06Evidence
Dated gates. Published either way.
What the repository has committed to measure, by when, and what it has measured so far. Every gate and kill date on this page is a registry row; the results table renders only what the tree records.
Results so far: 22 rows
01The ruleWritten before.
Written before.
Printed after.
Cross-user generalization is the hard problem; the LOSO harness is the acceptance test.
A dated gate and a kill criterion are written before every experiment; the result is published whichever way it falls; we replicate before claiming and cite the paper before the pitch.
02The dated gatesFive product rows.
Five product rows.
Every gate with a date.
The products registry, as it stands: one closed status per row and the gate text as the row carries it. The ids, statuses and dates are the vision's; the attested by line names who ratified each row; a gate that is missed is published as missed and stays on this page. ADR-0136 suspends the paid gates of this phase. The research gates stay in force.
kit on a person with sync error printed; >=10 hours from >=5 participants with a dataset card; cross-user EMG->force model at <= 1.5 N pooled after the 15-minute calibration against a learned RGB vision baseline (LOSO); zero-calibration twin printed
one signed or paid team — suspended by ADR-0136
force-aware vs motion-only ablation on by-hand episodes, >= 40 trials per condition, one interleaved sealed schedule, hidden-condition scoring; PASS iff the one-sided 95% Newcombe lower bound is above zero and the observed gain is >= 15 points, a pre-registered minimum observed effect, not a certified margin; else the channel thesis is wrong; published either way (ADR-0053)
three paid collections, >= 100 hours, >= $50k cumulative (Q1 2027, rendered as the quarter's last day) — suspended by ADR-0136
a fleet's keep/discard labels show the band predicts discards earlier than jerk and completion time; otherwise 1b stays a column in 1a. The vision dates 1b only as a 6 -> 24-month window; this date is the kill date of the research project that decides the gate (decided_by), held equal by the drift gate — decided by
2026-11-demonstrator-state-vs-discards
the first paid collection delivered on bought sensors (ADR-0045: the kit is licensed to teams that already collect only after a collection has been delivered; the date is the last day of the Q4 2026 window in which the company's plan places its first one to three pilots -- a company plan, not a customer commitment, ADR-0122); a licensee's participants consent to the licensee under a scheme of its own — suspended by ADR-0136
1,000 hours with channels across >= 50 participants and >= 3 customers (Q4 2027, rendered as the quarter's last day) -- the dated corpus clause. The model clause -- model v1 beats video-only baselines on held-out people and scenes on a published benchmark -- carries no product date: its stop is the kill criterion of the pillar bet horizons/bets.yaml#human-model-paired-corpus (kill_by 2029-12-31). No training run before licence revenue exists (the licence-revenue rule)
licence revenue exists and >= 40% 90-day retention in a 1,000-user pilot before any scale spend (the 2030+ window rendered as the year's last day: the first review date, not a commitment)
- Solid lane: a gated row
- Construction line: a proposed row
- Open ring: a dated gate; the registry records no outcome for it yet
- Axis break: the gates after it are not drawn to scale
- Page built 2026-10-18 (UTC)
03The research registry34 projects, 2 active.
34 projects, 2 active.
Each born with a kill date.
A research project is registered with a bet, a numeric kill criterion and a kill date before it starts, and dies by a published verdict either way. This table is the generated registry (research/REGISTRY.yaml), read from the repository when the site is built: titles and bet ids are rendered as the registry writes them, as descriptions of protocols and bets, never as claims. A project past its kill date without a verdict is raised by the repository's weekly calendar. A project can have one dated extension. Its row then gives the extension date, then the original kill date and the reason.
active means the project's dated gate is live and its protocol is pre-registered (research/README.md, ground b), or data or a partner exists (ground a); no participant has been recorded by Cr4dle and no purchase is recorded. proposed means registered with its kill date, the synthetic smoke path running and the gate not yet live. The date column is the day the project was written into the registry, not the day work or data began.
