TS-Reasoner v10.0.0 is the current verifier-first reasoning flagship.
Release: https://github.com/BoggersTheFish/TS-Reasoner-v0/releases/tag/v10.0.0
Core line:
Candidate events enter.
TS verifies.
Runtime integrity is not claim truth.
Typed receipts show why.
v3.5.0 adds the public surface, cold-reader demo, external mini-benchmark adapter, verifier-first draft, and TensionLM proposer-boundary smoke.
Boundary: this is not a broad NLP, general theorem-proving, external benchmark victory, live TensionLM runtime, or confidence-as-proof claim.
See: docs/ts_reasoner_v3_5_public_sync.md
Canonical ecosystem map for the BoggersTheFish TS / Thinking System research stack.
TS is an engineering framework for modelling information transfer through graph structure, constraint pressure, contradiction handling, provenance, and relaxation.
The current public artifacts are mostly toy-scope or narrow-scope, but they are real, inspectable, runnable, and increasingly receipt-backed. The goal is not to claim a finished general reasoning model. The goal is to make the current systems readable: what exists, what can run, what is still limited, and what would weaken the approach.
Core docs:
- README First
- TS Stack Doctrine
- Typed Tension Spec
- TS Receipt Schema
- TS Ecosystem Map
- Credibility Roadmap
- First Contact
- Repo Taxonomy
- Flagship Route
- 20-Minute Technical Tour
- Public Signal Reduction Checklist
- GitHub Cleanup Plan
- Repo Metadata Recommendations
- Read README First for the current public stack and claim discipline.
- Run TS-Reasoner-v0 v10.0.0, inspect the runtime OS receipt, and verify that runtime integrity, generated text, and confidence remain non-proof without typed verifier support.
- Inspect TS-Benchmarks for falsification-oriented scaling receipts, diagnostics, and failure reports.
- Inspect TensionLM for controlled sigmoid pairwise tension-attention comparisons and the public runner.
- Inspect BoggersTheCIG or cig-ts-engine for provenance-aware claim/evidence graph work.
- Start with TS-Reasoner-v0 if you want runnable reasoning traces.
- Start with TS-Core if you want graph/tension runtime experiments.
- Start with TS-Benchmarks if you want the falsification and scaling harness.
- Start with TS-Codex-OS if you want project graph, tension ledger, planner, and release receipts for Codex-driven development.
- Start with CIG repos if you want provenance-aware claim/evidence graph infrastructure.
TS-Start-Here
-> TS-Reasoner-v0: verifier-first runtime OS, typed traces, replay, ledger, checkpoint, restore
-> candidate events: quarantine / repair / branch / check, typed verifier decides
-> limitations: runtime integrity is not claim truth, not a chatbot, not general reasoning
-> TS-Codex-OS: project-control substrate
-> TS-Benchmarks: falsification and scaling harness
-> TensionLM: model substrate experiment
Read this as an independent research stack with a sober receipt path:
- Research theme: inspectable reasoning systems through constraint graphs, tension telemetry, provenance, and verifier-backed traces.
- Best proof of discipline: TS-Reasoner release receipts and benchmark artifacts.
- Best falsification path: TS-Benchmarks receipts tied to clean commits, including visible failure cases.
- Best public orientation: this repo and the profile README.
- Current risk: the work is early, narrow, and partly toy-scope.
- Current upside: the artifacts are concrete enough to inspect, run, criticize, and improve.
- TS-Reasoner v10.0.0 packages the verifier-first runtime line into one auditable runtime surface: policy contracts, ordered replay, tamper-evident ledger, checkpoint/restore, recovery drill, CLI, docs, tests, and receipts. Earlier releases added typed channels, candidate containment, exported-output ingestion, learned candidate ranking, runtime kernel, replay, ledger, and checkpoint restore.
- TS-Benchmarks exists as a falsification-oriented scaling harness with receipts, diagnostics, schema validation, plots, and an explicit scale-free failure issue.
- TS-Codex-OS exists as a local-first project-control substrate for graphing repo state, tensions, planned actions, and receipts.
- TS-Core exists as a graph/tension runtime kernel.
- TensionLM exists as a model-mechanism experiment around sigmoid pairwise tension attention.
- CIG repos exist as provenance-aware claim/evidence graph infrastructure.
- The public website and profile README now route visitors toward the clean research ladder.
- TS-Reasoner is not a broad theorem prover or general reasoning model.
- TS-Benchmarks v0.1 is synthetic graph evidence, not real-world knowledge graph scalability.
- TS-Reasoner v10.0.0 is still bounded: runtime integrity is not claim truth, generated text is not proof, model confidence is not proof, and this is not a chatbot or general theorem prover.
- Current learned components are mechanism experiments, not production models.
- CIG confidence and contradiction handling are heuristic.
- TS-Codex-OS is local project-control infrastructure, not an autonomous engineering agent.
The approach would be weakened if:
- tension telemetry does not improve repair, debugging, or verification over simpler baselines,
- low-tension states routinely hide incomplete proofs or unsupported abstentions,
- provenance graphs add complexity without improving contradiction handling,
- verifier-backed traces become too noisy or expensive to inspect,
- scaling diagnostics show the graph substrate fails on common graph families without a bounded fix,
- stronger external benchmarks show no advantage over conventional ranker/verifier baselines.
TS-Start-Here
-> TS-Reasoner-v0
-> TS-Benchmarks
-> TensionLM export set boundary
-> CIG / TS-Codex-OS
The current TS-Reasoner release is v10.0.0:
Verifier-first runtime OS:
candidate event
-> runtime policy contract
-> typed verifier boundary
-> quarantine / repair / branch / check
-> replay + tamper-evident ledger + checkpoint/restore
-> receipt with zero candidate graph contamination
Historical verifier repair beneath the runtime line:
Deeper-chain support repair:
v1.6 deeper-chain current-limit case
-> structural verifier repair
-> A -> B -> C -> D acceptance tests
-> same candidate containment boundary
The current TS-Benchmarks receipt is v0.1:
Synthetic graph scaling harness, clean-commit receipts, plots, baseline comparison, and visible scale-free failure.
Lead with bounded artifacts:
Here is a small bounded reasoning system.
Here are the traces.
Here are the failures.
Here is the falsification harness.
Here is what changed from v3 to v10.
- Website: https://www.boggersthefish.com/
- GitHub: https://github.com/BoggersTheFish
- Hugging Face: https://huggingface.co/BoggersTheFish