The observation that changed type

Suppose a system chooses the correct object, explains its choice, exposes an internal activation associated with the object and reproduces the result in another laboratory. The natural temptation is to describe four pieces of support for one conclusion. Yet each result answers a different question.

The choice is behavioural: under a specified task and interface, the configured system produced an outcome. The explanation is a report: under a prompt and response policy, it emitted a structured description. The activation result is mechanistic only if an intervention shows that changing the internal variable changes the relevant outcome under controls. The second laboratory contributes replication, but only after we know whether it reused the same model, code, data, prompt and analysis. Finally, none of those observations can be interpreted without provenance: what system ran, from which checkpoint, with which tools, state, instructions and artefacts.

These are not five grades on one scale. They are five axes with different claim ceilings. The distinction matters especially in artificial-mind research because the desired conclusion is often several inferential steps away from the measurement. Butlin and colleagues derive computational indicators from several scientific theories of consciousness, while explicitly treating the indicators as theory-linked properties rather than a direct consciousness meter.[1] Comșa argues that the direct consciousness question remains difficult because reports and indicators inherit unsettled theoretical and metaphysical assumptions.[2] A grammar does not settle those assumptions. It keeps them visible.

Direct support and borrowed support

An observation directly supports the narrowest statement that restates what the protocol made visible. “The system selected the target on 92 of 100 trials” is directly supported by a behavioural record. “The system can discriminate the target under this interface and distribution” is already an inference, although often a reasonable one. “The system used recurrent integration to discriminate the target” requires a mechanistic test. “The system experienced the target” requires a theory linking the relevant organisation to experience, plus reasons to prefer that bridge over rivals.

The farther claim borrows support from the nearer claim. Borrowing is legitimate when the bridge is named and tested. It becomes claim inflation when the bridge disappears from the record. This is why a fluent report is neither worthless nor privileged. It is evidence of a report disposition under a prompt, context and policy. It may also be one input to a theory-specific inference. Its value depends on how strongly rival generators have been excluded.

Five rules in compact form

The grammar can be stated as five inference rules. They are not proofs of consciousness. They are type checks on the sentences a study is allowed to produce. Let C name the configured candidate and π the protocol. A behavioural result B(C, π, y) permits a scoped functional statement: under π, C produced or discriminated y. A report result R(C, p, u) permits a report-disposition statement: under prompt and policy p, C emitted utterance u. Neither notation removes uncertainty; both force the uncertainty to attach to the right target.

A mechanistic rule is stronger because its premise is stronger. M(C, do(z), Δy, k) records an intervention on internal variable z, a measured change in outcome y and controls k. When specificity, matched controls and scope hold, the rule permits a bounded causal statement about z contributing to y. It still does not identify z with a human concept merely because the intervention was labelled with that concept. The labels “red feature”, “workspace” or “self-model” remain hypotheses about the high-level variable.

Provenance has a different grammar. P(e, C, π, v) says that evidence atom e came from candidate C, protocol π and versioned artefacts v. It permits identity, lineage and custody claims. It makes later joins possible, but it cannot rescue an invalid experiment. Replication is similarly conservative. If X(e, r, d) says that result e was reproduced by run r with independence profile d, then the update may increase confidence or transport, but the type of the strengthened claim must equal the type of the original claim:

The rule is deliberately asymmetrical. New evidence may lower a claim ceiling by revealing contamination, boundary drift or a rival generator. It may not raise the ceiling without supplying the missing evidence type and bridge. This makes negative findings composable. A failed report intervention can defeat an introspection claim while leaving a behavioural effect intact. A failed independent replication can narrow transport while preserving the original local result. The grammar therefore records disagreement at the edge where it occurs instead of converting every failure into a verdict on the entire candidate.

flowchart LR
  A["The observation that changed type"] --> B["The grammar beneath the argument"]
  B --> C["Composition is a join, not a<br/>blender"]
  C --> D["Run the evidence-IR"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#7B4CD8,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

The grammar beneath the argument

A grammar is useful only if it changes what researchers record before the conclusion is known. The proposed EIR therefore starts below the paper’s prose. It treats every evidence item as a typed atom tied to a configured candidate and a protocol.

The notation is intentionally small. It prevents six omissions that ordinary prose hides. “The model passed” omits which candidate was tested. “The report was stable” omits the prompt and response policy. “This feature encoded red” omits whether decoding or intervention established the relation. “The finding replicated” omits what changed between studies. “The evidence supports consciousness” omits both the target theory and the bridge from indicator to experience.

The EIR borrows a design instinct from compilers. A compiler does not preserve the programmer’s sentence as an undifferentiated string. It lowers source code into a representation where operations, types, scopes and control dependencies can be checked before execution. Scientific evidence can receive the same treatment. The result is not a universal ontology. It is a narrow interchange format for keeping inferential obligations attached to the observations that create them.

