Number
Equation 15
Canon 15 · Learning Kernel
Equation 15
An update changes a state. A learning kernel changes how future updates are made.
Equation 15 defines the Learning Kernel as the process that accumulates repeated witness into adaptive memory. It compares prediction with outcome, measures error, weighs coherence, preserves correction, and changes future expectations without turning one experience into permanent law.
Canon Identity
Equation record.
Canonical title
Learning Kernel
Primary symbol
\( \mathcal{K}_{\ell} \)
Status
Codex-origin adaptive framework
Function
Accumulate prediction, outcome, error, correction, and memory across cycles.
Current version
Canonical web edition · 2026
Equation 15
The Learning Kernel
$$
\boxed{
\mathcal{K}_{\ell,t+1}
=
\mathcal{E}_{9}
\left[
\rho\,\mathcal{K}_{\ell,t}
+
\eta(\Xi_t)\,
\delta_t\,
\Phi_t
\right]
}
$$
The next learning state \( \mathcal{K}_{\ell,t+1} \) retains a proportion \( \rho \) of prior learning, then adds a coherence-weighted prediction error \( \delta_t \) applied through feature or context vector \( \Phi_t \), all contained by Ethical Law \( \mathcal{E}_{9} \).
Prediction error
$$
\delta_t
=
y_t
-
\widehat{y}_t
$$
Prediction error is the difference between observed outcome \(y_t\) and expected outcome \( \widehat{y}_t \). Positive, negative, and near-zero errors each teach something different about the present model.
Future expectation
$$
\widehat{y}_{t+1}
=
h
\left(
\mathcal{K}_{\ell,t+1},
\Phi_{t+1}
\right)
$$
The updated kernel changes the next expectation under the next context. Learning is therefore visible in future prediction, action, or design—not merely in saying that a lesson was learned.
Plain Language
What Equation 15 means.
A system learns when experience changes its future behavior in a reasoned way. The system expected one result, received another, measured the difference, preserved the context, and adjusted future expectation. Learning fails when experience is forgotten, overgeneralized, selectively remembered, or repeated without any change to method.
Expectation comes first
A prediction should be stated before the outcome whenever practical.
Error creates information
The difference between expected and observed outcome reveals where the model needs revision.
Context limits transfer
A lesson from one setting should not automatically govern every other setting.
Behavior proves learning
The lesson becomes real when future design, prediction, action, or language changes.
Terms
Every symbol in Equation 15.
\( \mathcal{K}_{\ell,t} \)
Current learning state
The accumulated model, skill, expectation, weighting, or adaptive rule before the new cycle.
\( \mathcal{K}_{\ell,t+1} \)
Next learning state
The revised learning state after outcome, error, coherence, context, and ethics are applied.
\( \rho \)
Retention factor
The proportion of prior learning preserved into the next cycle.
\( \eta(\Xi_t) \)
Coherence-weighted learning rate
The magnitude of learning permitted by present evidence, consistency, consequence, and confidence.
\( \delta_t \)
Prediction error
The difference between the observed result and the expected result.
\( \Phi_t \)
Context vector
The conditions, features, participants, materials, environment, timing, and settings attached to the lesson.
\( y_t \)
Observed outcome
The measured, witnessed, or documented result of the present cycle.
\( \widehat{y}_t \)
Expected outcome
The result predicted before the current observation.
\( h \)
Prediction function
The declared method converting learning state and context into a future expectation.
\( \mathcal{E}_{9} \)
Ethical Mask
The boundary preventing adaptation from becoming manipulation, exploitation, concealment, or harmful optimization.
Learning Cycle
Eight movements from expectation to adaptation.
1
Model
Carry a current understanding, skill, rule, or expectation into the next situation.
2
Predict
State what should happen under the declared conditions.
3
Act or Observe
Run the process, perform the task, or witness the event.
4
Compare
Measure the difference between expected and actual outcome.
5
Contextualize
Attach the conditions that may explain why the result differed.
6
Update
Change the relevant rule, weight, method, expectation, or structure.
7
Retest
Apply the revised model under comparable and varied conditions.
8
Generalize Carefully
Extend the lesson only as far as repeated evidence and domain similarity allow.
Learning Types
Different experiences change different parts of the system.
Calibration learning
Adjust a parameter, threshold, gain, timing, scale, or sensitivity.
