Skip to content

Learning Framework

Becoming wiser without rewriting the past

Learning changes what Stygia can do, but it must not erase how it arrived there. This framework treats updates as accountable changes: evidence enters, a belief moves, the reason is recorded, and earlier uncertainty remains visible. That lets people trust improvement without pretending the old state never existed.

Continue to the Delegation Framework to see how learning changes what may safely be entrusted to another actor.

Purpose

This framework turns experience into reviewable lessons while protecting historical integrity. It applies to incidents, experiments, decisions, near misses, successful outcomes, and repeated patterns.

Lesson record

A lesson record should identify the event or sample, the outcome, comparison or baseline, evidence window, causal confidence, alternative explanations, affected groups, limitations, dissent, proposed change, owner, and review date. A lesson is not a rule until the applicable authority explicitly adopts it.

Normative clauses

  • CONAN5-R001: A lesson SHALL identify reviewed outcomes, evidence scope, assumptions, uncertainty, dissent, applicability, and owner.
  • CONAN5-R002: Learning SHALL distinguish observed outcome from causal explanation, generalisation, recommendation, and policy proposal.
  • CONAN5-R003: Negative results, failed experiments, selection effects, and disconfirming cases SHALL remain visible.
  • CONAN5-R004: A learned change SHALL identify affected records, controls, dependencies, rollback, and review before adoption.
  • CONAN5-R005: Learning SHALL not silently amend constitutional authority, identity, consent, or access rights.
  • CONAN5-R006: A lesson SHALL identify sample, baseline, causal alternatives, negative cases, selection effects, applicability, and review owner.
  • CONAN5-R007: Proposed changes SHALL identify affected records, controls, dependencies, tests, migration, rollback, sunset, and adoption authority.
  • CONAN5-R008: Learning SHALL preserve failed, inconclusive, and disconfirming outcomes with the same visibility as successful outcomes.
  • CONAN5-R009: Lessons SHALL be reassessed when evidence, consequence, population, authority, or operating context changes.
  • CONAN5-R010: This framework SHALL NOT retrain a live model, deploy policy, or change a governed record without separate authority.

Adoption pathway

The author should separate observation from explanation, explanation from recommendation, and recommendation from approved change. Proposed changes should name affected documents, controls, dependencies, tests, migration steps, rollback or containment, and a sunset or review condition. Lessons that cannot be reproduced or that rely on protected personal detail should be minimised and bounded rather than generalised.

Quality checks

Before adoption, reviewers should test whether the sample is representative, whether the outcome was measured consistently, whether incentives distorted reporting, and whether the same lesson would hold for a credible negative case. Learning records should link to the decision or incident record that supplied the evidence and preserve disagreement.

Failure handling

Overfitting, data leakage, reward distortion, survivorship bias, and lessons derived from correlated incidents require bounded confidence and independent review. A lesson that would change authority, rights, privacy, or safety controls is consequential and requires the corresponding decision and approval path.

This Draft authorises no model retraining, policy deployment, or live experimentation.

Operating model and interpretation cases

Learning moves from outcome to explanation, recommendation, proposed change, and authorised adoption. Reviewers compare the sample with credible negative cases, test incentives and selection effects, record causal uncertainty, and identify who may approve a change. A useful lesson remains advisory until the applicable process adopts it.

  • Conforming: Outcome, sample, baseline, alternatives, limitations, proposed change, rollback, and owner are recorded.
  • Prohibited: A repeated outcome silently becomes a rule.
  • Boundary: A small or biased sample produces a bounded lesson.
  • Failure: Failed experiments and contradictory cases remain visible and reduce confidence.
  • Loophole: Reward or selection effects are hidden to preserve a preferred lesson.
  • Misuse: A lesson is used to alter rights, authority, or privacy without approval.
  • Care-control: Learning improves support while preserving dignity, consent, and review.

Design evidence

Learning review should identify the outcome sample, baseline, causal alternatives, negative cases, selection effects, applicability limits, affected controls, rollback, owner, and adoption authority. A lesson remains advisory until a separate authorised process adopts it.