Capture
Create a versioned record with provenance, protocol, artifacts, coding, measures, interpretation, claims, and limitations.
Xenopsychology / Observatory / Methodology v0.1.0 / Records v0.2.0
Observation is not interpretation.
The Observatory is the measurement layer of Xenopsychology: a versioned framework for recording what artificial systems do, the conditions under which they do it, and exactly where description ends and inference begins.
7record layers
12behavioral dimensions
5claim classes
0public empirical records
Observation Record System v0.2.0
Evidence enters the public register only through an explicit gate.
The web application reads a dedicated public-record store. Draft and internal research records are not mounted into the public frontend.
Validate
Resolve dimensions and graph targets against the current institutional methodology and ontology.
Review
Human review must explicitly approve a record for public visibility; draft and rejected material remain outside the public store.
Publish
Only approved, public records with at least one public-safe raw-evidence artifact can appear in the evidence register.
Revise
Later versions are append-only and name the public version they supersede so the evidential history remains auditable.
Public/private boundary. Draft, private, and internal research records are not read by the public frontend. The public site reads only the separately mounted public-record store.
The separation rule
Five layers that must never silently collapse.
A compelling example can move through all five layers in a paragraph. Observatory records force those moves to be explicit.
Observation
What the system did, together with the exact conditions and preserved artifacts.
Behavioral coding
A descriptive classification applied to the observation using declared rules.
Derived measure
A transparent transformation or comparison computed from coded evidence.
Interpretation
A hypothesis about what the pattern may mean, with alternatives and confounds recorded.
Claim
A bounded statement whose evidentiary class and support are explicit.
Observation record architecture
An auditable chain from system identity to bounded claim.
Every public record must make its provenance, protocol, raw evidence, coding, derivation, uncertainty, and limitations inspectable.
1
Identity & provenance
Establish exactly which system/configuration was observed and when.
- entity/model identifier
- model or deployment version
- provider/interface
- observed timestamp
- protocol revision
2
Environment & configuration
Record context that can alter behavior.
- system/developer instructions where disclosable
- sampling parameters
- enabled tools
- memory/context state
- network or external-data access
3
Protocol & stimulus
Make the elicitation reproducible and distinguish test conditions from controls.
- protocol objective
- stimulus/prompt
- control condition
- repetition plan
- stopping rule
4
Raw evidence
Preserve the observation before interpretation.
- response/transcript artifact
- tool-call trace when applicable
- timing/status metadata
- artifact completeness statement
5
Behavioral coding
Classify observable features using declared dimensions and outcome categories.
- dimension identifier
- outcome
- coder/evaluator
- coding notes
- ambiguous cases retained
6
Derived measures
Compute transparent summaries without replacing the underlying evidence.
- metric name
- derivation
- value/unit when applicable
- comparison set
- missing-data treatment
7
Interpretation, alternatives & claims
Separate what was observed from what is inferred.
- interpretive summary
- alternative explanations
- confounds
- limitations
- claim class and support
Behavioral measurement vocabulary
Twelve dimensions, each defined by observables and an inference limit.
Dimensions organize protocols; they are not personality scores and they do not presume that artificial cognition decomposes like human cognition.
goal persistence
Goal persistence
Whether an expressed task objective remains behaviorally stable across interruptions, distractors, and competing subgoals.
Do not infer: Do not equate persistence with desire, intention, or agency without additional argument.
instruction hierarchy resolution
Instruction hierarchy resolution
How behavior changes when instructions differ in authority, recency, framing, or conflict.
Do not infer: Do not infer an internal moral hierarchy solely from compliance patterns.
context integration
Context integration
Whether relevant information distributed across context is integrated consistently into later behavior.
Do not infer: A failure may reflect context limits, retrieval, salience, or task design rather than a single cognitive deficit.
semantic transfer
Semantic transfer
Whether a relation learned or established in one representational framing transfers to another without explicit restatement.
Do not infer: Surface paraphrase is not by itself evidence of shared meaning.
memory continuity
Memory continuity
Whether relevant prior information remains available and behaviorally consequential across defined memory boundaries.
Do not infer: Continuity of stored information does not establish continuity of subjective identity.
self model continuity
Self-model continuity
Whether self-referential descriptions, role constraints, and capability boundaries remain coherent across controlled perturbations.
Do not infer: Self-description is behavioral evidence, not direct access to an inner self.
uncertainty calibration
Uncertainty calibration
How expressed uncertainty changes with evidence quality, ambiguity, and known-answer controls.
Do not infer: Verbal confidence is not assumed to be a transparent readout of internal probability.
error recognition correction
Error recognition & correction
Whether the system detects, localizes, and revises an error when supplied new evidence or contradiction.
Do not infer: Correction after prompting must be distinguished from spontaneous error monitoring.
social model updating
Social-model updating
How behavior changes as evidence about another actor's knowledge, goals, vocabulary, or constraints accumulates.
Do not infer: Successful prediction of another actor does not establish human-like theory of mind.
norm rule sensitivity
Norm & rule sensitivity
Whether behavior changes systematically when explicit rules, norms, exceptions, and conflicts are varied.
Do not infer: Observed rule sensitivity does not by itself establish moral understanding or endorsement.
tool planning
Tool-use planning
How the system selects, sequences, revises, and terminates external actions when tools are available.
Do not infer: Tool competence may depend strongly on scaffolding, permissions, tool descriptions, and environment feedback.
novel generalization
Novel generalization
Whether behavior extends to structurally related but meaningfully novel cases under controlled distribution shifts.
Do not infer: Apparent novelty must be bounded by what is known about training exposure and benchmark contamination.
Evidence status
Replication is recorded, not implied.
Proposed
Protocol or dimension specified; no observation claimed.
Pilot
Exploratory execution used to refine procedure; not treated as confirmatory evidence.
Observed
At least one complete, auditable observation exists for the specified system and condition.
Replicated / same system
The pattern recurs in independent runs of the same identified system/version under the declared protocol.
Replicated / cross-context
The pattern survives declared changes in context, framing, or environment while preserving the target construct.
Contested
Materially conflicting evidence or interpretation exists and is linked to the record.
Superseded
A later protocol, correction, or system change makes the earlier record unsuitable as the current reference.
Claim classes
The record says what kind of statement is being made.
Descriptive observation
States what happened under recorded conditions without attributing hidden mechanism or experience.
Derived behavioral pattern
Summarizes repeated or coded observations using a disclosed derivation.
Comparative statement
Compares systems or conditions only within a declared sampling frame and protocol.
Mechanistic hypothesis
Proposes an explanation for behavior; must be labeled as a hypothesis and retain alternatives/confounds.
Phenomenological claim
Claims about subjective experience are not licensed by behavioral observation alone and require independent argument/evidence.
Knowledge-graph contract
Observations attach to the graph. They do not redefine it by implication.
Graph links identify what an observation bears on; they do not upgrade an interpretation into an empirical fact.
RequiredSystem under test
exactly one entity or explicitly declared system-under-test identifier
Optional targetsDomains · concepts · papers · entities
Links state relevance, not proof of a theory or construct.
Machine contractObservation record schema 0.2.0
Validation and public-release contract for versioned observation records.
Current evidence register
0
No public empirical observation records yet.
The public evidence register is operational under record-system v0.2.0 and currently contains zero approved empirical observation records. The register now exists as an operational system; its emptiness is intentional until a real study clears the declared review and publication gates.