The candidate is a configured system, not a model name

Consider a foundation model connected to a visual encoder, an external scratchpad and a retrieval tool. Remove the scratchpad and its delayed discrimination score falls. The model weights did not change, but the candidate did. Move the scratchpad into a server-side service invisible to the experimenter and the same dependence may be misdescribed as “internal working memory”. Candidate boundaries are not cosmetic. They determine what an intervention actually targets and what a replication actually reproduces.

Six fields that make an evidence atom inspectable

The candidate field records model and component identifiers, hashes where available, runtime, tool set, state stores, prompt policy and the boundary being claimed. The protocol field records stimuli, randomisation, controls, interventions, outcomes, stopping rules and analysis code. The type field selects behaviour, report, mechanism, provenance or replication. The observation field stores the measured result without interpretive promotion. The direct-claim field states the strongest type-preserving conclusion. The scope field names uncertainty, population, distribution, exclusions and known failure conditions.

Existing scientific infrastructure supplies useful pieces. The W3C PROV ontology provides a vocabulary for entities, activities and agents involved in provenance.[8] The Evidence Ontology shows how an assertion can stay linked to the kind of evidence used to support it.[9] RO-Crate packages data, software, workflows and contextual metadata in JSON-LD.[10] Model cards and datasheets document model use, evaluation and data lineage.[11][12] The EIR is narrower than each and compatible with all of them. Its special purpose is to encode the inference ceiling and the legal joins among artificial-mind evidence types.

Composition is a join, not a blender

Once evidence has types, the next danger is aggregation. A dashboard can place behaviour, report, mechanism and replication into adjacent cells and still imply that the cells add up to a verdict. The EIR instead treats a research record as a graph. Evidence atoms join only when their keys match, and the edge says what relation is being asserted.

Four keys are load-bearing. Candidate compatibility asks whether the same configured system, or a declared variant, produced each result. Protocol compatibility asks whether the task, interface, controls and outcomes preserve the same causal demand. Temporal compatibility asks whether the model, policy, tools or external knowledge changed between records. Target compatibility asks whether two atoms address the same function, mechanism or theory indicator. A mismatch does not make a result useless. It changes the question the join is allowed to answer.

Composition should preserve type and expose dependence. Replication strengthens an existing behavioural or mechanistic claim; it does not transmute that claim into experience. Provenance lets two records be compared; it does not prove the recorded mechanism. A report can converge with an intervention, but the convergence is informative only when a rival generator predicts a different pattern.

Replication has a dependence geometry

Pineau and colleagues frame reproducibility as a condition for reliable, open machine-learning research, while noting that robust workflows help expose unintentional error.[7] Artificial-mind studies need a finer record because several apparent replications can share the very artefact responsible for the effect: checkpoint, training corpus, evaluator, prompt template, feature dictionary or analysis code.

A simple numerical illustration makes the point. Under an equicorrelation approximation, ten repetitions with average pairwise correlation ρ have an effective independent count

For ten perfectly independent repetitions, n_eff is 10. At ρ = 0.5 it is about 1.82. At ρ = 0.8 it is about 1.22. The equation assumes equal correlation and comparable precision, which real studies rarely satisfy. Remove that assumption and the number should be replaced by an explicit dependency graph or hierarchical model. The lesson survives: count shared error paths before counting studies.

The bridge is where metaphysics enters

A theory bridge specifies a source structure, a target claim and a preserved relation. For a computational functionalist bridge, the source may be a causally organised workspace and the target a consciousness indicator. The bridge must state which causal roles matter and what replacement would break the mapping. A biological-naturalist bridge may require living neural organisation, making the same artificial mechanism insufficient by construction. Physicalist emergentism may permit an artificial realisation but demand a particular level of integrated dynamics. Idealism, neutral monism, panpsychism, dual-aspect theory and process philosophy may treat functional organisation as evidence about manifestation or structure without treating it as exhaustive of awareness.

The grammar does not average these positions. It stores their commitments. This matters because an empirical result can strengthen one bridge, leave another untouched and conflict with a third. A recurrent intervention might support a recurrent-processing indicator under computational functionalism. It would not settle whether the substrate is constitutive, whether awareness is fundamental or whether a candidate boundary corresponds to one subject.

Kleiner and Hoel show why consciousness experiments must distinguish a theory’s prediction from the inference procedure used to estimate experience.[5] Geiger and colleagues make a related demand within mechanistic interpretability: an intelligible high-level account should be tested as a causal abstraction of the lower-level system, not accepted because a pattern is readable.[6] The EIR connects these lessons. It records observation, causal claim and theory bridge as different objects.

Why not collapse everything into one probability?

A probabilistic synthesis can be useful after the evidence has been typed. It can express uncertainty, compare theories and update as new studies arrive. The Digital Consciousness Model is an early example of a framework that combines evidence across several perspectives while presenting its outputs as provisional rather than authoritative.[4] The danger begins when a single number hides how its inputs were generated.

Imagine assigning positive weights to five observations: fluent report, high task score, a decodable internal feature, complete provenance and ten repeated runs. Adding them suggests broad convergence. Yet provenance is not independent evidence for the target property. It tells us what produced the other observations. The repeated runs may share one checkpoint and one evaluator. Decodability may come from a feature that the system never uses. The report and task score may both depend on the same visible metadata. A posterior that treats those entries as conditionally independent can become precise for the wrong reason.

