Learn the whole movement.
Keep correcting as it happens.

Fifty milliseconds here means simulated movement time, not a measured inference speed. First measure the full sensing, policy, 60-step prediction, selection and dispatch cost on the intended hardware. Select the number of candidate futures that fits that budget. In a simulator, inference may take longer in wall-clock time while simulation waits; record that slowdown. A physical real-time system needs a qualified deadline and a trained feedback controller that can continue if planning misses it. A frozen long command is not an acceptable substitute.

Physics still uses its own smaller integration step and contact constraints. Three seconds covers this example, not all tasks. A learned estimate of benefit after those three seconds keeps the planner from treating the horizon as the end of life.

Visual continuity. Use the temporal visual representation to carry appearance and motion information across successive eye images. For this design, train that representation using prediction of later visual features and consistency of feature correspondence forward and backward between nearby frames. Eye and head motion estimates enter that comparison so the model does not confuse moving its eyes with moving the object. Keep ambiguous correspondence marked uncertain instead of inventing an object identity. The stored-outcome record identifies an experience, not an object; an identity-swap test must separately verify that temporal perception can follow the same object through a partial occlusion. No explicit simulator object ID is supplied to acting inputs.

State and actions. Let s contain the causal temporal state, previous intention, physical estimates, command-filter state and wellness history. Let g be a selected intention from a finite memory shortlist that includes the previous one. Let a be the effective motor command. Store both the previous intention and selected g. The next real state includes g as its previous intention. Reset flags and actual duration are part of every transition.

Policy. Use a categorical distribution over the finite goal shortlist. For motor requests, use a diagonal Gaussian with learned mean and scale conditioned on s and g, transformed into the existing bounded request ranges. Although the random samples are independent given that state, their means share the same temporal state and goal. The resulting effective action also depends on the recorded command filter. Preserve the old motor weights as the initial mean where the schemas match. New variability parameters must respect the existing command ranges. Initially the finite candidate set is the existing stored-outcome ring, plus the carried intention if it has aged out of that ring. Do not introduce a new nearest-neighbour cutoff that silently removes the current intention. Each goal record contains its actual observation reference, timestamp, body/world identity, observed sensory/body/contact features, producer/model versions, and validity flags. Its learned embedding comes from re-encoding that observed outcome; it is not a target joint-angle script. The goal head computes a logit from current state and each candidate embedding, then normalizes those logits to select a candidate. These outcomes guide proposals but do not forbid an exploratory command from discovering a previously unseen useful result. Add that result only after it is actually experienced. The candidate set uses past experience only, with stable record identities and versioned representations; imagined outcomes are not inserted as real memories. Missing experience is acquired through the existing bounded exploration route, not a fabricated grasp example.

Open objective decision—required before implementing this learner. The source’s immediate motor drive uses the sum of grounded wellness outputs. Its normalized goal rank instead uses Σ w_c(stored_c − centre_c)/max(scale_c, floor_c), where each stored channel has already had its recent running average subtracted. Harmful remembered moments may receive −∞. These have different meanings. Neither is automatically a suitable finite reward to add across time. The existing source records a historical raw-score dominance problem, so reverting to the raw sum is not a neutral choice.

The architecture below requires a finite, time-defined utility derived from the existing channel producers. Its final formula is deliberately not approved here. Codex owns this decision, with Claude executing any later authorized saved-record analysis. The next design decision is to replay both existing scores on the same saved passive, exploratory and useful-control windows, identify each channel’s level/rate/event units and history dependence, and select the specification-consistent cumulative objective. The comparison must include sustained useful control: subtracting a running average can make a maintained benefit disappear from a ranking even while it remains useful. Keep raw harm limits outside the critic’s finite numeric target. Record any change in priorities as a project decision. This is an unresolved integration choice, not something more training will settle by itself.

Value learning, once that utility contract is fixed. Use two action-value networks Q₁(s,g,a), Q₂(s,g,a) and slowly updated target copies. For a real transition, use:

y = actual interval utility + discount(actual duration)
    × E[current goal and motor policy at next real state]
      min(Q₁_target(next state, next goal, next command),
          Q₂_target(next state, next goal, next command))

Train both critics toward y; do not send gradients through the target. On a genuine terminal state, omit the continuation. An intention switch is not a terminal state. On an artificial logging cutoff, retain the continuation when its next state is available. Define utility and its time units from the existing wellness contract; a rate is integrated over actual duration, whereas an interval measurement is not multiplied by time a second time.

Policy learning. Increase the current policy’s expected min(Q₁,Q₂), evaluated on real-state prefixes. Sum over the finite goal shortlist and sample the continuous motor distribution. Motor gradients pass through a differentiable version of the actual command decoder and filter. Keep decoder parameters fixed during policy optimization. First prove that its output agrees with dispatch and check its local derivatives against actual command changes; matching outputs alone does not validate gradients. If a required filter is not differentiable, use a likelihood-ratio policy gradient based on the probability of the sampled request u, with Q evaluating its effective action a = filter(u,s), instead; do not pretend the unfiltered request was executed. An exploration-entropy term is an explicit, separately logged training setting, not a new wellness channel.

This is a one-step off-policy learning target: experience collected by the planner can teach the current policy. Delayed benefit travels backward through repeated value updates across the recorded sequence. Do not substitute an uncorrected long return collected by a different policy. Twin critics reduce some overestimation; they do not certify unfamiliar actions. Compare their estimates with real outcomes and restrict unsupported planner proposals to the qualified exploration envelope.

