TS-Reasoner-v0

TS-Reasoner is a verifier-first reasoning system.

Core line:

LLMs propose.
TS verifies.
Confidence is not proof.
Typed traces show why.

It separates candidate generation, learned/advisory ranking, typed proof verification, and traceable rejection or abstention.

candidate proposer
-> learned/advisory ranking
-> typed verifier channels
-> accept / reject / abstain trace
-> receipt

Current flagship release

Current release:

https://github.com/BoggersTheFish/TS-Reasoner-v0/releases/tag/v4.3.0

v4.3.0 adds a bounded natural-language reasoning shell: prompts are converted into relation premises, candidate claims, verifier decisions, and natural-language answers.

It adds:

• v3.1 public surface artifact
• v3.2 cold-reader demo trace
• v3.3 external mini-benchmark adapter
• v3.4 verifier-first reasoning draft
• v3.5 TensionLM proposer boundary
• v3.6 scaled proposer boundary evaluation
• v3.7 real exported candidate batch
• v3.8 external benchmark translation pack
• v3.9 live proposer dry-run interface
• v4.0 live proposer sandbox
• v4.1 external JSONL backend proof
• v4.2 GPT-2 output fixture adapter
• v4.3 natural-language reasoning shell

Flagship evidence

v3.3 external adapter:

status_accuracy: 1.0
wrong_accept_count: 0
accepted_without_typed_support_count: 0
trace_schema_validity: 1.0

v4.3 natural-language reasoning shell:

nl_case_count: 12
extraction_success_rate: 1.0
candidate_generation_success_rate: 1.0
natural_language_answer_accuracy: 1.0
verifier_selection_accuracy: 1.0
wrong_accept_count: 0
accepted_without_typed_support_count: 0
candidate_graph_contamination_count: 0
abstention_correctness: 1.0
trace_schema_validity: 1.0
gpt2_comparison_claim: false
broad_nlp_claim: false
confidence_is_not_proof: true

Interpretation:

Across 12 bounded natural-language prompts, TS-Reasoner extracts relation structure, verifies candidate claims, and renders natural-language answers while keeping fluency outside the proof boundary.

Run the current receipts

python3 scripts/v3_2/run_cold_reader_demo.py
python3 scripts/v3_3/evaluate_external_minibench_v33.py
python3 scripts/v3_5/evaluate_tensionlm_proposer_boundary_v35.py
python3 scripts/v3_6/evaluate_scaled_proposer_boundary_v36.py
python3 scripts/v3_7/evaluate_real_exported_candidate_batch_v37.py
python3 scripts/v3_8/evaluate_external_benchmark_translation_pack_v38.py
python3 scripts/v3_9/evaluate_live_proposer_dry_run_interface_v39.py
python3 scripts/v4_0/run_live_proposer_sandbox_v40.py
python3 scripts/v4_0/run_live_proposer_sandbox_v40.py --external-jsonl-backend data/v4_1_external_jsonl_backend_proof/external_backend_candidates_v41.jsonl
python3 scripts/v4_2/adapt_gpt2_output_fixtures_v42.py
python3 scripts/v4_0/run_live_proposer_sandbox_v40.py --external-jsonl-backend data/v4_2_gpt2_output_fixture_adapter/adapted_gpt2_external_backend_candidates_v42.jsonl
python3 scripts/v4_3/run_natural_language_reasoning_shell_v43.py
python3 -m unittest discover -q

What this is

TS-Reasoner is:

• a bounded verifier-first reasoning artifact;
• a typed trace system;
• a candidate rejection and abstention system;
• a safe bridge for learned or language-model candidate proposers;
• a receipt-first research surface for verifier-first reasoning.

What this is not

TS-Reasoner is not:

• a chatbot;
• a general theorem prover;
• a broad natural-language understanding system;
• an external benchmark victory claim;
• live TensionLM runtime integration;
• a system where model confidence becomes proof authority.

Why this matters

Language models often entangle candidate generation, confidence, and proof.

TS-Reasoner keeps those roles separate.

A model may propose or rank a candidate claim, but the claim is not accepted unless typed verifier channels support it.

Core boundary

candidate generation != proof
model confidence != proof
typed verifier support = proof boundary

Key v3.5 files

• docs/v3_1/PUBLIC_SURFACE.md
• examples/cold_reader_demo/readable_trace.md
• docs/v3_3/EXTERNAL_MINIBENCH_ADAPTER.md
• docs/v3_4/VERIFIER_FIRST_REASONING_DRAFT.md
• docs/v3_5/TENSIONLM_PROPOSER_BOUNDARY.md
• docs/v3_6/SCALED_PROPOSER_BOUNDARY.md
• docs/v3_7/REAL_EXPORTED_CANDIDATE_BATCH.md
• docs/v3_8/EXTERNAL_BENCHMARK_TRANSLATION_PACK.md
• docs/v3_9/LIVE_PROPOSER_DRY_RUN_INTERFACE.md
• docs/v4_0/LIVE_PROPOSER_SANDBOX.md
• docs/v4_1/EXTERNAL_JSONL_BACKEND_PROOF.md
• docs/v4_2/GPT2_OUTPUT_FIXTURE_ADAPTER.md
• docs/v4_3/NATURAL_LANGUAGE_REASONING_SHELL.md

Release ladder

Version	Core addition	Boundary
v1.x	typed tension channels and TensionLM candidate bridge	TensionLM output remains candidate data
v2.x	learned candidate models and verifier-trace training	learned models remain advisory
v3.0	verifier-guided candidate model	typed verifier remains proof authority
v3.5	public surface, cold demo, external adapter, proposer boundary	confidence is not proof
v3.6	scaled proposer boundary evaluation	high-confidence candidates remain candidate data
v3.7	real exported candidate batch	provenance survives the proof boundary
v3.8	external benchmark translation pack	answer keys remain metadata, not proof
v3.9	live proposer dry-run interface	v4 runtime contract is ready, but not live yet
v4.0	live proposer sandbox	bounded backend emits candidates, TS verifies
v4.1	external JSONL backend proof	external candidate rows enter the sandbox safely
v4.2	GPT-2 output fixture adapter	GPT-2-shaped generated text becomes candidate data
v4.3	natural-language reasoning shell	bounded NL prompts become verified answers

One-command baseline

python3 inference.py --question "If all A are B and all B are C, are all A C?"

That writes artifacts/latest_trace.json and prints the selected answer, selected chain, and global tension.