TensionLM-117M-TS-Reasoner-v8

This is the explainable-boundary CPU TS reasoner for the frozen TensionLM-117M-Reasoning-v2 substrate.

v8 keeps the v6/v7 solver path and adds richer abstention decisions:

• explicit refusal rules such as graph:cycle_detected, graph:ambiguous_branch, arithmetic:division_by_zero, and code:unknown_function,
• low-confidence bands for unsupported or partial prompts,
• a mixed TAC-GEN distribution that interleaves solvable standard/paraphrase prompts with unsolvable unknown prompts.

The point is still no-GPU reasoning: frozen language substrate plus explicit TS graph/program operators, boundary detection, and inspectable confidence.

Eval receipts

Fixed benchmark scores:

System	TAC v2	TAC v3	TAC v4
TS-Reasoner-v8	120/120	120/120	120/120

Generated benchmark scores:

Distribution	Engine	Score	Solve rate
TAC-GEN paraphrase, seed 9101	v8	3000/3000	100%
TAC-GEN unknown, seeds 9201-9204	v8	12000/12000	0%
TAC-GEN mixed, seeds 9301-9304	v8	12000/12000	65.7%

For the unknown distribution, correctness means returning <ABSTAIN>. In the mixed distribution, standard/paraphrase prompts are solved and unknown prompts are refused in the same pressure field.

These are system scores, not raw LLM scores.

Usage

python inference.py --prompt "Handoff log: a hands off to b; b hands off to c. Ignore the separate handoff x hands off to b. The handoff chain beginning at a ends at" --category transitivity --show_trace

python inference.py --prompt "Graph ledger: main(a,b); main(a,c). Resolve main* from a; terminal node:" --category transitivity --show_trace

python demo_ts_reasoner_v8.py

Limitations

This artifact handles generated formal prompt families covered by the included operators. It is not a chat assistant, not raw model improvement, and not a claim of open-ended natural language understanding. The confidence values are rule-calibrated system signals over these families, not probabilities over all natural language.