The learned self-perception component

Self perception, named SelfState in the code, learns a changing representation of bodily condition. It supplies one part of the wider self described above. The body schema tracks control, and the longer-term self mechanisms track relations with wellness.

The self-perception network receives skin readings, motor capacity and recent sensory history. Motor capacity means how much of its normal motor force is available. Sensory history includes what its senses, such as vision and touch, have recently reported. It keeps a learned internal state: numbers that carry what recent experience says about this body.

A command is a set of numbers sent toward the motors. A timed command sequence is a list of those commands and how long each lasts. A forecast network uses the self state and that sequence to predict changes in bodily function. These self forecasts concern sensing quality and motor capacity; the separate movement predictions describe where body parts will move.

Read the loop: readings → self state → forecast → later measurement → learning → changed network weights. A weight is a stored number that learning changes. The report ends at the researcher; the learning arrow returns to the networks.

This is more than a checklist. The learned state combines inputs over time and relates them to earlier self states. The displayed condition fields are only some of its outputs. In the recorded grounded_residual setting, current function is anchored to measured function; the learned forecast predicts how it will change.

How self perception can affect movement

The forward model combines recent sensory information. Its prediction heads turn the combined information into predictions. The motor controller is a learned network that proposes commands. The planner compares command predictions before choosing a command.

The exact inputs, outputs and learning operation

SelfState contains small encoders, a recurrent network that carries recent state, learned attention over prior self states, and current/future function decoders. Each skin site's current function has three fields: sensing resolution, mechanical load and stimulation. Motor groups contribute available capacity. The future predictor reads the initial state, commands and their actual durations; it also predicts uncertainty.

In grounded-residual prediction, predicted function equals initial measured function plus elapsed time times a learned rate of change, limited to the allowed range. Later observed function supplies the learning target. Self-state input copies are detached from the main model's learning graph. Its errors train the self networks; in learn mode their gradients are clipped separately from the main model's gradients.

Self learning is an optimizer operation, not another neural decision maker. During evaluation, weights stay fixed. The report is output data, not a model of conscious speech.

How the self extends across body, objects and time

A hierarchy has smaller parts inside larger parts. UEMA uses several different hierarchies. One groups body parts; another groups experiences across time. The name alone does not explain what a part learns.

These parts serve different jobs within a wider self. The body schema tracks immediate control. Named associations track what accompanies wellness over time. Self perception learns bodily state and command consequences. Latent-body learns how parts move together. None of their names alone describes the whole UEMA self.

Why the older diagram omitted parts

The detailed diagram hid disabled components, so the old hierarchical-self table disappeared. It also omitted latent-body, even though the saved launch arguments record shadow learning. The revised diagram shows the disabled table and the separate latent-body branch. The later causal-loop work's learned body head belongs to the forward/efference path; it is not this latent-body hierarchy.

The inspected planner searches through learned predictions

A command is the set of numbers sent toward the body's motors. The motor controller is a learned network that proposes one. The forward model and prediction heads predict what a command will cause. The intention planner changes the command numbers, predicts again, and compares the result with a goal. During that search, the learned network weights stay fixed.

The ordinary planner chooses a goal from remembered sensory moments ranked using wellness. It then searches for a command whose prediction matches that goal. A separate prospective_wellness setting can add predicted wellness to the search; it is off in the reviewed October 15 run. The self-choice route below is another distinct option.

This is UEMA's model inversion: start with a desired predicted consequence and search for a command that produces it. The search itself need not own a network; it uses the learned forward model. Section 10.4 also specifies a learned attention module to compare several imagined sequences. That module exists in the source as CrossSimulationAttention, but is not used by the inspected MIMo bridge path. Saying that all of UEMA's intention system has no networks would be incorrect.

Exactly what self-choice does

Self-choice is one optional implementation experiment. UEMA's spec calls for bodily consequences to influence commands. It does not require this particular six-candidate chooser. Its activation checks answer whether that implementation is useful; they do not define the whole solution.

It forecasts bodily function for six command sequences: the proposed sequence, a hold at the current pose, the remaining commands from the previous plan, and three sampled variations. It checks uncertainty and compares expected loss of sensing or motor capacity over time.

The additional learning-opportunity comparison

This optional comparison asks where recent experience has improved prediction while older predictions remain accurate. It first checks that predictions using the real command beat predictions with shuffled commands and predictions that simply repeat the present reading. It then scores each candidate's predicted sensory change against holding still, weighted by that measured learning progress. Only candidates whose predicted bodily function stays within the configured range can receive this preference.

The implementation uses the first command's sensory prediction, not a forecast of the whole movement's learning benefit. Its score is a reason to try an experience. Repeated physical ability tests must measure whether that experience actually teaches reaching or grasping.

This comparison uses learned forecasts, but its thresholds and comparison rule are supplied by code. The comparison rule is fixed; its predictions are learned. Its purpose is to test whether predictions of this body's consequences improve its choices.

