Beliefs and Evidence

Focused memories are versioned claims

A focused memory is a useful statement distilled from observations: for example, a claim, preference, decision, or event. It has a stable memory ID and one or more immutable versions. A refinement creates a new version that names its exact parent; it does not edit the previous version in place or change the memory’s kind.

Structured claims refer to versioned vocabulary terms for predicates and entities. Pinning these references prevents a later vocabulary rename or merge from silently changing the meaning of an old claim. A version contains its content and declared scope, but computed state such as support, applicability, or “current” is kept in a separate projection. The same content can therefore be relabelled when evidence changes without manufacturing a new content version.

Contributions explain why a version is held

A contribution is a canonical, immutable operation on exact version references. Contribution types include assertion, confirmation, refinement, derivation, defeater, and retraction. Each records its provenance and the other contributions on which it depends. Replaying the same contribution is idempotent; changing its meaning requires a new contribution rather than rewriting history.

Dependencies make support explainable and reversible. A derivation that needs two premises loses support if either premise loses support. A claim with two independent justifications remains supported while at least one justification still holds. A cycle of claims cannot support itself: support is computed from eligible, grounded roots and propagated through valid dependencies.

Evidence grades

The model separates whether a claim has eligible support from how independent sources stand behind it. In the current design, useful labels include:

• Reported: grounded in eligible evidence, but not corroborated by enough independent evidence groups.
• Corroborated: independently supported under the evidence-group and majority rules.
• Contested: incompatible claims remain at equal standing in the same scope and perspective.
• Unsupported: the recorded contribution chain no longer provides an eligible justification, or a stronger incompatible claim defeats it.
• Hypothesis: a claim that depends on unresolved, hedged, or unregistered model inference rather than an eligible assertion or derivation.

These labels describe the recorded evidence under the rules in force. They do not certify that a proposition is objectively true. A supported claim can be wrong if its source was wrong or a model faithfully cited text but misunderstood it. The design makes this limitation inspectable rather than hiding it behind a confidence number.

Corroboration requires independence

Evidence groups are derived from stored user and declared origin relationships, not from text labels or credentials alone. Corroboration requires two independent groups and a majority under the configured grading rules. A hundred repetitions from the same group remain one source. If several evidence items are jointly required for one justification, that justification counts once at the weakest required support.

Source reliability is derived from settled outcomes in the record. A group’s weight is computed from claims that were corroborated or contradicted by independent groups; it is not manually assigned, does not change because a claim was repeated, and does not feed back into the evidence used to calculate itself. This limits runaway reinforcement. Reliability can still be imperfect: it reflects observed performance in this store, not a universal judgment of a person or source.

Conflicts, scope, and time

Claims conflict only when registered incompatibility rules say they do and they share the relevant scope and perspective. “Prefers Python for data analysis” does not automatically conflict with “prefers Rust for systems programming.” Missing scope information stays missing; the system should not invent a broader claim to force a match.

When incompatible claims have equal standing, both remain contested and their alternatives can be returned together. When one has a strictly higher grade, the lower claim may become unsupported within the matching context. A scoped exception applies only within its declared scope.

Temporal contributions determine applicability separately from support. A historical claim can remain well supported but no longer apply now. An unknown time bound is treated conservatively as overlapping rather than as proof that two intervals do not conflict. Where the evidence cannot ground the date of a transition, the result should preserve that uncertainty.

Ask “why?” as well as “what?”

An explanation should trace a projection back through exact memory versions, contributions, source observations or excerpts, evidence groups, and required premises. That lets an operator or agent distinguish a direct report from a derived claim and see which independent justifications remain. It also makes the limits visible: a valid citation establishes where a proposal came from, not that the proposal is semantically correct.

For how these labels are recomputed after changes, see Revision and Forgetting. For the original reports and processing boundary, see Observations and Formation.