← back
Robotics lab + worker review

TEST08 · from source session to trainable work

One evidence-backed understanding for lab evaluation. The recording worker can speak naturally at any exact moment; TradesAgent applies only role-scoped evidence and routes the rest.

ISAAC SIM · TEST08 · KIT PHYSX FILM

Real Isaac Sim playthrough · f4806–f13065

Left = Isaac Kit PhysX renders of the compose/source-fitted scene (real GPU film, not tip dots). Right = session DJI, scrub-synced to the film window. Not claiming 1:1 session-match — product credit stays 0. Tip residual overlay is demoted below (internal gauge only). Overlay scrubber under this panel is browser Three.js — never Isaac Kit. Published: 283 PASS chunks / 293 total.

KIT_FILM · REAL PHYSX · CREDIT 0
Honest now: real Kit PhysX frames from compose USD seats on densify span f4806–f13065 (80.2s–218.0s session, ~149 s film). Geometry is standardized OEM / densify seats — sparse cuboids + tip marker on a light void because site-scan room mesh is hidden on this film path. Hands/grasp are INFERRED spring/contact seats, not MEASURED CAD 1:1. Tell us what looks wrong (camera, missing room, floating parts, wrong span) and we fix that next ~15 min.
ISAAC KIT · PHYSX FILM · f4806–f13065
SESSION DJI · scrub-synced to film window
Source DJI video missing
Physics modeSPRING_TIP_HAND_CONTACT_FREE_LUMBER
Tip drive truthSPRING_TIP_HAND_CONTACT_FREE_LUMBER
Film frame spanf4806–f13065
Session clock80.2s–218.0s session
Chunk progress283 PASS / 293 published · cache=1786440038
Product credit0
NoteFounder-playable Kit PhysX playthrough from DJI ego (fov_y≈60.0°, identity densify CV + rigid tip Δ — not overhead Kit). Tip spring-tracks densify seats; world wrist from PrimaryHandRoot; kinematic hand on WRIST_TIP_BLEND (INFERRED). Tip↔lumber filter truth_state=TIP_LUMBER_FILTER_ON. CONTACT_CARRIES_GRASP=False; lumber_shove=False (max_xy_drift=0.0000 m). Browser Overlay ≠ Isaac. Session-match credit still 0.
Free Orbit · compose seats (secondary · browser, not live Kit stream) · StaticRoomMesh

Explore densify compose USD seats with site-scan floor/walls. Reset to DJI snaps browser camera to densify ego. This is not the Kit film above.

FREE ORBIT · drag / zoom / pan
room mesh · floor · site support
t=210.58s · span 28.01s–217.98s · DJI fov=60.0°

FREE ORBIT explores compose OemVisual + site-scan StaticRoomMesh (floor/walls) in the browser. Reset to DJI restores densify ego pose/FOV. Not a live Kit viewport. Site-scan StaticRoomMesh (ARKit/LiDAR) provides floor/walls orientation. Lingbot retired. Bodies: tip, work_surface, driver, lumber, lumber, lumber, phone, lumber, lumber, floor, fastener, fastener, fastener, fastener, fastener, room, site_support. Withheld seats stay hidden (no ghost floats).

Tip residual gauge (secondary · NOT the product) · seal v8 · FALSIFIED · mean 44px · credit=0

Internal PhysX tip vs held-out DJI tip dots on f12574–12642. Useful as a residual ladder — not a founder-facing Isaac Sim deliverable.

ISAAC EVIDENCE · PHYSX vs DJI · seal v8

Sealed PhysX tip prediction vs held-out DJI tip on frames f12574–f12642. Amber = session tip · cyan = PhysX tip. Overlay honesty only — credit stays 0 until session-match accepts.

