Principle 6: Failure Determines Where the System Evolves

A failure is useful evidence, not merely a reason to retry. Successful outcomes can also reveal learning about what should remain stable, be reinforced, or evolve.

Purpose

When execution reveals a weakness, uncertainty, successful pattern, or other learning opportunity, the Infoconex AI Flywheel should determine whether change is justified and where resulting learning belongs before changing the system.

Different outcomes reveal different kinds of learning. Treating every failure as a prompt problem, code problem, memory problem, or retry problem creates brittle systems and repeated mistakes. Likewise, forcing every successful execution to produce a change creates unnecessary churn.

Requirements

A conforming implementation must satisfy these requirements:

The system evaluates and classifies meaningful weaknesses, uncertainty, successful patterns, and other learning opportunities before deciding how they should affect future operation.
Classification determines whether adaptation is justified, whether evidence reinforces an existing validated operating pattern, or whether no new persistent learning is justified.
Different kinds of learning can be sent to different destinations rather than being forced into one fixed mechanism.
Supported destinations include, as appropriate, deterministic capability, procedural guidance, reasoning knowledge, validation, and governance.
The system can identify when responsibility is in the wrong place and move it across the Moving Determinism Boundary.
When adaptation is justified, the selected destination reflects the mechanism best suited to handle the correction reliably without removing flexibility that is still needed.
A failed or rejected adaptation remains distinct from any reusable lesson derived from that attempt.
Candidate changes and other learning intended for persistent future use are sufficiently validated before being trusted for that use.
Governance changes remain subject to human authorization.

Operational Model

Classification Can Produce Different Dispositions

Classification may determine that:

Adaptation is justified and resulting learning should change a persistent operational asset.
No adaptation is justified and the outcome reinforces an existing validated operating pattern.
No new persistent learning is justified even though the outcome and classification remain part of the evidence record.

A failure may also produce learning that is different from the failed candidate itself. For example, a rejected change may reveal a known-failure rule, applicability constraint, detection mechanism, validation check, procedural instruction, or alternative-strategy rule.

Possible Adaptation Destinations

Deterministic Capability — Use when the problem is repeatable and can be corrected through code, tools, scripts, services, or another deterministic implementation.

Procedural Guidance — Use when the process is incomplete, unclear, incorrectly ordered, or missing known exception handling.

Reasoning Knowledge — Use when the condition still requires judgment but would benefit from durable guidance, examples, heuristics, or lessons.

Validation — Use when the system lacks a reliable way to determine whether an action or outcome is correct.

Governance — Use when evidence shows that authority, approval requirements, prohibited actions, or escalation rules should change. Governance changes require human authorization.

The Evolution Routing Rule

When a correction is required:

Correct the weakness at the lowest layer capable of handling it reliably without unnecessarily removing adaptability.

This prevents the system from solving every recurring problem with more AI reasoning when a more reliable mechanism exists. It also prevents the system from hard-coding behavior that still requires judgment.

Routing Can Move Responsibility

Examples include:

AI reasoning → SOP when a recurring judgment becomes a known procedure
SOP → deterministic capability when a stable procedure becomes reliably automatable
Deterministic capability → SOP or AI reasoning when evidence shows that a hard-coded assumption is too brittle

Lifecycle Relationship

Observe and Evaluate: Evidence establishes what happened and what learning the outcome may contain, including successful reinforcing evidence.
Classify: The system determines whether adaptation is justified and where resulting learning should live.
Adapt: A candidate correction is created when change is justified, or the no-adaptation disposition is explicitly resolved.
Validate: Candidate improvements or other reusable learning intended for persistence are evaluated for sufficient support.
Persist and Reuse: Supported learning becomes part of future operation, reinforcing evidence may strengthen an existing validated pattern, or the lifecycle explicitly resolves that no new persistent learning is required.
Later cycles: Failure-derived learning and new evidence can challenge earlier classifications or persisted learning and route the system through the lifecycle again.

Evidence of Implementation

Evidence supporting this principle may include:

Classified failures, uncertainty, successful patterns, or learning opportunities with identified causes or meaning
Explicit decisions about whether adaptation is justified
Examples in which different lessons produced different destinations
Records showing why a correction was placed in code, procedure, reasoning guidance, validation, or governance
No-change classifications that reinforce an existing validated pattern without manufacturing an adaptation
Failure-derived learning that remains distinguishable from the failed or rejected candidate
Evidence that responsibility can move across the Moving Determinism Boundary when justified
Validation results for candidate changes or other reusable learning before persistence or reuse

Non-Conforming Patterns

This principle is not satisfied when:

Every failure produces the same response
The system only retries without diagnosis
Every successful outcome is forced to create a change
Every lesson is stored only as memory
Every failure is treated only as a code defect
Every weakness becomes additional prompt text
Learning is persisted without determining what part of the system should own it
The system has no intentional way to move responsibility when the current placement is wrong
A failed candidate is promoted as approved behavior instead of separating any validated lesson derived from it

Relationships to Other Principles

Principle 3: Work Is Distributed Across a Moving Determinism Boundary defines where responsibility can live; this principle determines whether and where specific learning should be applied.
Principle 5: Execution Must Produce Outcome Evidence provides the evidence required for responsible classification and routing.
Principle 7: Learning Must Change a Persistent Operational Asset ensures reusable supported learning becomes durable when persistence is justified.
Principle 1: Autonomy Is Bounded by Human Authority governs changes that affect authority or protected operational assets.
Principle 8: Improvement Must Compound Through Reuse depends on relevant current learning and validated operating patterns being inherited by future execution.

Principles Index
Previous Principle: Principle 5: Execution Must Produce Outcome Evidence
Next Principle: Principle 7: Learning Must Change a Persistent Operational Asset