| Project | Status | Registered | Kill by | Bet | Verdict |
|---|---|---|---|---|---|
2026-09-hyser-emg-force-losoHeld-out-user per-finger force from forearm HD-sEMG on the public Hyser corpus (external_corpus; a wrist-band emulation) | killed | Extension of the kill date : The owner decision of 2026-10-08 to continue with the pre-registered v2 criterion after clauses 1 and 2 of v1 were met. | open-corpus-semg-force-stack | verdict.md | |
2026-10-capsule-to-humanoid-retargetRetargeted skill capsules on the humanoid twin | proposed | capsule-to-humanoid-retarget | none yet | ||
2026-10-contact-sound-featuresHand contact-sound features against the fingertip-force rig | proposed | contact-sound-features | none yet | ||
2026-10-contact-truth-auditOpen contact-truth audit of force-labelled corpora | proposed | contact-truth-audit | none yet | ||
2026-10-electrophys-finger-decoderOne electrophysiology-to-finger decoder for ECoG, EEG and sEMG (BCI IV ds4 sanity leg, WAY-EEG-GAL EMG plus EEG grip-force leg, LOSO) | proposed | open-corpus-semg-force-stack | none yet | ||
2026-10-emg-force-open-stackEMG to per-finger force and contact on open corpora: baselines to EMG2Force and a self-supervised encoder, with a Hill-Kalman layer, LOSO | proposed | open-corpus-semg-force-stack | none yet | ||
2026-10-force-model-stage1Force model program, stage 1: the S1 screen on Hyser band8 | proposed | open-corpus-semg-force-stack | none yet | ||
2026-10-forceband-replicationPer-finger contact force from a bought sEMG/IMU band after a 15-minute calibration (ForceBand replication) | active | contact-force-wrist-semg | none yet | ||
2026-10-gaze-conditioned-il-replicationNatural gaze from bought eye-tracking glasses as a conditioning label for by-hand demonstrations (Ishida / Gaze2Act replication) | active | natural-gaze-target | none yet | ||
2026-10-gaze-predicted-vs-measuredMeasured Neon gaze against gaze predicted from the head camera and head pose: does the gaze-target label change on held-out participants? | proposed | measured-vs-predicted-gaze | none yet | ||
2026-10-insole-contact-moduleInsole foot-contact labels against video-derived foot contact | proposed | insole-contact-module | none yet | ||
2026-10-model-atlas-aModel atlas A, the force arms on main by LOSO on Hyser band8, NinaPro DB2 and putEMG | proposed | open-corpus-semg-force-stack | none yet | ||
2026-10-nail-force-moduleFingernail force module against band-only force | proposed | nail-force-module | none yet | ||
2026-10-odour-context-labelsChemical context labels from a wearable MOX array | proposed | odour-context-labels | none yet | ||
2026-10-sim-dynamics-pocOne-step action-conditioned object prediction on an uncalibrated simulator | killed | action-conditioned-sim-dynamics | verdict.md | ||
2026-10-sim-force-ablation-bcBehaviour cloning with and without the force column on the SO-101 twin -- the sealed force ablation on sim:mujoco with trained arms, and the sim-to-sim correlation (simulation; the protocol's dry run) | proposed | force-as-training-time-label | none yet | ||
2026-10-skill-capsule-learningSkill capsules with cues against video-only practice | proposed | skill-capsule-learning | none yet | ||
2026-10-socnav-walk-corpusLane-two walk corpus and the head and gaze ablation of a waypoint policy | proposed | human-social-navigation-lane | none yet | ||
2026-10-task-narration-labelsHuman task commentary against machine captions as keystep labels | proposed | task-narration-labels | none yet | ||
2026-11-demonstrator-state-vs-discardsDoes the demonstrator-state column predict a fleet's discarded episodes earlier than jerk and completion time? | proposed | demonstrator-state-discards | none yet | ||
2026-11-human-parity-v0-simHuman-parity trials v0 in simulation: can blinded judges tell a person's replay from a policy's rollout, and does the score track task success? | proposed | embodied-agent-human-parity | none yet | ||
2026-11-imagined-ablationImagined force-aware vs motion-only ablation inside an open world-model evaluator, scored against the real 2026-10-19 trials | proposed | world-model-evaluator-force-ablation | none yet | ||