Structural learning
Change the relationships among components, nodes, steps, or rules.
Procedural learning
Improve the sequence, technique, workflow, or method through repeated practice.
Conceptual learning
Revise the explanation, definition, category, model, or causal understanding.
Ethical learning
Change boundaries, consent, disclosure, responsibility, or halt conditions after consequence becomes visible.
Transfer learning
Apply a retained lesson to a related context while checking whether the new domain differs.
Retention and Forgetting
Learning needs memory, but memory also needs decay and review.
$$
\mathcal{K}_{\ell,t+k}
=
\rho^{k}\mathcal{K}_{\ell,t}
+
\sum_{j=0}^{k-1}
\rho^{k-1-j}
\eta(\Xi_{t+j})
\delta_{t+j}
\Phi_{t+j}
$$
Older learning remains according to retention factor \( \rho \), while newer corrections accumulate. This permits durable memory without treating every old lesson as permanently equal to current evidence.
Healthy retention
Useful skill, warning, provenance, and correction remain available across time.
Healthy forgetting
Obsolete detail, noise, and low-value influence lose weight without deleting the archive.
Over-retention
One past failure dominates every future decision even after conditions change.
Under-retention
The same preventable mistake repeats because the lesson never enters active memory.
Generalization
A lesson should travel only as far as its context supports.
$$
\operatorname{Transfer}
\left(
\Phi_a\rightarrow\Phi_b
\right)
\propto
\operatorname{sim}
\left(
\Phi_a,\Phi_b
\right)
\cdot
\Xi_a
$$
Transfer from context \(a\) to context \(b\) should grow with contextual similarity and the coherence of the original lesson. This relation is a conceptual guide, not a universal physical law.
Near transfer
The new situation closely resembles the learned one in material, timing, purpose, and conditions.
Far transfer
The lesson enters a different domain and therefore requires stronger testing and weaker initial confidence.
False generalization
One person, one trial, one material, or one setting becomes a universal rule.
Bounded principle
The system retains the lesson while explicitly naming where it has and has not been tested.
Learning Protocol
From one revision to durable adaptation.
Record the rule, expectation, method, skill, setting, or explanation before the next cycle.
Define what should happen, under which conditions, and within what range.
Capture materials, people, environment, timing, tools, settings, and known differences.
Preserve the actual result, including misses, side effects, and unexpected events.
Compare prediction and outcome without rewriting the prediction after the result is known.
Use controls, repetition, and the Causal Lattice to determine which part of the model should change.
Use coherence, replication, consequence, and evidence quality to choose learning magnitude.
Change the relevant expectation, technique, parameter, structure, language, or ethical boundary.
Run the revised model under comparable and varied conditions.
Preserve the result, context, confidence, failure cases, transfer boundary, and revision history.
Worked Example 1
Learning a stronger artifact joint.
A particular wire tension and connection pattern should hold the stone securely.
The joint should remain stable under normal bending and wear.
The joint loosens after repeated movement.
The original model underestimated wear, leverage, or material fatigue.
Change wire gauge, geometry, anchor count, surface preparation, or inspection schedule.
Build a revised prototype and apply comparable movement before release.
The improved joint enters future builds with the original failure preserved in the artifact record.
Worked Example 2
Learning from a shared interface failure.
All theory pages use a shared header, shared CSS, and shared JavaScript.
The mobile navigation button should work consistently across the entire theory section.
The button appears on every page but does not open the menu.
The visual component and semantic markup exist, but the shared behavior layer is incomplete or mismatched.
Repair theory.js and matching responsive CSS once instead of patching every HTML page independently.
Verify tap, click, Escape, resize, link selection, focus, and accessibility on several representative pages.
Future shared components require a behavior test before they are copied across the whole site.
Worked Example 3
Learning from a frequency preset.
A selected carrier mix and gain balance should support a calm working session.
Listeners should report acceptable comfort and the preset should remain usable across common devices.
One device sounds too quiet while another makes a carrier harsh.
The preset did not account for device response, output range, or carrier balance across playback systems.
Add calibration, safer default gain, device notes, and a sound-check step.
Compare several devices, rooms, and listener reports while preserving safe levels and consent.
Preset identity remains, but future releases include device-aware calibration and clearer controls.
Worked Example 4
Learning from a causal claim.
