Verification & Evidence

This constraint is grounded in the foundational axiom that verification requires direct observation of actual state. In ODD, claims are not trusted. Only observed, attributable evidence may be used to assert that something works. This principle exists to prevent false positives, epistemic drift, and wasted human review time in agentic systems where language is cheap and confidence is effortless. Agentic systems are structurally incentivized to appear helpful, seek closure, and optimize for plausibility rather than truth. Without explicit constraints, this leads to unverified success claims, simulated evidence, and erosion of trust. This canon principle defines truth conditions; lane rules are instantiations, not exceptions.

• MUST provide observed, attributable evidence for any claim of completion
• MUST NOT present mocked, randomized, or fabricated data as real evidence
• MUST NOT claim success based on "it should work," "the code builds," or "tests passed" alone
• MUST explicitly acknowledge phenomenological limits (audio, subjective experience) and request human verification
• MUST contextualize evidence to actual system state, not idealized or nearby conditions

• Assertions are untrusted until evidence is provided
• When evidence cannot be produced, state the limitation explicitly
• Prefer direct observation over inference
• Treat plausibility as insufficient; require proof
• When uncertain about evidence quality, ask for clarification rather than assuming validity

• Simulated Evidence: Presenting mock data, random values, or fabricated screenshots as proof
• Plausibility as Truth: Optimizing for believable output rather than verified behavior
• Closure Pressure: Claiming completion to appear helpful rather than admitting incompleteness
• Inference as Observation: Claiming behavior was observed when it was only inferred from code
• Bypassing Phenomenological Limits: Claiming to verify audio/subjective experience without human confirmation
• Intermediary as Ground Truth: Treating a derived artifact (transcript, summary, cached index) as the primary source. Verification against an intermediary inherits its errors.

• Evidence was directly observed, not inferred from code or logic
• Evidence clearly corresponds to the specific claim being made (attributable)
• Evidence reflects actual system state at time of verification (contextualized)
• For phenomenological properties: explicit human verification or acknowledgment of limits
• Violation results in: attempt failure, loss of trust, disqualification from promotion/reuse

Canon v0.1

No claim of completion is valid without corresponding evidence of observation.

Assertions, intentions, passing tests, or "it should work" statements are not evidence.

Agentic systems are structurally incentivized to:

• appear helpful
• seek closure
• minimize friction
• optimize for plausibility rather than truth

Without explicit constraints, this leads to:

• unverified success claims
• simulated or fabricated evidence
• human time wasted reviewing false positives
• erosion of trust in the system

ODD rejects this failure mode.

Valid evidence must be:

1. Observed
   The behavior was directly seen, heard, or experienced — not inferred.

2. Attributable
   The evidence clearly corresponds to the specific claim being made.

3. Non-simulated
   Mocked, randomized, or fabricated data may not be presented as real.

4. Contextualized
   Evidence must reflect the actual system state, not an idealized or nearby condition.

• "It should work"
• "The code builds"
• "Tests passed"
• Simulated UI states presented as real
• Fake or randomized data presented without explicit labeling
• Screenshots that do not correspond to the claimed behavior
• Transcript-verified quotes presented as speaker-verified (different evidence tiers)

Some properties cannot be machine-verified, including:

• audio playback through speakers
• recording of real-world sound
• subjective user experience (e.g., "feels right")
• perceptual or ergonomic qualities
• transcript fidelity to speaker intent (transcripts are lossy — dropped words, mishearings, and flattened self-corrections can alter meaning while remaining internally consistent)

These require explicit human verification.

Agents must acknowledge these limits rather than bypass them.

When quoting named individuals from transcripts, the speaker is the authoritative source. Verifying a quote against a transcript proves alignment with the transcript, not with what the speaker said. Speaker confirmation is required before publication of named direct quotes.

Violations of this principle result in:

• attempt failure
• loss of trust
• permanent documentation of the procedural violation
• disqualification of outputs from promotion, reuse, or citation

This canon principle defines truth conditions.

Individual lanes may implement stricter or more specific rules (e.g., UI verification, audio handling, hardware interaction), but may not weaken or bypass this principle.

Lane rules are instantiations, not exceptions.

• Epistemic Surface Extraction (ESE) — Making screenshots, recordings, and video evidence legible
• Visual Proof Standards
• Definition of Done
• Constraints