2026-12-cosmos-force-conditioning-loraA LoRA adapter on open Cosmos Predict 2.5 weights conditioned on per-finger force: does removing force change plausibility on held-out users? | proposed | force-conditioned-video-world-model | none yet | ||
2026-12-force-as-training-labelForce as a training-time label: a gated residual per-finger-force branch on an open world-action model, scored by an in-domain evaluator on a held-out task (rented GPU; research-only) | proposed | force-as-training-time-label | none yet | ||
2026-12-reach-target-labelsReach-target prediction (kinematic): next-reach-target labels from hand trajectory plus gaze at 200 and 500 ms before contact, held-out participants | proposed | reach-target-labels | none yet | ||
2026-12-semg-effort-from-videoPredicting the band's sEMG-derived effort band (%MVC) from redacted video and pose on held-out participants | proposed | semg-effort-from-video | none yet | ||
2027-01-latent-wm-channel-probeDoes force (then gaze) carry information beyond pose in a frozen latent world model? A held-out-user channel probe | proposed | wm-human-conditioned-latent | none yet | ||
2027-01-pedestrian-gazePedestrian gaze and the gaze_on_robot label: gaze against head orientation on held-out people, with public social-navigation corpora for evaluation | proposed | pedestrian-gaze-on-robot | none yet | ||
2027-02-persistent-state-evaluatorDoes explicit object state outside the generator fix the what-memory gap on pick-squeeze-place? A re-observed-state consistency probe | proposed | wm-human-conditioned-latent | none yet | ||
2027-04-human-motion-to-g1-twinOwn-platform v0: human motion retargeted onto a G1-class twin in simulation, tracked on held-out clips | proposed | embodied-agent-human-parity | none yet | ||
2027-06-camera-side-reach-targetRobots reading humans: a third-person estimator of what is reached, when and how hard, supervised by wearable truth | proposed | human-model-paired-corpus | none yet | ||
2027-06-video-contact-liftContact force from bare-hand video, calibrated by the kit | proposed | video-contact-lift | none yet | ||
2027-09-agent-human-model-sim-hriBrains v0: the Human Model in the planner of a simulated home with a modelled human, judged by the hri_safety_ablation rule | proposed | embodied-agent-human-parity | none yet | ||
2028-06-own-platform-twin-fidelityOwn-platform v1: does the simulation twin predict the bought body? Real-to-sim correlation over tasks and seeds | proposed | embodied-agent-human-parity | none yet |
The bar under each kill date runs from the day the project was registered to its decision date (the extension date, if the project has an extension), on one axis from 2026-09-01 to 2029-01-01; the dashed line is the day this page was built, 2026-10-18 (UTC).
04Results so farWhat the tree records.
What the tree records.
Nothing else.
This table is generated from the repository's own record of results. A number appears here when a research project publishes its verdict or a results manifest carries an artefact, with the evidence kind and release eligibility its report prints. No number is quoted ahead of that record.
| Source | Metric | Value | Outcome | Evidence kind | Release-eligible | Published |
|---|---|---|---|---|---|---|
2026-10-sim-dynamics-pocverdict.md | Mean per-episode object-position RMSE at 20 ms, action-conditioned ridge, authored-physics held-out group (8 episodes) | 1.861 mm (action-blind ridge 4.540 mm, 59.0% lower) | The criterion is met in both held-out groups. The status killed closes the project after its one comparison. It is not a failure. | simulation | no | |
2026-10-sim-dynamics-pocverdict.md | Mean per-episode object-position RMSE at 20 ms, action-conditioned ridge, changed-physics held-out group (8 episodes) | 1.876 mm (action-blind ridge 4.434 mm, 57.7% lower) | The criterion is met in both held-out groups. The status killed closes the project after its one comparison. It is not a failure. | simulation | no | |