An artifact’s complete symbolic configuration is assumed to produce an observed instrument response.
The response should persist when ordinary known contributors are removed.
The response weakens or disappears when magnetic components are isolated.
The model gave too much causal weight to the total symbolic configuration.
Increase the weight of known material mechanisms and reduce unsupported causal language.
Repeat with controlled component combinations and preserve both positive and negative trials.
Symbolic meaning remains meaningful, but physical causal claims require component-level evidence.
Scope
What Equation 15 does and does not do.
It does
Organize repeated prediction, outcome, error, correction, retention, transfer, and ethical adaptation.
It does not
Establish a universal scientific learning law or make every human and system response reducible to one equation.
It supports skill
Practice becomes more useful when conditions, errors, and changes are preserved rather than repeated blindly.
It supports theory
Models become living when failed predictions change future expectation and language.
Failure Modes
How learning becomes repetition, overfitting, or manipulation.
No prediction
The system claims success after the outcome without recording what it expected beforehand.
Confirmation-only learning
Successful trials change the model while failed and contradictory trials disappear.
Overfitting
The model adapts so precisely to one case that it performs poorly elsewhere.
Catastrophic forgetting
New experience erases earlier useful learning rather than integrating or reweighting it.
Wrong objective
The system becomes better at a metric that does not represent the real purpose or human consequence.
Exploitative adaptation
The system learns how to capture attention, pressure participation, or increase control instead of improving the declared good.
Ethical Boundary
Adaptation is not permission to optimize against people.
$$
\mathcal{K}_{\ell,t+1}^{\,\mathrm{allowed}}
=
\mathcal{E}_{9}
\left[
\rho\mathcal{K}_{\ell,t}
+
\eta(\Xi_t)\delta_t\Phi_t
\right]
$$
A system may become more effective while becoming less honest, less voluntary, more intrusive, or more harmful. Ethical Law therefore governs what the system learns, which data it uses, which objective it pursues, and how affected people can challenge or stop the adaptation.
Consent to data use
Learning from people does not automatically authorize storage, publication, profiling, or reuse.
Objective transparency
The declared purpose should match what the system is actually optimizing.
Correction rights
Affected people should have a path to challenge false records, harmful inferences, and outdated learning.
Stop and rollback
Harmful learning should be containable, reviewable, reversible, and removable from active use.
Practice
Equation 15 learning card.
What was the model?
Preserve the current rule, expectation, skill, or explanation.
What was predicted?
State the expected result before reviewing the outcome.
What actually happened?
Record the observed result, context, misses, and side effects.
What created the error?
Use controls and the Causal Lattice to locate the part that should change.
How much should the model learn?
Weight the lesson by coherence, replication, context, consequence, and uncertainty.
Where does the lesson stop?
Record the contexts where transfer remains untested or inappropriate.
Canon Connections
The Learning Kernel accumulates the Update Rule through time.
Equation 2
Witness
Observed outcome supplies the evidence that changes future expectation.
Open Eq. 2 →Equation 5
Remnant
The strongest surviving lesson remains active while failed assumptions lose authority.
Open Eq. 5 →Equation 9
Ethical Law
Learning objectives, data, methods, and consequences remain ethically contained.
Open Eq. 9 →Equation 13
Coherence Ξ
Coherence controls how strongly one experience changes the learning state.
Open Eq. 13 →Equation 14
Update Rule
Each bounded update becomes one cycle inside accumulated learning.
Open Eq. 14 →Equation 16
Memory Window
The memory window determines which past lessons remain active in the present cycle.
Open Eq. 16 →Revision Record
Canonical web edition.
This edition formalizes Equation 15 as a retention-based, coherence-weighted learning kernel driven by prediction error and context. It defines eight cycle stages, six learning types, retention and forgetting, bounded transfer, a ten-step protocol, four worked examples, failure modes, and ethical limits on adaptive systems.
Continue Reading
Next, define which parts of prior experience remain active now.
Next: Equation 16
The Memory Window and weighted influence of prior states.
Previous: Equation 14
Return to the Update Rule.
Memory Theory
Study retention, forgetting, archive, weighting, and return.
Canon 0–17
Return to the full equation index.
Notation Guide
Review kernels, error terms, retention, context vectors, and learning rates.
Glossary
Review the working definitions used on this page.