TensionLM-117M-TS-Reasoner-v9

v9 turns the v8 boundary-aware CPU reasoner into a usable local system:

• stable trace JSON via solve_trace,
• CLI entrypoint,
• local HTTP API,
• browser demo,
• 20-example public prompt pack,
• repeatable example and server smoke verifiers.

The reasoning path remains CPU-only and uses the frozen TensionLM-117M-Reasoning-v2 substrate plus explicit TS graph/program operators.

Trace Shape

Each solve emits:

• prompt and requested/detected category,
• answer or <ABSTAIN>,
• solved flag,
• confidence and confidence band,
• active rule,
• rationale,
• boundary reason for refusals,
• graph nodes and graph edges where available.

Eval Receipts

Fixed benchmark scores:

System	TAC v2	TAC v3	TAC v4
TS-Reasoner-v9	120/120	120/120	120/120

Interactive/system receipts:

Receipt	Score
Public examples pack	20/20
Server smoke	pass
TAC-GEN mixed seed 9401	3000/3000
TAC-GEN unknown seed 9402	3000/3000

These are system scores, not raw LLM scores.

CLI

python ts_reasoner_v9.py solve --prompt "Trace Python: x=7; y=mystery(x); y is" --category code_reasoning --json
python ts_reasoner_v9.py examples
python ts_reasoner_v9.py serve --host 127.0.0.1 --port 7860

HTTP API

curl -s -X POST http://127.0.0.1:7860/solve \
  -H 'Content-Type: application/json' \
  --data '{"prompt":"Graph ledger: main(a,b); main(a,c). Resolve main* from a; terminal node:","category":"transitivity"}'

Endpoints:

• GET /
• GET /healthz
• GET /examples
• POST /solve

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.