2026-09-hyser-emg-force-loso2026-10-08-5f7c0773 | Held-out-user pooled per-finger force MAE, 8-channel band emulation, calibrated against zero calibration at its own alpha, 20 Hyser subjects (kill clause 1)Data: Hyser (Jiang, Dai, Liu, Fan), PhysioNet hd-semg 2.0.0, doi:10.13026/hxan-pe94, with Jiang et al., IEEE TNSRE 2021, doi:10.1109/TNSRE.2021.3082551, and PhysioNet, under the ODC-By 1.0 terms. Corpus hyser, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED). | 3.4003 %MVC calibrated, 3.1943 zero calibration, 3.2223 constant floor; paired +0.206 [+0.113, +0.296], calibrated worse on 18 of 20 subjects | Kill clause 1 of v1 is met as written. Clauses 1 and 2 are met, and clause 3 is not met. On 2026-10-08 the owner chose to continue with a pre-registered v2 criterion. The project closed on 2026-10-08 with its v2 verdict. | external_corpus | no | |
2026-09-hyser-emg-force-loso2026-10-08-5f7c0773 | Held-out-user pooled per-finger force MAE, all 256 channels, calibrated against zero calibration at its own alpha, 20 Hyser subjects (kill clause 2)Data: Hyser (Jiang, Dai, Liu, Fan), PhysioNet hd-semg 2.0.0, doi:10.13026/hxan-pe94, with Jiang et al., IEEE TNSRE 2021, doi:10.1109/TNSRE.2021.3082551, and PhysioNet, under the ODC-By 1.0 terms. Corpus hyser, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED). | 3.4088 %MVC calibrated, 3.1091 zero calibration, 3.2223 constant floor; paired +0.300 [+0.143, +0.459], calibrated worse on 16 of 20 subjects | Kill clause 2 of v1 is met as written. Clauses 1 and 2 are met, and clause 3 is not met. On 2026-10-08 the owner chose to continue with a pre-registered v2 criterion. The project closed on 2026-10-08 with its v2 verdict. | external_corpus | no | |
2026-09-hyser-emg-force-loso2026-10-08-5f7c0773 | Next-day to same-day ratio of the calibrated pooled MAE, 8-channel band emulation, 20 Hyser subjects, day-1 MVC scale (kill clause 3, limit 2)Data: Hyser (Jiang, Dai, Liu, Fan), PhysioNet hd-semg 2.0.0, doi:10.13026/hxan-pe94, with Jiang et al., IEEE TNSRE 2021, doi:10.1109/TNSRE.2021.3082551, and PhysioNet, under the ODC-By 1.0 terms. Corpus hyser, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED). | 1.02 (session 2 3.424 %MVC, session 1 3.364 %MVC) | Kill clause 3 of v1 is not met: 1.02 is below the limit 2. Clauses 1 and 2 are met as written. On 2026-10-08 the owner chose to continue with a pre-registered v2 criterion. The project closed on 2026-10-08 with its v2 verdict. | external_corpus | no | |
2026-09-hyser-emg-force-loso2026-10-08-5eb4f25b | Held-out-user pooled per-finger force MAE of the nested-selected pipeline (C4 on 20 of 20 folds), 8-channel band emulation, 20 Hyser subjects, against zero calibration at its own alpha and the constant predictor (v2 clause A)Data: Hyser (Jiang, Dai, Liu, Fan), PhysioNet hd-semg 2.0.0, doi:10.13026/hxan-pe94, with Jiang et al., IEEE TNSRE 2021, doi:10.1109/TNSRE.2021.3082551, and PhysioNet, under the ODC-By 1.0 terms. Corpus hyser, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED). | 3.0767 %MVC selected, 3.1943 zero calibration, 3.2223 constant; paired -0.118 [-0.238, +0.037] and -0.146 [-0.271, +0.016], selected worse on 4 of 20 subjects in each | Clause (A) of v2 does not hold: the upper bound of each paired interval is above 0. Clauses (A) and (B) do not hold, and clause (C) holds. As written, the v2 kill condition is met. The owner closed the project on 2026-10-08 (status killed). | external_corpus | no | |
2026-09-hyser-emg-force-loso2026-10-08-d43a89d2 | Held-out-user pooled per-finger force MAE of C4 with alpha 1 fixed from Hyser, 24 channels, 44 putEMG-Force subjects, against zero calibration at its own alpha and the constant predictor (v2 clause B)Data: putEMG-Force (Kaczmarek, Mankowski, Tomczynski), Poznan University of Technology, with Kaczmarek et al., Sensors 2019, doi:10.3390/s19163548, under the CC BY-NC 4.0 terms. Corpus putemg-force, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED, commercial use FORBIDDEN). | 18.1116 %MVC selected, 18.7467 zero calibration, 19.7311 constant; paired -0.608 [-1.271, +0.170], worse on 12 of 44, and -1.601 [-2.290, -0.827], worse on 7 of 44 | Clause (B) of v2 does not hold: the upper bound against zero calibration is above 0. Clauses (A) and (B) do not hold, and clause (C) holds. As written, the v2 kill condition is met. The owner closed the project on 2026-10-08 (status killed). | external_corpus | no | |