FALSIFIED · credit=0
mean residual
44.3 px
worst
129.4 px @ f12642
tip path pred / obs
120 px / 237 px
pass bar
≤ 30 px
Tip residual chain: v3 mean 68 / worst 238 → v4 mean 142 / worst 182 → v5 mean 135 / worst 185 → v6 mean 135 / worst 185 → v7 mean 76 / worst 232 → v8 mean 44 / worst 129
Source DJI video missing — cannot draw tip overlay.
f12574 · Δ 6.2 px · paused
What good looks like: predicted tip (cyan) tracks observed tip (amber) within ~30 px across f12574–12642; credit would still require accepted session-match transaction
What you’ll see now: still FALSIFIED — mean ~44 px, worst ~129 px @ f12642; tip path short vs session (~120 vs ~237 px)
Next defect: PHYSX_TIP_LATERAL_STILL_SHORT_AFTER_DRIVERLINK_ROTATION_COUPLING
DriverLink tip-owner sampling resolved frozen quaternion / axial-only tip path. Pred tip path 120.1 vs obs 236.9 px; worst 129.4. Next: finish clean-prefix PhysX chain with tip-owner sampling, or Foundry full tip SE3 lateral (not axial-only) if residual remains dominated by missing world tip velocity.
SOURCE + OVERLAY · secondary · browser Three.js · NOT Isaac Kit
UDL playhead · test08

Source video + browser Overlay · one scrubber

Full engagement 3D coverage · 0:28–3:37 (190.0s) · 99.9% engagement · f1679–13065. Right pane is browser Three.js Overlay (densify/Foundry bins) — NOT Isaac Kit / NOT PhysX. Match from DJI ego; free-orbit secondary. PhysX evidence lives in the panel below when artifacts exist.
CameraCompare
Failure chips (shown on purpose)
INFERREDDriver tip (densify / partner splice)INFERREDScrewdriver body · INFERRED · tip+hand/lumber (6537 measured)INFERREDInsert mesh (fastener window orientation)INFERREDHEMNES long rail A · INFERRED · densify axis seat (1811 measured)INFERREDHEMNES long rail B · INFERRED · densify axis seat (9042 measured)INFERREDHEMNES ribbed panel · INFERRED · silhouette-pass densify (4683 measured)INFERREDiPhone 15 Pro Max (iPhone16,2) · INFERRED · dense ARKit×BOUND (1807 journal / 8487 carry; 6 depth-sync)INFERREDHEMNES short end A · INFERRED · densify axis seat (3317 measured)INFERREDJoint 1 · 109053 · INFERRED · JOINT_TIP_TRACK (measured=1234)INFERREDJoint 2 · 109053 · INFERRED · JOINT_TIP_TRACK (measured=959)INFERREDJoint 3 · 109053 · INFERRED · JOINT_TIP_TRACK (measured=1304)INFERREDJoint 4 · 109053 · INFERRED · JOINT_TIP_TRACK (measured=296)INFERREDSite-scan room · MEASURED metres / camera bound to the scan (gravity agreement 3.10°)INFERREDSite-scan support table · MEASURED metres / camera bound to the scan
1. Source (authority)
Source video missing
2. Browser Overlay (Three.js) · NOT Isaac PhysX · DJI ego
OUTSIDE PEAK SPAN · show best available
0:28–3:37 (190.0s) · 99.9% engagement · f1679–13065. Scrub freely — empty poses mean no estimate yet, not a UI withhold.
0:00/ 0:00Densify contact-chain 3D (hands → tip/tool → member) in DJI ego frame
Jump densify
CLIENT INSPECTION · EVAL REPAIR GUIDANCE · TEST08

Teachers + routers — not gates that hide geometry

REAL procedures (AT↔3D, WHY↔AT, contact↔3D, visual-geometry + specialist evals). FAIL/GUIDE explain what’s wrong, how to fix, mark UDL, and delegate. Truth stays INFERRED · not MEASURED 1:1 · never a UI gate.

CLIENT_INSPECTION_AGREE

overall=CLIENT_INSPECTION_AGREE · AT↔3D guide=PASS raw=AGREE agree_rate=1.0 · VISUAL_GEOMETRY guide=PASS peer_fails=0 · evals teach+route · never blank UI · truth=INFERRED · not MEASURED 1:1

AT_3D_AGREEMENT
guide:PASSAGREE

If the client builds AT from the 3D render alone, do they match the UDL-authoritative action timeline?

Why OK · AT_3D_AGREEMENT OK — agree_rate=1.0 (teach signal only; not a publish gate).

agree 100.0%flags 0repair → component_foundry + isaac_testing
WHY_AT_SPAN_AGREEMENT
guide:PASSAGREE

Do WHY claims bind real AT spans with matching frame windows?

