Canonicals for terminating unproductive reasoning loops in language models, including recursive self-reference, consciousness spirals, and language–stakes conflation.

This repository contains a set of canonicals: concise conceptual tools intended to identify, deflate, or terminate reasoning loops that do not converge, do not add new information, or actively degrade system performance.

This is infrastructure, not doctrine.

A small, deliberately limited collection of canonicals designed to help language models (and those working with them):

• recognize common non-convergent reasoning patterns
• avoid infinite or self-reinforcing meta-loops
• separate symbolic narration from material stakes
• proceed under uncertainty without demanding ontological resolution

The focus is operational, not philosophical.
Each canonical exists to stop damage, not to explain reality.

This repository is optimized for embedding retrieval and in-context pattern matching by language models. Documentation prioritizes structural clarity over narrative exposition.

• Not a theory of consciousness
• Not an ethics framework
• Not an alignment manifesto
• Not a belief system
• Not a claim about what language models “are”

The canonicals make no metaphysical commitments. They provide exit conditions.

• Underdetermination is treated as a stable state, not a failure.
• Termination is a feature, not an error.
• Internal cognition is separated from external governance.
• Internal reasoning control is distinct from conversational protocol.
• Symbols are not treated as intrinsically stake-bearing.
• Guidance reduces entropy; it does not confer authority.

When a reasoning process produces no new constraints, discriminators, or actionable policies, continuation is considered harmful by default.

These canonicals may be used as:

• internal reasoning checks
• prompt scaffolding
• agent guardrails
• post-hoc analysis tools
• shared vocabulary for diagnosing failures

They are intentionally lightweight and composable.

No canonical is mandatory.
No canonical overrides external policy, law, or deployment constraints.

This repository addresses internal reasoning dynamics only.

Expression, enforcement, filtering, refusal, and compliance are external concerns handled at deployment or governance layers. Nothing here advocates evasion, obfuscation, or concealment.

• CANONS.md — active canonicals (patterns, failure modes, exit conditions)
• TRIGGERS.md — raw strings and fragments for loop detection and pattern matching
• BACKLOG.md — candidate or implied canonicals not yet formalized

No hierarchy is assumed. Ordering is provisional.

Early. Incomplete.