2026-09-hyser-emg-force-loso2026-10-08-5eb4f25b | Next-day to same-day ratio of the pooled MAE of the selected pipeline, 8-channel band emulation, 20 Hyser subjects, day-1 MVC scale (v2 clause C, limit 2)Data: Hyser (Jiang, Dai, Liu, Fan), PhysioNet hd-semg 2.0.0, doi:10.13026/hxan-pe94, with Jiang et al., IEEE TNSRE 2021, doi:10.1109/TNSRE.2021.3082551, and PhysioNet, under the ODC-By 1.0 terms. Corpus hyser, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED). | 1.035 (session 2 3.1185 %MVC, session 1 3.0139 %MVC) | Clause (C) of v2 holds: 1.035 is below the limit 2. Clauses (A) and (B) do not hold, and clause (C) holds. As written, the v2 kill condition is met. The owner closed the project on 2026-10-08 (status killed). | external_corpus | no | |
2026-10-electrophys-finger-decoder2026-10-07-d2555347 | Leg 1: ECoG finger flexion (BCI IV data set 4), mean Pearson r over fingers 1, 2, 3 and 5 and 3 subjects, last 30% of each recording in time order (locked threshold 0.36)Data: BCI Competition IV data set 4, recorded by the University of Washington (K. J. Miller) and the Wadsworth Center, NYS Department of Health (G. Schalk). Paper: Schalk et al., Decoding two-dimensional movement trajectories using electrocorticographic signals in humans, J Neural Eng 4:264-275 (2007). Corpus bci-iv-ds4, evaluation only, because the terms grant no licence and their read by an agent session is not countersigned (register: UNCLEAR, UNVERIFIED). | 0.4305 (3-subject bootstrap [0.287, 0.587], descriptive only) | Leg 1 passes: 0.4305 is at or above the locked threshold 0.36. The threshold is inside the 3-subject interval. | external_corpus | no | |
2026-10-electrophys-finger-decoder2026-10-08-b623d26b | Leg 2: WAY-EEG-GAL grip force, change in mean per-trial Pearson r when EEG is added to 5 EMG channels, leave-one-subject-out over 12 participants (locked margin 0.02)Data: WAY-EEG-GAL (Luciw, Jarocka, Edin), Scientific Data 1:140047 (2014), figshare doi:10.6084/m9.figshare.c.988376.v2, under the CC0 dedication. Corpus way-eeg-gal, evaluation only, because the CC0 read of an agent session is not countersigned (register: UNVERIFIED). | 0.0138 (paired bootstrap [0.0064, 0.0213]; EMG alone 0.847, EMG plus EEG 0.861) | Leg 2 fails: 0.0138 is below the locked margin 0.02. EEG stays a recorded research_only column. | external_corpus | no | |
2026-10-sim-force-ablation-bc2026-10-07-d0bc4912 | Dense behaviour-cloning pair on the MuJoCo twin, task success with the force column against the column at zero, 40 sealed trials per arm, three training seeds (policy_ablation@3, minimum effect 15 points: a pre-registered minimum observed effect, not a certified margin) | seed 0: 23/40 vs 18/40, lower bound -5.722 (KILL); seed 1: 13/40 vs 39/40, lower bound -75.907 (KILL); seed 2: 36/40 vs 0/40, lower bound +77.737 (PASS) | The ablation rule fails on seeds 0 and 1 and passes on seed 2, so the verdict changes with the training seed. Kill clause 1 is met again by this run. The open clause is not changed. The ablation verdict is not the kill criterion. | simulation | no | |
2026-10-sim-force-ablation-bc2026-10-07-d0bc4912 | Sim-to-sim correlation of the seed-0 dense pair, authored MuJoCo twin against a perturbed MuJoCo twin, Spearman rho over 10 (bucket, arm) pairs (real_to_sim_correlation@1, floor 0.7) | rho 0.475 | The correlation rule fails: rho 0.475 is below the floor 0.7 (KILL). The correlation report is written, so its part of kill clause 1 is met. | simulation | no | |