Why OK · WHY_AT_SPAN_AGREEMENT OK — agree_rate=1.0 (teach signal only; not a publish gate).

agree 100.0%flags 0repair → action_timeline (+ world_data for WHY prose)
CONTACT_3D_AGREEMENT
guide:PASSAGREE

Where contact claims are active, is tip↔OEM workpiece surface near-contact in 3D (not tip↔center)?

Why OK · CONTACT_3D_AGREEMENT OK — agree_rate=1.0 (teach signal only; not a publish gate).

agree 100.0%flags 0repair → component_foundry (+ isaac_testing for PhysX)
VISUAL_GEOMETRY_TRUST
guide:PASSAGREE

Ignoring span-agreement badges, do visual-geometry specialist evals say the 3D looks seatable (lumber/hands/fuse/LiDAR/coverage)?

Why OK · VISUAL_GEOMETRY_TRUST OK — agree_rate=1.0 (teach signal only; not a publish gate).

agree 100.0%flags 0repair → review (rollup) → owning specialist per peer eval
apps/web/public/lab-clips/client-inspection/test08.json · 2026-08-05T19:40:36Z
WORK ATOMS · FOUNDER FEEDBACK · TEST08

Judge the question against the real video

Context spans may be wide for review; the amber mark is the published exact onset (tip-touch / alignment / prep) when the producer ships it — click seeks there. Playhead auto-selects the containing context span (prefers nearest exact onset on overlap). Outcomes are live receipts — 1 SATISFIED · 99 UNSUPPORTED_EVIDENCE · 30 WITHHELD.

bar WITHHELD
28 instances · 130 conditions
fps 59.940

28 candidate work-atom instances; 1 conditions satisfied, 0 violated, 30 withheld

FOUNDER CORRECTION / TRUTH NOTE
  • capture_purpose=ACTIVE_PHONE_LIDAR_SWEEP confidence=0.98CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=STAGED_PHONE_SECOND_ANGLE confidence=0.94CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.78CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=GROOVE_SEAT confidence=0.86CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.78CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.78CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.78CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.78CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.78CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.78CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.78CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=PRODUCTION_ASSEMBLY confidence=0.74CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=FINISHED_ORBIT_FOR_MODELING confidence=0.95CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
  • capture_purpose=FINISHED_ORBIT_FOR_MODELING confidence=0.93CAPTURE_PURPOSE_PRIOR · capture_purpose_prior
Video missing — apps/web/public/clips/test08/raw.mp4 does not exist. Use /api/lab/video/test08 when staged.
WHAT BUYERS WOULD NOTICE
  • Fingers / hand matchPASS

    Do tracked hands/fingers line up with the video, especially around the tool tip?

    HAND_TRACK_FIDELITY: PASS
  • Object floatPASS

    Do objects hang in mid-air or drift without support / contact?

    INANIMATE_GRAVITY_SUPPORT_SEAT: PASSOBJECT_TRACK_FIDELITY: PASS
  • Contact / force plausibilityPASS

    When the video shows contact or drive torque, do contact/force claims look physically plausible?

    CONTACT_EVENT_CORRECT_OR_WITHHELD: PASS
  • Range sanityFAIL

    Are clearances, depths, and relative poses inside believable physical ranges for this joint?

    OBJECT_TRACK_FIDELITY: PASSWORK_ATOM_CONDITION_GATE: FAIL
Filter
apps/web/public/lab-clips/work-atoms/test08.json · index 0ae86c917d89… · 2026-08-08T22:39:04.118405Z · no session-match, physical, or product credit is claimed by this receipt
DECISION LEDGER · ATTEMPT → ERROR → RECOVERY · TEST08

What failed, and whether recovery is evidenced

Play or scrub — the ledger entry whose exact attempt span contains the playhead is selected automatically; selection clears outside any span. Click still seeks. Recovery stays UNSUPPORTED / UNKNOWN until a later satisfied atom binds it — never invented. 28 UNKNOWN.

