docs: update audited SR results

README.md

@@ -18,28 +18,21 @@ Search-Free World Models**.
 [Code](https://github.com/zju3dv/INTACT-JEPA) ·
 [Project page](https://zju3dv.github.io/INTACT-JEPA/)
 
-## Recommended releases
-
-| Release | Actor context | Evaluation contract | PushT | Cube | Reacher | TwoRoom | Macro SR |
-|---|---|---|---:|---:|---:|---:|---:|
-| [**INTACT**](https://huggingface.co/INTACT-JEPA/INTACT/tree/main/INTACT) | Causal previous action | Stateful continuation | 86.44 +/- 1.07 | 100.00 +/- 0.00 | 97.78 +/- 0.51 | 98.78 +/- 0.69 | **95.75 +/- 0.52** |
-| [**INTACT, no previous action**](https://huggingface.co/INTACT-JEPA/INTACT/tree/main/INTACT-no-previous-action) | None | History-free cold start | 85.33 +/- 0.58 | 99.44 +/- 0.38 | 93.44 +/- 0.38 | 98.78 +/- 0.69 | **94.25 +/- 0.08** |
-| [**INTACT, unified E5**](https://huggingface.co/INTACT-JEPA/INTACT/tree/main/INTACT-unified) | Shared encoder, task heads | Official Direct | 80.22 +/- 1.26 | 99.56 +/- 0.77 | 95.67 +/- 1.76 | 82.11 +/- 4.11 | **89.39 +/- 0.77** |
-
-The first two releases are task-specific, full-data, one-epoch models evaluated with
-Official Direct control, `H=5`, `RH=5`, and zero search. Each value averages
-evaluation seeds `0/1/42` within training seeds `0/42/3072`; `+/-` is the
-sample standard deviation across the three training-seed means.
-
-The unified E5 release is the paper's controlled four-task shared-encoder
-experiment. It uses one shared encoder with task-specific heads and is listed
-separately because its epoch count and model-sharing protocol differ from the
-two task-specific E1 releases.
-
-The first two variants isolate the previous-action question cleanly. Standard INTACT
-uses strictly causal pre-`t` history at an explicitly declared continuation
-point and only policy-executed actions afterward. The history-free actor omits
-the previous-action channel in both training and evaluation.
+## Audited results
+
+| Setting | Inference | PushT | Cube | Reacher | TwoRoom | Macro SR |
+|---|---|---:|---:|---:|---:|---:|
+| **Task-specific E1 INTACT** | **Direct, zero search** | 87.44 +/- 1.26 | 100.00 +/- 0.00 | 97.33 +/- 0.00 | 97.67 +/- 1.20 | **95.61 +/- 0.59** |
+| **Task-specific E1 INTACT** | **Guarded A 128x3** | 91.56 +/- 0.69 | 99.67 +/- 0.33 | 97.56 +/- 0.51 | 97.56 +/- 1.26 | **96.58 +/- 0.44** |
+| **Shared-encoder E5 INTACT** | **Direct, zero search** | 86.11 +/- 0.96 | 100.00 +/- 0.00 | 97.22 +/- 0.84 | 81.56 +/- 3.67 | **91.22 +/- 0.51** |
+| [History-free E1 ablation](https://huggingface.co/INTACT-JEPA/INTACT/tree/main/INTACT-no-previous-action) | Direct, zero search | 85.33 +/- 0.58 | 99.44 +/- 0.38 | 93.44 +/- 0.38 | 98.78 +/- 0.69 | **94.25 +/- 0.08** |
+
+All values use the corrected Official evaluation implementation. Standard
+INTACT uses strictly causal pre-`t` action history at a declared continuation
+point and appends every policy-executed action exactly once. The history-free
+ablation removes the previous-action channel from both training and evaluation.
+Guarded A is an optional local verifier around the Direct plan; it evaluates
+384 sampled candidates rather than the 9,000 used by CEM 300x30.
 
 ## Download
 
@@ -58,20 +51,18 @@ hf download INTACT-JEPA/INTACT --include "INTACT-unified/*" \
 
 Each directory contains a model-specific README, auditable manifests, and
 SHA-256 checksums. The unified directory additionally contains the complete
-six-cell, 72-checkpoint comparison matrix.
+shared-encoder comparison matrix.
 
 ## Unified E5 comparison
 
 | Controlled cell | Native interface | Official macro SR |
 |---|---|---:|
 | LeWM | CEM 300x30 | 66.17 |
-| Inverse only | Direct | 56.67 |
-| Waypoint intent only | Direct | 74.22 |
-| Goal intent only | Direct | 80.61 |
-| Waypoint INTACT | Direct | 76.42 |
-| **Goal-displacement INTACT** | **Direct** | **89.39** |
+| Inverse only | Direct | 68.19 |
+| Goal intent only | Direct | 84.92 |
+| **Goal-displacement INTACT** | **Direct** | **91.22** |
 
-All six cells use the same E5 shared-encoder protocol. Task-level results,
+All cells use the same E5 shared-encoder protocol. Task-level results,
 per-seed statistics, checkpoint hashes, and runtime fingerprints are provided
 inside [`INTACT-unified`](https://huggingface.co/INTACT-JEPA/INTACT/tree/main/INTACT-unified).
 The immutable original layout also remains available at the frozen