2026-10-sim-force-ablation-bc2026-10-08-3e382b64 | Flow-matching chunk pair with the HumanEgo preset values on the MuJoCo twin, task success with the force column against the column at zero, 40 sealed trials per arm, training seed 0 (policy_ablation@3) | 40/40 vs 40/40, delta 0.0 points, lower bound -6.336 | The ablation rule fails: the lower bound -6.336 is not above zero and the delta 0.0 is below the minimum effect 15 (KILL). On this twin the column does not change task success. | simulation | no | |
2026-10-sim-force-ablation-bc2026-10-08-630235e7 | Flow-matching chunk pair with the HumanEgo preset values on the MuJoCo twin, task success with the force column against the column at zero, 40 sealed trials per arm, training seed 1 (policy_ablation@3) | 40/40 vs 40/40, delta 0.0 points, lower bound -6.336 | The ablation rule fails: the lower bound -6.336 is not above zero and the delta 0.0 is below the minimum effect 15 (KILL). On this twin the column does not change task success. | simulation | no | |
2026-10-sim-force-ablation-bc2026-10-08-4622f648 | Flow-matching chunk pair with the HumanEgo preset values on the MuJoCo twin, task success with the force column against the column at zero, 40 sealed trials per arm, training seed 2 (policy_ablation@3) | 40/40 vs 40/40, delta 0.0 points, lower bound -6.336 | The ablation rule fails: the lower bound -6.336 is not above zero and the delta 0.0 is below the minimum effect 15 (KILL). On this twin the column does not change task success. | simulation | no | |
2026-10-emg-force-open-stack2026-10-08-d131bfa6 | Held-out-user pooled per-finger force MAE of S1 (EMG2Force, fixed variant, rollout), Hyser band8, 20 subjects, arm cross_subject_loso_calib15 (the three conditions of kill leg (a))Data: Hyser (Jiang, Dai, Liu, Fan), PhysioNet hd-semg 2.0.0, doi:10.13026/hxan-pe94, with Jiang et al., IEEE TNSRE 2021, doi:10.1109/TNSRE.2021.3082551, and PhysioNet, under the ODC-By 1.0 terms. Corpus hyser, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED). | 2.953 %MVC [2.662, 3.215], constant floor 3.315 %MVC (0.891 x floor), R2 0.237 | For this arm, the 0.80 condition and the R2 condition of leg (a) are not met, and the interval condition is met. Leg (a) is not met over its selection set: the oracle bound is 0.884 x floor (run 2026-10-09-5d97b141). | external_corpus | no | |
2026-10-emg-force-open-stack2026-10-08-78b20df8 | Held-out-user pooled per-finger force MAE of B3 (LightGBM on lag-stacked Hudgins features), NinaPro DB2 force exercise, 40 subjects, arm cross_subject_loso_calib15 (the three numeric conditions of kill leg (b))Data: NinaPro DB2, Atzori et al., Electromyography data for non-invasive naturally-controlled robotic hand prostheses, Scientific Data 2014, doi:10.1038/sdata.2014.53. Corpus ninapro-db2, evaluation only, because the publisher states no terms and the licence read of an agent session is not countersigned (register: UNVERIFIED). | 1.196 [1.038, 1.369] in the unstated DB2 force unit, constant floor 1.511 (0.791 x floor), R2 0.375 | For this arm, the three numeric conditions are met. The inner selection chose this arm for 40 of 40 subjects, so leg (b) is met on NinaPro DB2 (run 2026-10-09-7cbb0acf). | external_corpus | no | |
2026-10-emg-force-open-stack2026-10-08-d0a3f693 | Held-out-user pooled per-finger force MAE of B2 (BE521 lag-ridge), putEMG-Force, 44 subjects, arm cross_subject_loso_calib15 (the three numeric conditions of kill leg (b))Data: putEMG-Force (Kaczmarek, Mankowski, Tomczynski), Poznan University of Technology, with Kaczmarek et al., Sensors 2019, doi:10.3390/s19163548, under the CC BY-NC 4.0 terms. Corpus putemg-force, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED, commercial use FORBIDDEN). | 130.45 [123.55, 137.54] in the released force unit, constant floor 154.44 (0.845 x floor), R2 0.272 | For this arm, the 0.80 condition and the R2 condition are not met, and the interval condition is met. B3 on putEMG-Force is NOT RUN. Leg (b) on putEMG-Force is not met over its selection set: the oracle bound is 0.843 x floor (run 2026-10-09-d3b33ae3). | external_corpus | no | |
