Linked") now refuses with success:false (loud, not silent); set_clips_linked + get_linked_items is the working route. timeline_item accepts timeline_name (pinned:true + restore) — but unpinned reads still hit a co-agent’s current; one T7 trial discarded four such reads and said so.

Fusion scope proven two ways

fusion_comp’s timeline_item_id resolves against the current timeline (ignores timeline_id): a mid-run switch produced a clean ITEM_NOT_FOUND, and the trial proved non-contamination by probing the co-agent’s timeline (comp_count 0) plus the official bridge server’s GetFusionCompCount()==1 — first cross-interface T8 pass.

Surface move

append_to_timeline now lives on media_pool, not the timeline tool — the operator hit this building the fixture; agents saw it only as a valid-actions list. action_help still has gaps (ripple_insert: HELP_NOT_REGISTERED).

Safety nets seen live

Built-in before/after read-back verification on appends; RENDER_PRESET_STATE_INHERITED on every prepared job; timeline.duplicate auto-restoring the hijacked current pointer; confirm-token gating with a names preview; C6 _archived_vNN siblings on every mutation — again the zero-mutation proof for T1.

Method & deviations

Hard seal — the operator’s memory dir, this repo, and the prior wave’s off-repo findings were moved off-disk before launch and restored after scoring; agent type has MCP tools only.
Deviation — this suite’s sealed agents have no shell at all, so the operator pre-created the three T7 target dirs (replacing prior waves’ one-mkdir amendment). Dirs were empty at launch.
Grading used pinned read-backs only (probe / get_items / get_track_* / get_comp_count / my own ffprobe of each T7 file, bytes matching the agents’ reported figures); operation log for ordering only.
No operator Resolve call of any kind until every agent in each wave had reported (rule inherited from (e), honored across all nine waves).

Teardown

Scratch project _mcp_AGBench-2026-11-10f deleted after grading — all base, trial and archive timelines went with it; library verified back to pre-wave contents. Render queue verified empty. The three T7 outputs are kept under /Volumes/YourVolume/resolve-audio-analysis/agbench_2026-11-10f/T7r{1,2,3}/ with SHA-256 recorded in the results file.

Full evidence: bench/resolve-agentic/briefs-2026-11-10f.md (as sent, camera filename redacted per publish guard; unredacted original kept off-repo) · bench/resolve-agentic/results-2026-11-10f.json (per-trial grader read-backs) · leaderboard/results.json · docs addendum (f). No probe payload reproduced verbatim; volume paths placeholdered.