bench(worldmodel): the fourth condition is rank preservation, not accuracy

WorldModel/EXPERIMENT_LEDGER.md

@@ -458,3 +458,28 @@ Two reviewers verified the delivered synthetic A/Bs are sound (symlog threaded i
 - Decision: **REVERTED per the registered rule, without tuning K.** `synthetic_1f.py` and `eeg_jepa.py` are unchanged on main; this node is the record. Reverting is also right on the merits: the change adds a corpus mode and two parameters for no demonstrated benefit, and node 20 set the precedent.
 - **NODE 22's THREE-CONDITION MODEL IS NOW IN DOUBT.** It predicted that treating the last untreated condition would approach 76%. The condition was treated — measurably, 5.7x — and control did not move. Two readings remain and they are distinguishable: either (a) the residual gap between 0.00546 rollout error and the true env's exact zero is still decisive, in which case control is a near-discontinuous function of forward-model error and the 76% arm was only ever reachable with perfect dynamics; or (b) the three-condition model is incomplete and a fourth condition is unidentified. Node 22's 2x2 cannot separate these because its true-env arm used EXACT dynamics, not merely better ones.
 - Next question: **measure control as a function of rollout error directly**, rather than adding another candidate fix. Interpolate between the learned predictor and the true env — e.g. roll the true env but inject calibrated noise at each step, sweeping the injected magnitude across the range spanned by 0.00546 (multi-step trained), 0.0231 (incumbent) and 0 (exact). If control collapses somewhere between 0 and 0.00546, reading (a) holds and the requirement is a far more accurate forward model than any training change here will deliver — which is itself a design finding about latent-space MPC. If control stays flat until the error is large, reading (b) holds and there is a fourth condition still unfound. This is one sweep with no new model and it adjudicates a question five nodes have now circled.
+
+## 24 — The fourth condition is RANK PRESERVATION, and MSE does not deliver it (2026-08-02, commit <pending>)
+- Category: Benchmark
+- Hypothesis: (node 23) node 22's three-condition model predicted that treating the forward model would approach 76%; it was treated (rollout error cut 5.7x) and control did not move. Two readings remained: **(a)** the residual gap between 0.00546 and exact zero is still decisive, so control is near-discontinuous in forward-model error; **(b)** a fourth condition is unidentified. Node 22's 2x2 could not separate them because its true-env arm used EXACT dynamics, not merely better ones.
+- Prediction: sweep control against INJECTED latent error on an otherwise-perfect rollout, spanning the measured error levels. If control collapses below 0.00546, (a) holds. If it degrades gracefully, (b) holds.
+- Implementation: true-env rollout, one-frame goal, Gaussian noise added to the encoded final latent at swept magnitude. Per-dimension sd converted to the reported MSE scale, and to a latent distance as `sqrt(d * MSE)` with d=32, against the measured goal signal.
+- Evidence:
+  - **Latent goal signal:** median `||enc(start) - enc(goal)||` = **0.5303**. The incumbent's rollout error corresponds to a latent distance of **0.860 — 1.62x the entire goal signal**; the multi-step-trained model's to **0.418 = 0.79x**.
+  - **READING (a) IS REFUTED. Control degrades gracefully:**
+
+    | noise sd | equiv MSE | error/signal | success |
+    |---|---|---|---|
+    | 0.000 | 0 | 0.00 | 80.5% |
+    | 0.040 | 0.00160 | 0.43 | 81.0% |
+    | **0.074** | **0.00548** | 0.79 | **77.0%** (multi-step level) |
+    | 0.100 | 0.01000 | 1.07 | 72.0% |
+    | **0.152** | **0.02311** | 1.62 | **65.0%** (incumbent level) |
+    | 0.250 | 0.06250 | 2.67 | 59.0% |
+
+    At the INCUMBENT's own error magnitude, random error still yields **65.0%** — where the real predictor yields **35.0%**. Magnitude is not the explanation. Even error 2.67x the goal signal leaves control at 59%.
+  - **THE FOURTH CONDITION IS RANK PRESERVATION.** Over 120 episodes x 64 candidates: Spearman(predictor cost, true LATENT cost) = **+0.389** median, but Spearman(predictor cost, true POSITION cost) = **-0.018** median — zero. The predictor ranks candidates acceptably in latent space and its ranking is uncorrelated with the objective that decides success. Only **60.8%** of episodes reach even rho > 0.3 against the latent cost.
+  - **Why injected noise is benign and the predictor is not:** random noise perturbs each candidate INDEPENDENTLY, so the ordering survives in expectation and the CEM's elite set stays informative. A predictor whose error is structured — concentrated in the 31 latent directions that do not encode position, while the one direction that does is swamped — destroys the ordering at ANY MSE. The two failure modes are not comparable by magnitude, which is precisely why five nodes of accuracy-improving repairs changed nothing.
+- Decision: **REJECT "improve the forward model's accuracy" as a route to control**, and with it the framing that has driven nodes 19-23. MSE over a 32-dimensional latent rewards average accuracy across dimensions dominated by the frozen factors; the planner needs correct ORDERING along the single direction that encodes position. Accuracy is neither necessary nor sufficient for that — 65% from noisy-but-rank-preserving, 35% from accurate-but-rank-destroying, at the same error magnitude.
+- **This vindicates node 13's parked conclusion, three years of nodes later.** Node 13 named "a control-aware goal metric" as one of two fixes and explicitly reserved it as a DESIGN decision for the user rather than an autonomous build. Nodes 14-23 pursued the other branch — representation, benchmark, frame, action channel, forward model — found five real defects, fixed four, and moved control by nothing. The measured reason is now on record.
+- Next question: this is a design decision, not a bounded build, and per the node-13 stop rule it belongs to the user. The measured requirement is a training signal that optimises RANKING with respect to the control objective rather than reconstruction error in latent space — e.g. a contrastive or ranking loss over candidate rollouts, or a cost anchored on a decoded task variable rather than on full-latent distance. One bounded thing IS available first and should precede any of it: verify the remaining link by measuring Spearman(true latent cost, true position cost) directly. Node 22 showed the true latent cost yields 76%, so that correlation should be substantial; if it is NOT, the latent cost is a weaker proxy than node 22 implied and the goal metric is the defect rather than the predictor's use of it.