bench(worldmodel): the goal metric is also a defect — rho 0.25 at its best

WorldModel/EXPERIMENT_LEDGER.md

@@ -483,3 +483,22 @@ Two reviewers verified the delivered synthetic A/Bs are sound (symlog threaded i
 - 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.
+
+## 25 — The goal metric is ALSO a defect: the latent cost ranks at rho = 0.25 at its best (2026-08-02, commit <pending>)
+- Category: Benchmark
+- Hypothesis: (node 24) the bounded check before any design decision — node 22 showed the TRUE latent cost yields 76%, so Spearman(true latent cost, true position cost) should be substantial. If it is NOT, the goal metric is the defect rather than the predictor's use of it.
+- Prediction: substantial correlation, given 76% success.
+- Implementation: 150 episodes x 64 random candidates x 3 seeds, TRUE-env rollouts so no predictor is involved. Spearman within each episode between `||enc(s_H) - z_goal||^2` and `|pos_H - goal|`, for both goal encodings.
+- Evidence — **the prediction was wrong; the metric is weak even at its best**:
+
+  | goal latent | median rho | mean rho | rho > 0.5 | rho < 0 | node 22 success |
+  |---|---|---|---|---|---|
+  | one-frame (online goal) | **0.2514** | 0.2735 | 22.4% | 17.8% | 76.0% |
+  | incumbent (target goal) | 0.1570 | **0.0744** | 26.7% | **41.3%** | 37.5% |
+  | predictor cost (node 24) | **-0.018** | — | — | — | 35.0% |
+
+  - **The latent metric is a WEAK proxy at its best.** Even one-frame reaches only rho = 0.25. It still yields 76% because the CEM needs only to find a good candidate among 64, and a rho of 0.25 enriches the elite set enough — but the metric is not measuring what success measures.
+  - **The two-encoder frame does not merely weaken the ranking, it INVERTS it.** With the target-encoded goal, mean rho collapses to 0.074 and **41.3% of episodes have NEGATIVE correlation** — in four episodes in ten the incumbent cost ranks candidates backwards. That is the mechanism behind node 22's 76% vs 37.5%, and it is worse than "a weaker signal": it is a signal with the wrong sign nearly half the time.
+- Decision: **THERE ARE TWO INDEPENDENT COST DEFECTS, not one.** (i) the latent distance is a weak proxy for the control objective, rho ~ 0.25 at best, and (ii) the predictor destroys even that, rho -0.018. Node 24 correctly rejected forward-model ACCURACY as the lever and named rank preservation; this refines it. **A perfectly rank-preserving predictor on this metric would cap near 76%, not the 93.5% ceiling.** The remaining ~17.5 points are the goal metric's own weakness and no forward-model change can recover them.
+- **The complete measured chain, ceiling downward:** true position cost 93.5% -> true latent cost, one frame (rho 0.25) 76.0% -> true latent cost, two frames (rho 0.074, 41% inverted) 37.5% -> predictor cost (rho -0.018) 35.0%. Each step is a measured loss with a named cause.
+- Next question: unchanged in kind and now fully specified. This is the design decision node 13 reserved for the user, and both halves of it now have numbers. A control-aware goal metric must lift rho from 0.25; a rank-preserving training signal must stop the predictor driving it to zero. **They are separable and should be attempted separately**, since fixing only the predictor caps at 76% and fixing only the metric leaves the predictor destroying it. The one-frame goal is a prerequisite for either — it is free, it removes a 41.3% sign inversion, and nodes 22 and 25 both show it is necessary though not sufficient.