TacEx v4 · PegInsert · backbone generality

What carries tactile
outcome sensitivity?

Six policy families, two simulator-session seeds per study, and a 25× inference-depth reduction. The dashboard keeps every frozen verdict—and every boundary on what it means—visible.

Policy seeds 5–9 256 standard · rerun control 512 / seed-condition Non-privileged force history
BenchmarkTacEx v4PegInsert · t64
Session seeds / study2Separate S1 1000 · S2 2000 pair in each study
Paired episodes / family2,5601,280 per session and condition
Definitive new familyFlowDIFFUSION-LIKE in both sessions

01 · Family sweep

Architecture changes alone did not erase the pattern.

Points are pooled live-minus-ablated success counts from each family's own S1/S2 sessions; rows are not one shared cohort. Whiskers are exact paired-binomial 95% intervals. The orange guide is the frozen +30/1280 diffusion-like threshold; zero is the ACT-like boundary.

S1 S2 exact 95% interval +30 gate
Perception regime: onboard proprioception plus bilateral eight-sample PhysX force history (non-privileged). Policy seeds: 5, 6, 7, 8, 9. Each study has its own S1 seed 1000 and S2 seed 2000 simulator sessions; cross-family rows are unmatched context. Intervals: exact paired-binomial 95% confidence intervals.
FamilyParametersStudy session pairS1 Δ [95% CI] · signsS2 Δ [95% CI] · signsFrozen status

02 · Inference depth

One invalid probe, then one separately frozen answer.

The original study stays VOID. The later rerun is shown beside it, not blended into it: the higher-powered control passed and dp4 classified PERSISTS in both sessions.

A

What the rerun eliminates

The amount of inference-time iteration is not the load-bearing axis down to four DDIM steps.

B

What remains

Training-time corruption and noise-seeded generation remain bundled candidate mechanisms.

!

What this cannot establish

No zero-iteration claim, no universal policy theorem, and no certified tactile-reliance claim.

03 · Audit trail

The context travels with the numbers.

PerceptionProprioception + bilateral 8-sample PhysX force historyNon-privileged; no GT segmentation
Policy seeds5, 6, 7, 8, 9Every family and depth arm
PairingReset hashes → episode indexWithin-session contrasts only
UncertaintyPaired-binomial inversionEqual-tailed Clopper–Pearson 95% intervals
Frozen sources and artifact hashes

Family sweep

Artifact-only paper statistics. No study verdict was recomputed for this Space.

Open statistics JSON ↗

Original depth probe · VOID

f317700e…3465b02e15d44ecf…fbaaa86d Open frozen result ↗

Preregistered rerun · PERSISTS

9ba74fcf…15512061e9bd1d2d…5ac00e07 Open frozen result ↗