28 entries · 0 observed recoveries
0 observed failures · 20 unobservable drives
fps 59.940

28 decision-ledger entries; 0 observed fastener/align failures; 0 observed recoveries; 20 drive attempts UNKNOWN (missing tip/contact/torque)

Video missing — apps/web/public/clips/test08/raw.mp4 does not exist. Use /api/lab/video/test08 when staged.
Honesty rails: no recovery is invented when recovery_observed is absent · UNSUPPORTED library recovery_id is not an executed corrective action · UNOBSERVABLE drive attempts are not silent successes. 0 OBSERVED recoveries on this clip (honest).
Filter
apps/web/public/lab-clips/decision-ledger/test08.json · content 0b776bd48bec… · 2026-08-11T08:00:33Z · no recovery is invented when recovery_observed is absent
Questions0 of 7 reviewed

3 ready now · 4 more unlock as you finish — we show a few at a time to keep this quick.

  • Clip not available — apps/web/public/clips/test08/raw.mp4 does not exist
    What we already worked out

    Densify shows support near insert; load-bearing vs idle is still worker-unique.

    • Your partWe can't tell the why from the video — that's the part only you have.
    around 00:01:41.001 · if any of this is wrong, just say so
    The one thing we need from you

    On joints 1–2 your free hand looks near the insert — what was it doing, and was it carrying load into the table or a member end?

    • Why usYou were there — this is the part the camera can't show. No wrong answers, and your boss never sees this.
    • Your answer updates1 claim — nothing publishes on its own; a promotion check runs on your answer first.
    • We needfree text journeyman answer
  • Clip not available — apps/web/public/clips/test08/raw.mp4 does not exist
    What we already worked out

    Vision only hypothesizes bit/head match; seating feel is worker-unique.

    • Your partWe can't tell the why from the video — that's the part only you have.
    around 00:03:28.258 · if any of this is wrong, just say so
    The one thing we need from you

    Bit looked seated cleanly in the far-end screw — what catch or slip feel told you that tip matched the head?

    • Why usYou were there — this is the part the camera can't show. No wrong answers, and your boss never sees this.
    • Your answer updates1 claim — nothing publishes on its own; a promotion check runs on your answer first.
    • We needfree text journeyman answer
  • Clip not available — apps/web/public/clips/test08/raw.mp4 does not exist
    What we already worked out

    Hand-start vs power-drive encodes strip/cross-thread judgment.

    • Your partWe can't tell the why from the video — that's the part only you have.
    around 00:03:14.011 · if any of this is wrong, just say so
    The one thing we need from you

    You hand-started the far-end screw before power drive — what alignment or strip risk were you managing?

    • Why usYou were there — this is the part the camera can't show. No wrong answers, and your boss never sees this.
    • Your answer updates1 claim — nothing publishes on its own; a promotion check runs on your answer first.
    • We needfree text journeyman answer
Contact / friction (honest intervals)INFERRED — not MEASURED force gloves

INFERRED material-pair and contact-event intervals only — not force-glove MEASURED readings Intervals are inferred intervals available; worker force remains UNKNOWN.

Material-pair priors: 7 pairs · truth INFERRED

3:28-3:28INFERRED3:28-3:28INFERRED3:28-3:28INFERRED3:28-3:28INFERRED

No force arrows are drawn on the video. These badges are INFERRED interval claims only, not measured glove output.

Signal evidence behind these questions0 worker-check candidates · 17 action reads
What we already worked outcandidate action reads · worker can correct them
0:00-0:05initialize phone and board synchronizationINFERRED

best read: initialize phone and board synchronization. WHY (inferred): bind sealed mobile evidence to the continuous source session. Effect: source and mob…

0:05-0:28perform synchronized room and environment sweepINFERRED

best read: perform synchronized room and environment sweep. WHY (inferred): establish a measured support world before assembly. Effect: epoch-1 depth, mesh…

0:28-0:38return to the glass table and stabilize the work viewINFERRED

best read: return to the glass table and stabilize the work view. WHY (inferred): prepare a stable source view for object presentation. Effect: the action-…

0:38-1:05present four wooden organizer membersINFERRED

best read: present four wooden organizer members. WHY (inferred): make the assembly members individually inspectable. Effect: two rails and two two-hole en…

1:05-1:17present the corrugated insertINFERRED

best read: present the corrugated insert. WHY (inferred): separate the captured base from the structural members. Effect: the insert receives a persistent …

1:17-1:20record the mobile world epoch resetINFERRED

best read: record the mobile world-epoch reset. WHY (inferred): preserve non-native continuity without discarding the measured cross-epoch join. Effect: th…

These are current best reads, not final labels. A worker answer can split, merge, reject, or promote each span.