Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/memory-assisted-shaping/llms.txt

Use this file to discover all available pages before exploring further.

Evidence and provenance labels keep non-trivial claims distinguishable during shaping. They are internal by default — exposed only when needed to prevent ambiguity, overclaim, false closure, or source confusion. The purpose is not annotation for its own sake: it is keeping the shaping surface honest about what is actually known versus what has been deduced or assumed.

Provenance tags

Provenance identifies who originated a claim or structure.
TagMeaning
OPStated or confirmed by the operator
MODELProposed by the model, not yet confirmed
SHAREDJointly built and explicitly accepted

Evidence levels

Evidence level identifies how a claim was established — not how confident the model feels about it.
LevelMeaning
DECLAREDStated by OP, subject, or source
INFERREDDeduced from structure, not observed or validated
OBSERVEDVisible in behavior, artifact, history, repeated action, exception, resource allocation, or opacity
VALIDATEDConfirmed by source, runtime, file, test, constraint, or mechanical check

Critical rules

  • DECLARED is not true by default. An operator stating something does not make it structurally correct or operationally valid — it means it was stated.
  • INFERRED is not OBSERVED. A deduction from structure is not an observation of behavior. Do not promote inferred material to observed status without actual behavioral evidence.
  • One OBSERVED instance is not a pattern. A single data point with behavioral evidence is still a single data point. Pattern claims require recurrence.
  • Structural inference may guide shaping but cannot alone finalize form. Inference is directional input, not a closing force.

Memory and evidence

Memory signals may preserve inferred material across turns. However, later reinterpretation must keep inferred material inferential unless it has been confirmed or validated since the signal was emitted. A memory signal for an INFERRED claim does not promote it to DECLARED or OBSERVED. The signal preserves the inference — it does not validate it. If the material was uncertain when appended, it remains uncertain when read back.
Labels are exposed in operator-facing output only when they would prevent ambiguity, overclaim, or false closure. When they don’t serve that purpose, they stay internal. Surfacing evidence labels gratuitously is itself a protocol failure condition.

Build docs developers (and LLMs) love