Audited Results

September 14, 2026 ยท View on GitHub

This document records the headline values currently used by the INTACT project page. It is a summary, not a substitute for the complete paper tables or raw evaluation manifests.

History-Free E1 Ablation

The released history-free actor removes previous action from both training and evaluation. With full task data, one epoch, training seeds 0, 42, and 3072, evaluation seeds 0, 1, and 42, and 100 episodes per cell, it reaches the following Official Direct results:

Released interfacePushTCubeReacherTwoRoomMacro
No previous action85.33 +/- 0.5899.44 +/- 0.3893.44 +/- 0.3898.78 +/- 0.6994.25 +/- 0.08

The checkpoint manifests and complete evaluation rows are available in INTACT-no-previous-action. This is a separate ablation cohort; the headline task-specific INTACT results are reported below.

Task-Specific Models

Protocol: full task data, one end-to-end epoch, training seeds 3072, 3073, and 3074 per task, evaluation seeds 0, 1, and 42 per checkpoint, and 100 episodes per evaluation seed. Success rates are percentages on the official LeWM protocol. The task-specific Guarded A row uses H=5, RH=5, 128 samples x 3 rounds, raw-action sigma=0.25, top-k 16, and the same causal action history as Direct.

Training coordinateInferenceSequencesPushTCubeReacherTwoRoomMacro
Waypoint INTACTDirect077.67 +/- 0.8899.89 +/- 0.1988.11 +/- 0.6998.00 +/- 0.3390.92 +/- 0.30
Goal-displacement INTACTDirect087.44 +/- 1.26100.00 +/- 0.0097.33 +/- 0.0097.67 +/- 1.2095.61 +/- 0.59
Goal-displacement INTACTPure CEM 300x309,00088.44 +/- 1.1768.44 +/- 0.7783.67 +/- 0.6782.89 +/- 0.8480.86 +/- 0.51
Goal-displacement INTACTActor-on CEM 300x309,00093.89 +/- 1.5897.67 +/- 0.6788.89 +/- 1.7198.00 +/- 1.4594.61 +/- 0.82
Goal-displacement INTACTGuarded A 128x338491.56 +/- 0.6999.67 +/- 0.3397.56 +/- 0.5197.56 +/- 1.2696.58 +/- 0.44

Interpretation:

  1. Goal displacement is the strongest final Direct coordinate.
  2. Broad actor-on CEM spends 9,000 sequences yet trails zero-search Direct.
  3. Small local verification improves Direct while preserving the learned plan.
  4. Pure CEM is intentionally retained as an actor-disabled representation control, not as INTACT's native deployment interface.

Shared Encoder

Protocol: one visual encoder is jointly updated by four task processes; forward and action heads remain task-specific. Values are means and sample standard deviations over three training seeds, each evaluated on seeds 0, 1, and 42 with 100 episodes.

Training cellEvaluationPushTCubeReacherTwoRoomMacro
LeWMCEM 300x3074.56 +/- 3.6767.33 +/- 1.8683.11 +/- 0.9639.67 +/- 8.3966.17 +/- 2.67
Goal intent onlyDirect81.78 +/- 0.96100.00 +/- 0.0088.67 +/- 0.3369.22 +/- 7.0484.92 +/- 1.86
Goal-displacement INTACTDirect86.11 +/- 0.96100.00 +/- 0.0097.22 +/- 0.8481.56 +/- 3.6791.22 +/- 0.51
Goal-displacement INTACTActor-disabled CEM78.11 +/- 1.1770.33 +/- 3.0080.33 +/- 2.0351.56 +/- 4.5370.08 +/- 1.13
Goal-displacement INTACTGuarded A 128x385.44 +/- 2.7899.22 +/- 0.3897.00 +/- 0.8880.44 +/- 3.3690.53 +/- 0.32

The shared-encoder Guarded A row is a separate E5 cohort; its 90.53% macro is not the task-specific one-epoch Guarded A result of 96.58% above. Its training seeds are 0, 42, and 3072. The complete 72-cell Guarded A identity, including per-seed SR and checkpoint SHA-256 values, is frozen in results/guarded_a_official.json.

Adding the attached physical call to matched goal-intent-only training improves the Direct macro by 6.31 points. The actor-disabled comparison improves LeWM by 3.91 points overall, showing that representation and direct readout effects are related but distinct.

Theory-Linked Diagnostics

Across 15 eligible goal-displacement E1-E5 shared-encoder checkpoints:

DiagnosticPearson correlation with Direct SRIntended role
Predicted-expert action-family kNN overlap0.968local family-neighborhood agreement
Predicted-expert linear CKA0.988global centered family geometry
Pointwise action R20.983single-action recoverability

Effective rank is reported as a capacity diagnostic, not a quality score. In a controlled E5 comparison, rank 93.87 accompanies 74.22% SR while rank 89.26 accompanies 91.22% SR. Higher latent spread alone is therefore not a semantic certificate.

Scope

  • Published external rows retain their original training and evaluation protocols and are landscape context, not paired controls.
  • Official LeWM SR and CLEAR-LeWM SR answer different evaluation questions and are never pooled.
  • +/- denotes sample standard deviation over training seeds after averaging each checkpoint's evaluation seeds.
  • Candidate counts are sequences per planning solve, not environment steps.