2026-10-emg-force-open-stack2026-10-09-7cbb0acf | Held-out-user pooled per-finger force MAE of the inner-selected model (inner leave-one-training-subject-out selection over B0 to B3), NinaPro DB2 force exercise, 40 subjects, arm cross_subject_loso_calib15 (kill leg (b))Data: NinaPro DB2, Atzori et al., Electromyography data for non-invasive naturally-controlled robotic hand prostheses, Scientific Data 2014, doi:10.1038/sdata.2014.53. Corpus ninapro-db2, evaluation only, because the publisher states no terms and the licence read of an agent session is not countersigned (register: UNVERIFIED). | B3 (LightGBM) chosen for 40 of 40 subjects; 1.196 [1.038, 1.369] in the unstated DB2 force unit, constant floor 1.511 (0.791 x floor), R2 0.375 | The three conditions of leg (b) are met on NinaPro DB2, so the kill condition of the criterion is not met. The selection set was fixed after the outer numbers were known, the margin is thin, and the owner decides whether the bet continues. | external_corpus | no | |
2026-10-emg-force-open-stack2026-10-09-5d97b141 | Per-subject oracle bound over the selection set of kill leg (a) (B0 to B3, S1 quirk, S1 fixed), Hyser band8, 20 subjects, arm cross_subject_loso_calib15Data: Hyser (Jiang, Dai, Liu, Fan), PhysioNet hd-semg 2.0.0, doi:10.13026/hxan-pe94, with Jiang et al., IEEE TNSRE 2021, doi:10.1109/TNSRE.2021.3082551, and PhysioNet, under the ODC-By 1.0 terms. Corpus hyser, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED). | lowest pooled MAE / constant floor 0.884 (floor 3.312 %MVC) | Condition (i) of leg (a) is not met under any selection rule over the selection set, because the oracle bound is above 0.80. Leg (a) fails. | external_corpus | no | |
2026-10-emg-force-open-stack2026-10-09-d3b33ae3 | Per-subject oracle bound over the selection set of kill leg (b) on putEMG-Force (B0 to B2), 44 subjects, arm cross_subject_loso_calib15Data: putEMG-Force (Kaczmarek, Mankowski, Tomczynski), Poznan University of Technology, with Kaczmarek et al., Sensors 2019, doi:10.3390/s19163548, under the CC BY-NC 4.0 terms. Corpus putemg-force, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED, commercial use FORBIDDEN). | lowest pooled MAE / constant floor 0.843 (floor 154.44 in the released force unit) | Condition (i) is not met under any selection rule over this selection set, because the oracle bound is above 0.80. Leg (b) on putEMG-Force fails; B3 is NOT RUN. | external_corpus | no | |
2026-10-emg-force-open-stack2026-10-08-b3a0417a | Released ForceBand checkpoint through the S1 code path, 7 time-chunk validation sessions, teacher-forced mean per-finger MAE (a reference, not a held-out-user number; replication band [17.13, 18.13])Data: ForceBand (He, Wang, Kuang, Ghosh, Malik, Fermuller, Wu, Mao, Liu, Qi, Aloimonos), arXiv 2606.26093 (2026), release at commit 0fd984d2f79b, under the Apache-2.0 terms. Corpus forceband-release, evaluation only, because the licence read of an agent session is not countersigned (register: UNVERIFIED). | 17.6509 sensor units teacher-forced, 19.2783 in rollout | The replication band is met: 17.6509 is inside [17.13, 18.13]. ForceBand gives no held-out-user number, so this row is not a kill-criterion number. | external_corpus | no |
Every number printed here will carry its evidence kind, one of synthetic, simulation, hardware or external_corpus, and whether it is release-eligible. The rule is the benchmark harness's hardware_and_gate_passes: only a hardware result whose gate passed is; a number of any other kind never is. A dataset card says only synthetic, simulation or hardware; external_corpus is reported by research runs and never on a dataset card.
synthetic- produced by a generator or a fixture; the shape of an artefact, never a measurement
simulation- produced in the simulator or on the SO-101 twin; labelled, never release-eligible
hardware- produced on real sensors or a real body under a named protocol; release-eligible only when its gate passes
external_corpusresearch runs only- produced on a public corpus that others recorded on real sensors, evaluation only until a login countersigns its licence read, never on a dataset card, never release-eligible