Definition
A process believes p to the extent that it uses p as a premise when interpreting observations, generating expectations and evaluating counterfactual possibilities across the contexts p bears on, and maintains or revises that use through its epistemic learning. Truth-directed marks p’s role as a claim about how things are. It does not require the belief to be true, rational, explicit, conscious or easily revised.
Three parts carry the definition.
- Inferential commitment. Interpretation alone is not belief (CLAIM-336’s interpretive operators can entertain a frame). An interpretation is belief-like when it is maintained and reused as a premise across anticipations, inferences and revisions. Strength is how broadly and robustly p is so used, not a universal threshold of credence (THEORY-647’s context-relative thresholds are one formal option, not the definition).
- Revision profile. Two processes can agree now and diverge under evidence: CLAIM-294’s agents with P(H) = 0.9 each, and P(H | C) = 1.0 and 0.8. A belief state is the pair of present commitment and disposition to revise. That is operational, not a claim that each belief stores its own update rule.
- Relative to a process. Two processes can believe differently about one proposition without the organization being incoherent. The organization’s effective belief is the commitment maintained through its governance and legibility, which need not be any aggregate of the local credences (THEORY-127, THEORY-536). Private belief, organizational belief and official assertion can diverge.
Probes. Anticipation (forecasts of p’s observable consequences under independent scoring, not inference from approach or avoidance); inferential generalization (does the process use p to predict consequences it was not trained on?); epistemic intervention (supply diagnostic evidence or remove an assumption, and see what changes, keeping access apart from uptake). Then vary incentives: a report that moves with reward while independent predictions do not is a reporting or governance effect. The diagnostic is fallible: finite behaviour underdetermines belief.
What it is not
- Wanting (CLAIM-447): what draws pursuit. Liking: how an experienced outcome is evaluated. Feeling (CLAIM-439): an affective evaluation made available to the system’s further processing. Authority: which evaluations become consequential. An agent can believe an activity safe, want to avoid it, enjoy it when it does it, and be governed by its defensive processes.
- CASE-124’s m, a scalar lean in one interpretive dimension, which summarizes the content of the prevailing interpretation and is not the whole belief state.
- Dennett’s intentional stance (THEORY-538), which earns belief attribution from predictive success across contexts. The candidate adds a process-level hypothesis about what organization sustains the pattern.
- Knowledge. The exchange went on to ask whether belief and knowledge can be told apart in the model, and the owner set the distinction aside (U26–U27). Nothing here defines knowledge.
What would settle it
Adoption as stated, a narrower definition, or a rival. It would be adopted if claims about belief in the will and legibility lines can be stated in it, and the three probes distinguish cases the record needs (an accurate belief denied authority, QUESTION-270; a false belief kept by a legibility regime, CASE-156).