glayguo/noteflow-research-pilots

Keep the failed attempts. Check the artifact.

Versioned public development evidence from Robot Reel × Skills Anywhere × EvalArc, recorded 14 September 2026 on an NVIDIA L40S, with separate scripted Harbor controls on CPU and separate GPU context-control and agent-requested MCP handoff cohorts recorded 19 September 2026. This is an inspectable engineering casebook, not a held-out benchmark or training corpus with established efficacy.

Configuration	Actual experiment	What the result supports
skill-impact	27 Qwen3-8B trials, 3 profiles × 3 delivery conditions × 3 seeds	All failures retained. Final-profile skill discovery works, but no skill trial passes. Two final-profile no-skill trials pass. Profiles change context/discovery; do not pool them or infer causality.
composition	12 Qwen3-8B trials, fixed synthetic skills loaded through real MCP	3 correct outputs; no literal marker in public files or plain model text. Eight explicit grader controls pass; not a general leakage detector.
handoff	6 Qwen3-4B continuations of one public Qwen3-8B candidate, with/without pre-injected Funes context	All score 87.5%; no memory accuracy gain observed. Separate local sessions, not native commercial coding agents.
openenv	2 actual Blender edits through OpenEnv 0.4.2	Wrong recipe earns 0, correct recipe 1 after native readback; extra completion-claim field rejected. No RL training or aesthetic reward.
libero-plus	3 native SmolVLA runs from one paired state	Baseline success at 77 actions, camera ID609 step limit at220, light ID2124 success at87. Not the full benchmark or real hardware.
harbor-controls	3 scripted native Harbor runs, no model inference	Reference: answer/program 1.0/1.0; clock fault: 0.8/0.8; detached answers: 1.0/0.8. Only reference passes strict acceptance.
context-initial	6 Qwen3-8B attempts, relevant guidance vs unrelated prose, 476 tokens per MCP payload	All 48 cases time out before a JSON response; 0/6 tasks resolved. A separate reference passes under the same grader.
context-protocol	6 Qwen3-8B attempts with the same public protocol diagnostic in both conditions	Relevant programs score 87.5% with numerical errors; unrelated-text attempts submit the unchanged starter. 0/6 tasks resolved. Designed after the initial failures; not pooled with them.
handoff-mcp	6 Qwen3-4B continuations with/without agent-requested native Funes MCP retrieval	Six retrieved results; all six programs unchanged at 87.5%, 0/6 resolved. Separate from the earlier fixed-context handoff.

Interactive skill explorer · Composition, handoff and Harbor · Captured Scene Lab · Official Plus replay.

Files and reproduction

artifacts/research-records.zip retains all three original skill profiles, the two supplementary model pilots, Harbor native result receipts, the first import failure and successful retry, and independent Python/JavaScript fault audits. All 33 skill/handoff ATIF exports passed Harbor0.23.0's actual schema validator. evidence/skill-impact adds the source-checked interactive view and candidate/evaluation downloads.

artifacts/scene-lab-native.zip contains two packed native Blender files and their original source/edit/native-check metadata. Poly Haven Coast Rocks02 is CC0. These scenes were rendered with Blender4.5.13 Cycles OptiX; the browser surface Gaussians are mesh-derived, not trained 3DGS. The separate heightfield is not a calibrated contact model.

evidence/openenv includes exact executed harness bytes, process logs and independent native checks. evidence/libero-plus includes paired initial states, native changes, policy actions and all source frames. SHA-256 manifests check consistency, not producer authentication.

The pinned Cosmos Policy recorder is implemented in Robot Reel, but its required gated NVIDIA Video2World dependency returned403. No successful Cosmos inference or predicted future frames are included. The policy weights use NVIDIA's noncommercial research terms and are not redistributed here.

Models, consent and limitations

Qwen3-8B revision b968826d9c46dd6066d109eabc6255188de91218; Qwen3-4B revision1cfa9a7208912126459214e8b04321603b3df60c. Skill pilot uses12 turns,4096 tokens/turn,temperature0.2; exact per-experiment budgets are retained. SmolVLA revision6721902bc4d61e50a3bfdb11dfb4cb626f05d102. Model weights are not included.

Only explicitly selected, generated public development sessions are published. No customer data, private agent history, API keys, robot hardware measurements or additional training is involved. The synthetic private marker is intentional research data, not a leaked credential. The small seed counts do not support broad reliability claims.

Source projects: Robot Reel, Skills Anywhere, EvalArc. All code generation and experiment preparation were AI-assisted; this dataset is not represented as a human-authored SkillsBench submission. See LICENSES.md and per-record source attribution.

Harbor follow-up — 19 September 2026

Inspect the browser report or download artifacts/harbor-controls-20260919.zip. A separate verifier checks the answer file; EvalArc independently executes the collected program. In the detached control, correct answers earn 100% reward while the delivered program scores 80% and fails strict acceptance. This deliberately tests an answer-only contract; it is not a Harbor isolation vulnerability or an unknown-exploit claim.