When was this intended to be turned on?

The recorded plan was: first make bodily forecasts accurate, then compare forecast-guided commands with the existing actor and control comparisons on matched starting worlds. Enable the tested configuration after it improves useful choice, keeps earlier abilities, and works in held-out worlds — worlds not used to fit the model.

There is a specific current blocker. The preserved branch uses remembered hand goals. The frozen self-choice actor rejects that hand-goal stage because it does not yet include its calibrated behavior in the frozen actor. The fixed copy must preserve the hand-goal settings and state, then pass physical comparisons before the switch can work with this branch.

If this self-choice experiment is selected when model work resumes, its next step is to complete the fixed actor's hand-goal support, verify preservation of its learned state, and run the matched physical comparison. Selecting a different implementation of UEMA's intended loop would need its own explicit design. Model work remains stopped for this documentation review.

What the activation receipt checks

The receipt must bind the exact self weights, other perception/action weights, body-function settings and command timing. It requires physical action effects, useful choice/acquisition gains, replay verification, retention and held-out worlds. The newer receipt also requires twice the prediction-error upper bound to be below the measured separation between command effects. This asks whether the forecast is accurate enough to tell the choices apart. It is more specific than a low average prediction error.

A reach example: follow the information

1. Before leaning: read current skin contact, available motor capacity and sensory context. Self perception updates its internal state.
2. Before sending a command: the planner predicts consequences of candidate command numbers. A self-choice comparison, when enabled, also forecasts their bodily-function consequences.
3. While the body moves: actual contact and movement change the next sensory readings. Those readings, not a desired success label, become the next prediction targets.
4. After contact changes: compare earlier forecasts with what happened. Update the prediction networks. At the next permitted command update, planning can use the changed readings and learned predictions.

Which coordinated command to use must come from learned predictions and choice. The four steps above describe the flow of information; they do not supply a fixed recipe of lean, reach and close fingers.

Looking ahead does not mean waiting to react

The history window is how much of the past the model reads. The planning horizon is how far into the future it predicts. The command duration is how long an order is applied before it can change. The observation interval is how often new readings arrive. These are four different times.

For the requested three-second reach, the design is to predict the whole movement, execute a short part, observe again, and revise what remains. A half-second phase in an illustration need not make the body blind to changes for half a second. The reviewed October 15 run's short-horizon settings are recorded separately in the detailed map; they do not implement the later three-second design.

Why future readings must not leak into an earlier prediction

To predict what happens just before a fall, the model may use everything sensed before that moment. It may imagine what comes next, but it may not read the later landing from the training example. That would teach it to use information unavailable when it must act.

The current forward model blocks later times in each sense's time-attention path. Its cross-sense attention path can mix all times supplied to it. Predicting the newest next step from a past-only input window is valid. Training a prediction at every position in a longer recorded sequence requires a past-only input for each position, or a time mask on every path that mixes times. The decisive check is to alter later observations and confirm that earlier predictions remain unchanged. This protects prediction-based planning; it does not prevent the planner from imagining a future.

Knowing a condition and choosing a movement are different lessons

Suppose the hand is sore. Three questions follow: What is happening now? What would this movement do? Which movement should happen? Training a network to answer one question does not automatically teach the other two.

UEMA does not ban learned motor networks. Sections 9 and 27 explicitly include them. Section 1 rejects collapsing perception, prediction, self evaluation, intention and action into one direct observation-to-command policy; it also rejects a permanent reinforcement-learning substitute. Section 2 says wellness should guide which action is chosen. Therefore, a learned network that receives information about wellness is not automatically a violation.

A gradient tells how changing one input would change a computed value. Searching with a gradient changes the command being considered. Learning with a gradient changes stored network weights. The spec describes both learned motor networks and command search, but does not fully specify every motor-network weight update. Any proposed lesson must name which weights change and show how prediction, grounded wellness and intention remain involved.

Intrinsic means that the drive comes from the system's own condition, rather than task points such as "receive points for this ball." An intrinsic signal can feed either UEMA's prediction-and-intention route or a reward-trained action policy. Its source alone does not make the two architectures equivalent. Learning to predict wellness also needs a path by which that prediction affects commands.

An MLP is a small layered neural network. It can predict bodily condition, reconstruct motion, or produce commands. The important distinction is what its lesson changes it to do. Latent-body and a wellness-seeking command network can both use MLPs while learning different jobs.

What the reviewed motor network actually learns

The reviewed October 15 run uses wellness_mode=baseline. The motor network's movement lesson compares an earlier output with the pose the body later reached, with output-limit penalties. It also receives a wellness gradient from the present handle-position formula. Skin discomfort, lost capacity and body-map changes do not provide their own command direction through that gradient. Receiving that position gradient is different from training weights to maximize future wellness.

