Evidence & setup.

Reviewed against main revision 07e50d0985bb91c4f613017df7a30e9ea68394c0 on 18 October 2026. The public site is an aggregate presentation of retained records. It does not rerun training or drive hardware to produce website metrics.

Recorded highlights and distinct evidence domains

The newest media index adds real FLUX red-pen inserts on Haller: RAW-5k task LoRA, 16 actions per request. The team reports 3/5 successes in the early block and 2/3 in the 07:40 block. Five selected head/wrist videos are retained. These are small session blocks with successful highlights, not a randomized held-out benchmark or independently labeled trial set. They are separate from the older 16-trial index below, and do not establish learned bimanual task completion.

Gallery media reviewed against the team Drive on 18 October 2026: the 60-second project film includes its original speed and scripted-replay labels. The 32-second outside-view cutter recording shows grasp and carry toward the box, without a completed release. Both clips retain their full duration; the cutter clip omits ambient audio. They add no trials to the reported success counts.

The five-seed test, tabletop dataset and current red-pen deployment adapter are different experiments. The simulation comparison uses 20 Hz. The archived tabletop replay was recorded at 30 Hz and exported at 15 body-pose samples per second. The lab’s websocket delivery cadence is 30 Hz; this is not an inference-speed claim.

Media labels

• Real red-pen highlights: single-arm FLUX task-LoRA inserts on physical Haller. Session counts are attributed to media/README.md; the videos are selected successful attempts.
• Real-camera prediction: imagined versus actual recorded views on held-out red-pen take 101. These imagined frames are model output.
• Scripted simulation expert: a successful teacher trajectory, left arm parked and lid fixed open. This is training-data generation.
• Matched checkpoint videos: red-marker seed 901001, zero-shot versus final twin LoRA. Final gripper distance is 20.6 cm versus 2.2 cm. Neither video completes placement.
• Observed versus imagined: real FLUX inference in simulation, with an imagined future output. The imagined frames are predictions.
• FLUX session: policy integration with simulated arms, not physical task completion.
• Recorded-video analysis: VLM scene inspection of a human demonstration. No robot commands.
• Setup reference / rendered scene: independently captured and estimated views. Camera calibration, arm spacing and finger compliance remain uncertain.

The interactive dependency timeline illustrates behavior-tree rules. The six showcase ordering checks were symbolic tests using deterministic feedback and synthetic checkpoints. The red-pen adapter’s finite 42×6 GPU smoke output used blank images and no motor commands. Full learned dual-arm hardware task completion remains unverified.

Connect the public lab to your local robot

GitHub Pages serves static files. A local research companion reads the Haller controller’s measured telemetry and existing camera snapshots. It does not open another serial or camera device.

# In the Liminal source checkout with the research UI changes:
uv pip install --python .venv/bin/python -e '.[research]'
./scripts/research_ui.sh

Keep Haller running locally on port 8000. No profile is needed to view cameras and measured joints. To launch FLUX prompts, add --profile PATH_TO_VERIFIED_PROFILE with the current arm, checkpoint, worker, units and camera assignment. The repository's official_haller_right.yaml example uses a base checkpoint, rather than the RAW-5k adapter in the archived pen videos. The lab displays the configured checkpoint after pairing.

The live viewer loads the companion's compiled MuJoCo scene and applies measured joint poses to it. It freezes on stale telemetry and reconnects automatically. HTTP snapshots provide a fallback if the websocket is unavailable. Haller's source timestamps and receive age both contribute to freshness; wrist labels follow the verified policy binding.

1. In the full-screen lab, select Pair local robot.
2. Open the local pairing page at http://127.0.0.1:18003/pair.
3. Use its private link to return to the public lab. The one-session token is carried in the URL fragment, removed by the lab and retained only in memory. It is not sent to GitHub.
4. Allow local-network access if your browser requests it. The page must run on the same machine as the loopback companion; use a local SSH tunnel when appropriate.

Observation-only is the default. Sending actions additionally requires fresh telemetry, a verified profile, and MANUAL mode plus torque engaged through Haller. The lab does not engage torque or clear STOP. Stop this trial cancels the trial started by this companion; Haller remains the hardware stop interface. Disconnecting the browser does not stop a local trial.

Trials save locally under runs/research/ with the existing runner’s observations, actions and reports. No telemetry, camera stream or recording is stored on GitHub Pages. The two-arm viewer mirrors measured joints; live objects and contact forces are not tracked.

Public export and attribution

Download the result summary. It contains aggregate counts and repository-relative provenance only. Raw logs, profiles, serial identifiers, machine paths and credentials are excluded.

Hardware CAD: TheRobotStudio SO-101, model license. Renderer: Three.js 0.180.0, MIT license. Fonts: DM Sans and Instrument Serif, with their bundled SIL Open Font Licenses. Page references: DexMimic and Nerfies; the Liminal layout, text and interactions are original.