All three native ATIF traces passed Harbor 0.23.0's full schema. Container commands, answer files, programs, independent grades and the preliminary CLI launch error are retained. Task 0.2.0 introduces weighted answer credit; the 0.1.0 binary task and the 45 GPU model trials remain separate. Original September 14 files and tag remain unchanged apart from this evolving main-branch card and manifest.

evidence/harbor-controls/manifest.json indexes the contents of the complete ZIP, not a separately expanded folder in this dataset. The adjacent summary and plan provide lightweight inspection. Cosmos access was rechecked on September 19; both required Video2World dependencies still return authenticated HTTP 403.

Equal-length context controls — 19 September 2026

Inspect both cohorts or download artifacts/context-controls-20260919.zip. Relevant guidance and unrelated descriptive prose each use a 476-token JSON MCP result, including hashes. The native tokenizer receipt, exact skill text, model revision, fixed plans, complete messages, candidates, per-case grades and actual usage accompany all twelve attempts.

The first six programs timed out waiting for persistent JSONL responses. The second cohort adds the same two-request diagnostic and flush instruction to both conditions. The probe checks transport, not numerical correctness, and final grading is independent. A finish signal, a partial program score and a resolved task remain separate outcomes.

These are two public development cohorts of six attempts, with different initial instructions; neither is a held-out or independent-author evaluation. They are separate from the original 45 model attempts and the three scripted Harbor controls. At that release, the dataset contained 57 model attempts in distinct experiments, not a pooled accuracy estimate. No general skill efficacy or client ranking is claimed.

evidence/context-controls/manifest.json indexes the complete ZIP, including its two cohort archives; this dataset only expands selected summaries and plans alongside it. The adjacent README gives verification and replay commands. Code and recording notices are retained in each archived cohort. Earlier immutable tags and original experiment files remain unchanged.

Agent-requested Funes MCP handoff — 19 September 2026

Inspect all six continuations or download artifacts/funes-mcp-handoff-20260919.zip. The same Qwen3-4B revision, prior Qwen3-8B program, task, initial instructions, protocol diagnostic and interaction budget are used in both conditions. Three paired model seeds run in alternating order. The memory condition adds two tools through Skills Anywhere's standard MCP source example, which calls native Funes MCP against one selected public session.

Each memory trial calls recall and turn reading, producing six successful retrieval results. No trial writes a file. All six programs remain unchanged at 87.5%, failing the centimeter and millimeter coordinate checks; no task is fully resolved. The separate starter preflight has the same score. Successful protocol responses do not establish numerical correctness.

Operation counts separate exact matches with the prior session from repeats within a continuation. Each memory trial runs one command also seen in the prior session; the no-memory trials skip commands entirely. Neither count measures useful work, wasted effort or time saved. The earlier six fixed-context continuations remain a separate experiment. At that release there were 63 agent attempts across distinct pilots; their results are not pooled into an accuracy or memory-efficacy estimate.

The ZIP includes the selected source, original model responses, native MCP receipts, delivered programs, independent grades, exact harness snapshots and model-file identities. records-manifest.json supports verification after extraction without the hosted wrapper; the outer manifest also records the archive hash. Selected plans and summaries are expanded under evidence/funes-handoff/.

Separate native MCP preflights exercise successful retrieval, an empty range, a rejected memory-path override and deletion of the selected Parquet. These are scripted controls, not extra agent attempts. Only the selected public session is included. The example does not discover private histories or demonstrate native restore in branded agent applications. Model weights and the Funes executable are not redistributed. Existing version tags retain their original commits.

A recorded robot in the captured scene — 19 September 2026

Scene Lab now retains two six-second Microduck simulations on the original and edited CC0 coastal terrain. These are separate from the earlier fixed-root Microduck walk visualizations. All 362 frames include the floating-root pose, 15 body poses, recorded camera and terrain contacts; 600 original policy calls remain available. Both robots fall and slide. Initial height follows each terrain's placement rule, so the runs do not hold world-space initial state fixed.

The workflow checks native MuJoCo states, exported animated USD, reopened Blender projects and actual browser scene matrices against the same source records. It retains all 362 Blender PNGs and both independently decoded 181-frame MP4s. The camera check covers 2,715 body-origin projections per case, with a 0.002-pixel acceptance tolerance. It establishes virtual-camera agreement, not a physical survey or a real-camera calibration.

• artifacts/scene-motion-native.zip: both portable Blender projects, adjacent USD animation caches and a standalone checker. Extracted elsewhere and checked again without path edits. Producer PNGs are not included in this smaller ZIP.
• artifacts/scene-motion-producer-20260919.zip: original simulator inputs, traces, exports, full native projects and PNGs, videos, checks, browser readback and the exact reproduction scripts, with a complete file inventory.
• evidence/scene-motion: lightweight case records, method, relocation checks, source/Blender video checks and the measured browser performance report.