The loss manager also defines direct and forward wellness objectives. Direct replaces the pose-matching term with negative current wellness; forward adds negative predicted wellness alongside its pose term. Forward explicitly scores the controller output through learned predictions. The current-wellness input used by direct comes from a detached achieved pose in this bridge, so its name alone does not establish a learning path into command weights. The reviewed October 15 run uses baseline.

The current measured-wellness path uses formulas. A separate learned approximate evaluator copies five of their results and can evaluate predicted tokens. Additional goal-ranking channels are not all included in those five forecasts. Neither this evaluator nor SelfState alone is the full wellness-and-intention mechanism.

What humans suggest we should keep separate

Human brains do have circuits for movement and action planning. Brain imaging finds both areas linked to particular body parts and a network associated with whole-body action planning. Design implication: having movement networks is compatible with human biology; the study does not prescribe UEMA's exact networks or learning rule. Gordon and colleagues, 2023.

People also sense their internal body condition. This is called interoception. In a heartbeat study, activity in part of the insula was related to how accurately people sensed their heartbeat. Design implication: keep a learned body-state representation that can affect later perception and choice. Critchley and colleagues, 2004.

Pain intensity and how unpleasant pain feels can change separately. Physical pain and social rejection also have distinguishable brain activity patterns. Design implication: keep damage, lost function, unpleasantness and social context distinguishable before combining them for choice. One skin-pressure number should not stand for all of them. Rainville and colleagues, 1997; Woo and colleagues, 2014.

Reflex, learned movement, and conscious self

A reflex is a rapid response that does not require deliberate planning. The inspected code has a built-in gaze response: peripheral visual motion supplies an eye-orienting target that is blended into the gaze command. Its rule has no trainable weights. The reviewed run enables it.

UEMA section 9 instead aims for fast and careful responses to develop within one prediction-and-wellness mechanism, using different amounts of computation. The gaze rule is an implementation difference to evaluate, not evidence that the spec requires a permanent reflex layer. A fast learned movement is not automatically a separate reflex.

UEMA's self also includes predictions about its own limits

Section 23 goes beyond a body-state list. It asks the system to predict what will happen, how wrong that prediction may be, and how imagined wellness may differ from later real wellness. In the future Many Minds design, each subordinate retains its own prediction-and-wellness machinery for a mastered skill and can report an unexpected event. Predicting those reports adds another form of self knowledge.

The present self-function network already learns uncertainty: an estimate fitted to the size of its bodily-function forecast errors. That is one part of this idea. The inspected bridge does not expose the full set of section 23's own-error, imagined-versus-real wellness-gap and subordinate-alert predictions. These are distinct connections to map and test, not new names for the existing self-state numbers.

UEMA already proposes emotional mechanisms. Section 22 describes stress as prediction errors that remain despite attempts to resolve them, and grief as the sustained consequences of losing something on which learned predictions and wellness depended. These are the spec's hypotheses about emergent experience. They do not require a programmer to label a state "fear" or "grief."

To evaluate that account, follow a concrete loss: a useful thing becomes unavailable; predictions that depended on it fail; attempts to restore the expected outcome have consequences; effort and capacity change; learning changes later predictions and choices. The named-thing associations and bodily self forecasts cover parts of this chain. The full sequence needs evidence linking those changes over time. A brief prediction error, by itself, does not distinguish surprise from the proposed persistent distress.

A body perspective can develop through experience

A small child and an adult have different reach, strength, memories and expectations. For UEMA, the useful design is to carry current capacities and learned history into the self state. Then the same event can predict different consequences as the body and experience change. Age alone should not prescribe a feeling or a movement.

The design direction I recommend

1. Preserve the whole UEMA loop. Keep bodily state, control of body and objects, remembered relations, and prediction reliability connected to intention. Teach forecasts from the consequences that actually follow commands.
2. Keep wellness learning grounded in experience. Predict bodily condition and its consequences before ranking movements. Persistent damage and short pain events need distinct inputs and timescales, so a brief successful catch is not treated like the fall it prevented.
3. Make perception usable by intention. Compare the consequences of a full timed movement, including a temporary cost that prevents a worse outcome. Predict ahead, execute part, read the body again and revise the remaining commands. The optional self-choice experiment is one candidate implementation; it is not the definition of UEMA.
4. State every command-learning lesson openly. Name its inputs, time horizon, teacher and exact weight update. Check whether real and imagined consequences still guide action. A reward-trained replacement or a network that only copies planner commands changes the design; it must not enter under the name of a prediction repair.
5. Give emotional context a measurable account. Before adding a new preference, specify what persists in self state, which observations update it, and how its predicted consequences affect choice. Test changed context with the same bodily condition, and changed bodily condition with the same context. This separates context learning from relabelling physical pain.

The operator has opened the policy restriction for careful evaluation. That permits comparison of an explicit revision; it does not make the alternatives part of the existing spec. The default design direction here is to preserve perception, prediction, grounded wellness and intention, and name any proposed departure before adopting it.