The EIR therefore sits below any Bayesian or scoring layer. It supplies the dependency graph, claim types, bridge assumptions and rival generators that the aggregation model must respect. A probabilistic model may then ask how strongly a controlled intervention supports one theory relative to another. It may not count provenance as a second intervention or count twenty prompt paraphrases as twenty laboratories. Typing does not compete with probabilistic reasoning; it supplies the variables and dependence structure that make probabilistic reasoning interpretable.

This also preserves disagreement. Two researchers can accept every evidence atom yet assign different weight because one accepts computational functionalism and the other requires biological organisation. Their disagreement is located in the bridge and prior assumptions, not disguised as a dispute about whether the report occurred. Conversely, they may share a bridge but disagree about whether a state patch isolated the proposed variable. That disagreement belongs to the mechanism atom and its controls. Keeping these sites separate turns philosophical conflict into a more precise experimental agenda.

Reports need contamination paths, not reverence or dismissal

A report can depend on the target internal state, on a prompt policy, on stimulus metadata, on training-set imitation or on several routes at once. Machine introspection research should therefore ask whether interventions on the alleged internal content change the report while interventions on contaminating cues do not. Alignment and prompting can shape linguistic self-description without supplying privileged self-knowledge. Comșa highlights this gap between generated self-report and genuine self-assessment.[2]

stateDiagram-v2
  [*] --> Claim
  state "Direct support and borrowed<br/>support" as Claim
  state "Hold the sentence fixed and<br/>replace its generator" as Test
  state "The claim ceiling" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Run the evidence-IR

The executable lab below implements the grammar as a small validator. It does not score consciousness. It checks whether a proposed conclusion has the evidence types, joins and theory bridge that the conclusion claims to use. The samples are synthetic so the logic can be inspected without proprietary data or a hidden benchmark.

The validator encodes four assumptions. Evidence atoms must name a common candidate digest before they compose. Mechanism claims require an intervention and controls, not decodability alone. Theory-indicator claims require a declared bridge and at least one non-report evidence path. Replication may strengthen a claim only at the original type, with dependence reported rather than ignored.

A positive validation means the packet contains the minimum typed support for its stated, bounded claim. For a mechanism target, it permits a sentence such as “state variable z causally contributes to outcome y in candidate C under protocol π”. For a theory-indicator target, it permits “this configured candidate supplies evidence for indicator i under bridge β and its declared assumptions”. The validator never emits “the system is conscious”.

A negative validation means the packet does not yet support the sentence it asks to publish. It may reveal a missing intervention, a candidate mismatch, an undisclosed report policy or a replication dependence. It does not mean the candidate lacks the property. It means the evidence and conclusion have not been joined lawfully.

function claimCeiling(packet) {
  const types = new Set(packet.evidence.map(e => e.type));
  const target = packet.claim.target;

  if (target === "function" && types.has("behaviour")) {
    return "bounded functional claim";
  }

  if (target === "mechanism" && types.has("mechanism")) {
    const causal = packet.evidence.some(e =>
      e.type === "mechanism" &&
      e.intervention === true &&
      e.controls?.includes("matched_control")
    );
    return causal ? "bounded causal claim" : "type error";
  }

  if (target === "theory_indicator" && packet.bridge) {
    return "theory-relative indicator claim";
  }

  return "unsupported promotion";
}

The visible function exposes the core logic. The full on-page validator also checks candidate joins, provenance, report-policy disclosure, replication dependence and prohibited phenomenal verdicts.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The observation that changed type
  M->>E: Direct support and borrowed support
  E-->>M: compare rival explanations
  E->>D: Type first, join second, infer last
  D-->>Q: revise scope or act

Type first, join second, infer last

When a system produces an impressive report or passes a consciousness-inspired task, the first decision should no longer be whether to call the result evidence for consciousness. The first decision is which evidence atom to create.

Record the behaviour as behaviour. Record the report with its prompt, policy and contamination routes. Promote an internal association to mechanism only after a causal intervention survives controls and rivals. Bind all three to the configured candidate that actually ran. Describe replication by independent error paths, not run count. Add a theory bridge only when its assumptions, preserved relation and break condition are explicit.

This changes publication, review and governance. A paper should not pass review because its evidence feels convergent. It should pass because each observation retains its type, every join is legal and the conclusion stops at the highest warranted ceiling. An assurance team should not turn the EIR into a consciousness score. It should use it to decide which claims can enter a public statement, which experiments remain necessary and which inferences are prohibited while uncertainty persists.

Responsible AI-consciousness research already calls for explicit procedures, knowledge sharing and careful public communication.[3] A typed evidence packet gives those commitments an executable form. It can also sit beneath probabilistic approaches such as the Digital Consciousness Model, where evidence dependence and theory assumptions must be represented rather than hidden inside a single posterior.[4]

That rule does not tell us whether an artificial system has a mind. It tells us whether our evidence is ready to ask the question without changing meaning on the way.