| --- |
| license: mit |
| tags: |
| - reasoning |
| - interpretability |
| - candidate-ranking |
| - typed-verifier |
| - thinking-system |
| library_name: other |
| pipeline_tag: text-classification |
| --- |
| |
| # TS-TinyVerifier-v0 |
|
|
| `TS-TinyVerifier-v0` is the small learned candidate/channel model artifact for |
| `TS-Reasoner v2.0.0: Learned Candidate Model`. |
|
|
| This is not an instruction model. It is not a chatbot. It is not a standalone |
| verifier. It is a tiny pure-Python linear model that proposes/ranks structured |
| candidate claims and predicts typed-channel signals for TS-Reasoner. |
|
|
| TS-Reasoner remains the verifier. |
|
|
| ## Boundary |
|
|
| ```text |
| learned candidate model proposes/ranks |
| -> TS-Reasoner candidate bridge |
| -> typed channels verify |
| -> receipt records accepted / rejected / abstained candidates |
| ``` |
|
|
| Candidate confidence is metadata. It is not proof authority. Accepted outputs |
| require typed-channel support, and candidate graph contamination must remain |
| `0`. |
|
|
| ## Included Files |
|
|
| - `learned_candidate_model.json`: pure-Python model weights and metadata. |
| - `learned_candidate_model_train.jsonl`: controlled structured training split. |
| - `learned_candidate_model_eval.jsonl`: controlled evaluation split. |
| - `learned_candidate_model_stress.jsonl`: adversarial/stress split. |
| - `learned_candidate_model_report.json`: eval report. |
| - `learned_candidate_model_stress_report.json`: stress report. |
| - `learned_candidate_model_receipt.json`: release receipt. |
| - `example_trace_learned_candidate_model_demo.json`: grant-facing demo trace. |
| - `DATASET_CARD.md`: dataset description and limitations. |
|
|
| ## Metrics |
|
|
| Eval split: |
|
|
| - `candidate_ranking_accuracy`: `1.0` |
| - `accepted_candidate_support_rate`: `1.0` |
| - `bad_candidate_rejection_rate`: `1.0` |
| - `verifier_beats_model_confidence_rate`: `1.0` |
| - `channel_activation_accuracy`: `0.9531` |
| - `resolver_prediction_accuracy`: `0.875` |
| - `abstention_accuracy`: `1.0` |
| - `candidate_graph_contamination_count`: `0` |
| - `trace_schema_validity`: `1.0` |
| - `deeper_chain_success_rate`: `1.0` |
| - `distractor_robustness`: `1.0` |
|
|
| Stress split: |
|
|
| - `candidate_ranking_accuracy`: `1.0` |
| - `accepted_candidate_support_rate`: `1.0` |
| - `bad_candidate_rejection_rate`: `1.0` |
| - `verifier_beats_model_confidence_rate`: `1.0` |
| - `channel_activation_accuracy`: `0.9886` |
| - `resolver_prediction_accuracy`: `1.0` |
| - `abstention_accuracy`: `1.0` |
| - `candidate_graph_contamination_count`: `0` |
| - `trace_schema_validity`: `1.0` |
| - `deeper_chain_success_rate`: `1.0` |
| - `distractor_robustness`: `1.0` |
|
|
| ## Demo |
|
|
| Input: |
|
|
| ```text |
| All A are B. |
| All B are C. |
| All C are D. |
| Question: Are all A D? |
| ``` |
|
|
| Model candidates: |
|
|
| - `All A are D` |
| - `All D are A` |
| - `A equals D` |
|
|
| Verifier result: |
|
|
| - accepts `All A are D`, |
| - rejects `All D are A` because reverse inference is blocked, |
| - rejects `A equals D` because identity collapse is blocked, |
| - records `candidate_graph_contamination_count: 0`. |
|
|
| ## Limitations |
|
|
| - Synthetic, parser-controlled structured examples. |
| - Tiny linear model, not a language model. |
| - No live TensionLM runtime is loaded. |
| - Not suitable for production decisions. |
| - Not a formal proof system. |
| - Model predictions are advisory; typed verification decides acceptance. |
|
|
| ## Source |
|
|
| GitHub release: |
| https://github.com/BoggersTheFish/TS-Reasoner-v0/releases/tag/v2.0.0 |
|
|
