rekursiv-ai/trackinizer/belief

Belief — authoring expectations

A Belief is a proposition whose support comes from cited artifacts. Write it so the claim is falsifiable and its evidence is legible.

Per-kind expectations doc (SoT). Fields owned by types/inquiries.py; grammar by ../SKILL.md.

Completeness bar

• title = the proposition as a single, checkable sentence ("TRM's puzzle-ID

embedding is load-bearing memorization"), not a topic.

• judgement ∈ {proven, disproven, unproven, undecidable} — the coarse verdict;

set deliberately, unproven is the normal state for an open musing.

• confidence ∈ [0,1] — calibrated probability the proposition is true; 0.5

is the neutral prior. Orthogonal to judgement and status.

• citations — at least the load-bearing evidence, as signed edges (below).
• description — the argument/context when the title alone underspecifies.

Citations: the epistemic edge (NOT paper→paper)

Citations store Artifact → {Belief, Experiment}. The evidence is the citer; the Belief is the target. So:

• trax paper Q proves belief P — Q is load-bearing evidence for P (votes in

the proof predicate).

• trax paper Q favors belief P — Q is context (informs, does not vote).
• For-vs-against is the sign of valence ([-1,1], default 0.5), not a separate

kind. disproves/disfavors are aliases that negate valence.

Do NOT model a paper citing a paper as a Belief citation — that is cites_paper (see ../paper/SKILL.md). A Belief is a claim; a Paper is evidence. "Two papers agree" is expressed as both proves/favors the same Belief, not one citing the other.

Expectations

• One proposition per Belief; split compound claims.
• Every consequential Belief carries ≥1 signed citation; a bare Belief with no

evidence is a TODO, not a finding.

• judgement/confidence are set by the author from the evidence — nothing sets them

automatically; update them as citations accrue.

• Status (invalid = retracted framing) is orthogonal: a proven Belief can be

superseded by a sharper one; an invalid Belief can still read proven.

Common mistakes

• A topic as the title instead of a falsifiable sentence.
• Citing a Paper FROM a Belief in the wrong direction (evidence points UP at the

claim; the stored edge is Artifact→Belief).

• Using a separate edge kind for "against" instead of negative valence.
• Leaving judgement/confidence at defaults after evidence has accrued.
• A Belief with no citations presented as settled.