TensionLM-117M-TS-Reasoner-v5

This is the CPU-only TS graph/program reasoner v5 for the frozen TensionLM-117M-Reasoning-v2 substrate.

v5 adds a separate interpreter engine:

• graph_parser: query-selected typed-edge graph following.
• arithmetic_parser: operation-log and word-trace arithmetic.
• safe_python: bounded Python trace/loop/branch/data-flow semantics.
• controller: tries v5 operators first, then legacy v4 fallback.

Eval receipts

All scores are system scores, not raw LLM scores.

System	TAC v2	TAC v3	TAC v4
GPT-2 124M	3/120	0/120	0/120
Base TensionLM 117M	7/120	1/120	2/120
TensionLM-117M-Reasoning-v2	20/120	2/120	0/120
TS-Reasoner-v4	120/120	120/120	0/120
TS-Reasoner-v5	120/120	120/120	120/120

TAC v4 is adversarial to v4: it uses query-selected graph chains, operation logs, and Python traces. The v4 failure ledger is included in eval/.

Usage

python inference.py --prompt "Graph ledger: main(alpha,beta); side(alpha,zeta); main(beta,gamma); main(gamma,delta). Resolve main* from alpha; terminal node:" --category transitivity --show_trace

python inference.py --prompt "Counter starts at 5. Ops: +2; *3; -1. Counter ends as" --category arithmetic --show_trace

python inference.py --prompt "Python loop: total=0; for i in range(5): total += i. total =" --category code_reasoning --show_trace

Limitations

This artifact is narrow and inspectable by design. It is not a chat assistant, not raw LLM improvement, and not proof that dense weights solved TAC v4. The claim is the no-GPU path: frozen language substrate plus explicit TS graph and program operators can move formal-task capability without retraining the model.