World-model learning. Reconstruct state from a real past prefix. Train one-step and multi-step predictions of later sensory/body/contact features with the recorded effective actions and durations. In the multi-step branch, feed predictions forward rather than the correct next observations. Include every state element consumed by the next policy or wellness calculation, including history-dependent wellness terms. Predict continuous joint positions/velocities, root pose, proprioceptive and visual features with conditional continuous-distribution losses; predict contact occurrence with a categorical loss and touch magnitude conditional on contact. Train against the corresponding later measured readings or their declared sensory-encoder features. Sample distinct contact possibilities when rolling forward rather than averaging touching and not-touching into one supposed physical event. Advance temporal sensory memory, root/joint estimates, contact history, eye pose, command-filter state, previous intention and all wellness-producer histories used by the next step. Known filter updates remain exact computations, not extra neural guesses. Test predicted uncertainty on sequences excluded from fitting. Use channel scaling from the existing observation contracts; do not silently trade away touch error because vision has more numbers.

Planning. Sample K continuations from the same goal and motor policies, repeatedly advancing the complete predicted state for 60 steps. Include a deterministic policy continuation as a comparison: at each predicted step choose the maximum-probability goal and the bounded motor mean, then apply the same command filter. Score each with predicted interval utility plus discounted terminal value of continuing the base learned policy. That terminal value is not a claim about an optimal future planner. For each calibrated horizon and contact condition, subtract the held-out 95th-percentile absolute error of the predicted finite-window utility sum from that predicted sum. For the three-second subtotal, compare with actual readings over the same three seconds. Add a qualified terminal continuation estimate only as described below, then choose the largest resulting estimate. This percentile is a proposed engineering choice, not a confidence guarantee for an individual trajectory. Log the held-out sample count and uncertainty of that percentile. Define those conditions before comparison using measured joint/speed ranges, contact class and prediction horizon in the calibration set. A candidate outside those recorded conditions has no qualified planning advantage. An observed three-second sum cannot certify an unlimited future value. Check the terminal value separately against measured later windows collected under the base policy. Once the utility contract gives finite interval bounds and a discount below one, bound the remaining discounted tail from those bounds; otherwise keep that terminal estimate out of qualified planner ranking. A bootstrapped critic target is not an observed return. Log finite-window prediction error and continuation uncertainty separately, rather than calling their sum a measured full-return error. Recalibrate after model or objective changes. If no candidate is qualified, continue the learned feedback policy within its existing physical limits and collect bounded real exploration; do not freeze learning until a grasp is already known. Send the chosen first command through the single dispatcher, and observe again. K, the discount and learning weights are declared configuration choices; measure computation and compare them before adopting values.

Learning ownership. Prediction losses update the world model and its representation. Real TD errors update the critics. Critic-guided policy losses update the goal and motor policies. Initial ownership is explicit: prediction and retained representation-learning losses may update the shared temporal core; policy and critic losses receive detached core outputs and update only their own heads. First update prediction from a real sequence batch. Re-encode its past prefixes with that version. Update critics, then policy heads, then slowly update the target encoder and target critics after finite accepted updates. For the TD target, the current goal/motor policies choose the next goal and request from current-encoder next-state features. The target critics evaluate that same selected goal and effective command using target-encoder state and goal features. Store record IDs to make this correspondence exact. Do not feed the target encoder’s lagged coordinates into the current policy. Never mix a stale stored latent vector with newly changed encoder coordinates. Re-encode stored goal observations under the declared encoder version too. Keep the existing representation losses that prevent a constant, uninformative representation from appearing to predict perfectly. Preserve old replay, optimizer state and model versions. Check retained abilities whenever representation updates change the meaning of stored outcomes.

Compare the preserved baseline, learned feedback policy alone, and that same policy with three-second planning. Start from matched saved physical states. Use an object just beyond arm-only reach and allow whole-body motion. Evaluate with learning disabled after acquisition, then repeat after checkpoint reload and with held-out positions and sizes.

Proposed physical success: lift the ball centre at least 2 cm above its resting height and retain it for at least 1 second with no stand contact, within the existing body/contact limits. Count all attempts, including drops. Record joint forces, allowed contact penetration, object motion, support, correction latency and guard intervention. Add small disturbances at varied times. Report success counts with uncertainty, and improvement versus actual training experience—not just lower prediction loss.

Compare a short-horizon and a three-second-horizon planner with the same trained components to test whether lookahead helps. Compare real observation feedback against a fixed precomputed sequence to test correction. These are proposed later comparisons, not a new diagnostic launch list. Their exact sample count, runtime and stopping decision must fit the named question’s cumulative budget; a longer run requires the existing permission process.

Independent architecture review · Final document review and corrections · Independent clarity and screenshot review · Desktop and narrow browser checks · Writing guide.

Source inspected in the preserved worktree: D:/UEMA_wt/causal_loop_20261025/Source/Python/lem/model/motor_controller.py, forward_model.py, intention_generator.py, and mimo_bridge/lem_model.py, native_contract.py, sequence_shadow.py. The source findings above describe inspected routes; the proposed connected learner is not presented as already implemented.

Document work: Codex; initial reporting bound 20 minutes, extended by 10 minutes to resolve the independent review findings. No completion estimate was invented. No model code, saved learning, original report or pose asset is modified by this document. Browser checks inspect this page only.