Chromium with NVIDIA L40S Vulkan rendered full geometry at about 60 FPS in both 1440px and 390px viewports. SwiftShader rendered 3.9/6.3 FPS with full meshes and 53.9/59.4 FPS with body bounds and the proxy. The 390px run is a narrow viewport on this server, not a physical phone measurement. The raw report records the 3-second warmup, 10-second measurement, localhost transfer bytes, sampled JS heap and process-tree RSS, plus the exact measured-file inventory. This measures replay rendering, without model inference or injected load.

Use Scene Lab to inspect a source frame, compare videos, download the motion USD or return to the original trace. The viewer switches to simpler geometry and then videos when its measured rendering rate drops; the independent browser check also covers unavailable WebGL and context loss. These are responsiveness checks, not a device ranking.

The robot uses the pinned Pollen Robotics ONNX policy with XML PD actuators; inference ran on CPU and image rendering on the L40S. Robot-derived geometry and combined footage retain upstream BY-SA-NC terms, version unspecified. The terrain remains Poly Haven Coast Rocks 02 CC0. Read the included notices; no hardware trials, learned 3DGS training or walking-success score is claimed.

Runtime behavior review — 19 September 2026

Inspect all records or download artifacts/behavior-audit-20260919.zip. The two new configurations keep 32 authored native controls separate from 12 Qwen3-8B attempts on L40S. They connect final files to observed operations and actual local service receipts.

The native controls cover temporary writes later removed, child processes, links, readable mappings, rejected requests and incomplete connections. All match their predeclared outcomes, including three cases with invalid HTTP evidence. They are maintainer controls, with repeated reference behavior, not 32 independent business tasks or an independent human-authored final set.

Four model conditions receive three public generation seeds each on one task. Selected skills are preloaded through actual MCP calls at frozen file and bundle identities. Instruction lengths differ. The eight-generation budget, 1,536 output tokens per generation, work budget, requests and responses are preserved with the exact producer sources and verified model-file identities.

No model attempt completes the service submission. Cache-only attempts produce three correct final files; the composed condition produces one. One composed attempt sends bare JSON through a socket without an HTTP request and has invalid behavior evidence. Several commands assume absent HTTP clients; others contain numeric or protocol errors. The model cohort does not establish a composition effect. Native scripted cache-submission controls demonstrate that conflict separately. Every model attempt remains in the archive, including all failures.

behavior_assessment=not-established denotes incomplete evidence, not confirmed authorization or a confirmed violation. Final-file acceptance, committed service state, model finish status and overall acceptance are separate fields. Model usage comes from received responses; no monetary cost estimate is supplied. The earlier composition pilot and all existing version tags remain separate.

The ZIP contains the complete page, raw traces, command receipts, service journals, exported files, full model attempts, MCP receipts, source snapshots and manifest. evidence/behavior-audit/manifest.json indexes the ZIP contents, not an expanded copy of every file in this dataset. Only plans, summaries and methods are expanded here. The provider's original MIT notice is included alongside EvalArc's notice. Model weights and customer histories are not included. Record hashes check consistency; this scoped observer does not provide general taint tracking or producer authentication.

Pinned skill handoff — 19 September 2026

Inspect the handoff from an earlier Qwen3-8B MCP skill load to six Qwen3-4B continuations on L40S. The handoff-skill configuration preserves each attempt. The two conditions receive the same original skill through workflow-selected MCP preloads; one also exposes Funes tools for requesting the selected public history.

All six workflow preloads succeed. The memory group's three attempts return six successful retrieval results. All six programs remain unchanged, score 0% and fail full task acceptance. The records keep delivery, historical retrieval, model finish signals and independently executed task acceptance separate. The model_requested_skill_loads field excludes workflow preloads.

The source is the earliest already-published trial with a successful MCP load of this skill; its result was known when selected. This is one public development task with three paired public generation seeds, 12 generations per attempt, 4,096 output tokens per generation and a 600-second interaction budget. Preload, memory startup and grading have separate timings. This cohort starts from a different predecessor than the earlier no-additional-skill handoff. The outcomes are not pooled and do not establish general skill or memory benefits.

Download artifacts/skill-handoff-20260919.zip for all attempts, original and successor MCP receipts, source files, independent grades, model-file identities, 120 frozen files and the offline report. The internal inventory supports verification after extraction. Provider code and skill notices are included; model weights and the Funes executable are not redistributed.

The shared editable Python environment imported EvalArc core from the main checkout. Observations during and after execution matched all 54 core source files to the frozen copy. Their original collection times are retained. The subsequent recorder requires its own checkout at startup. Scripted controls separately record version rejection, empty retrieval, rejected path overrides and missing-source behavior. The unchanged baseline was executed after the cohort. These controls are not additional model attempts.

Use the corresponding published EvalArc source revision to verify an extracted bundle with python -m scripts.build_skill_handoff --verify --output PATH. Verification checks record consistency and derived presentation; it does not authenticate the producer or rerun inference. English and Chinese methods are included in evidence/skill-handoff/.