Skip to content
This repository was archived by the owner on Jun 25, 2026. It is now read-only.

RELEASE_NOTES.md

Latest commit

 

History

History
1956 lines (1460 loc) · 73.2 KB

RELEASE_NOTES.md

File metadata and controls

1956 lines (1460 loc) · 73.2 KB
Raw

TS-OS Cognitive Physics Engine

The Cognitive Physics Engine adds the requested long-range TS-OS roadmap layer as deterministic, verifier-first substrate simulators. It implements Photonic_State_Ledger, Retrocausal_Fuzzer, Temporal_Tension_Bridge, Spectral_Coupling_Telepathy, Unified_Field_Kernel, and The_Lazy_Universe_Engine without granting any substrate proof authority.

Artifacts:

  • ts_reasoner/cognitive_physics_engine.py
  • ts_metacompute/scheduler.py
  • scripts/evaluate_cognitive_physics_engine.py
  • tests/test_cognitive_physics_engine.py
  • docs/cognitive_physics_engine.md
  • docs/TS_METACOMPUTE_STACK.md
  • artifacts/cognitive_physics_engine_report.json
  • artifacts/cognitive_physics_engine_receipt.json

Gates:

  • Photonic state is encoded as deterministic frequency slots.
  • The interference gate detects destructive cancellation in the contradiction case.
  • Temporal tension updates earlier assumption probabilities through a tamper-evident event ledger.
  • Resonance telemetry shares constraint shape, not answer packets.
  • The unified field kernel emits only zero-tension verifier-supported state.
  • The lazy universe engine composes the full stack.
  • Accepted without verifier support remains zero.
  • Candidate graph contamination remains zero.

Boundary:

  • No literal photonic chip is driven by this implementation.
  • No physical retrocausality is claimed.
  • No telepathy or instant communication is claimed.
  • No zero-point energy source is claimed.
  • No generated output is accepted without typed verifier support.

v40.0.0: Self-Hosting Verifier-First Research OS

v40.0.0 composes the v32-v39 roadmap into a bounded local research OS. It can compile taught domains into operational ontologies, maintain verifier-gated memory, compile language into typed plans, generate bounded research experiments, audit/repair reasoning state, stage confirmation-gated patches, audit the TS ecosystem graph, evaluate model proposers, and prepare a next release-candidate receipt.

v40 composes bounded v32-v39 surfaces into one self-hosting research OS receipt. It does not claim each subsystem is a fully mature autonomous implementation.

Artifacts:

  • ts_reasoner/research_os.py
  • ts_agl/compiler.py
  • scripts/evaluate_v32_v40_research_os.py
  • tests/test_research_os_v40.py
  • docs/v40_self_hosting_research_os.md
  • MODEL_CARD.md
  • RELEASE_SUMMARY_FULL.txt
  • artifacts/ontology_compiler_receipt.json
  • artifacts/verifier_gated_memory_receipt.json
  • artifacts/ts_agl_plan_receipt.json
  • artifacts/research_forge_receipt.json
  • artifacts/repair_kernel_receipt.json
  • artifacts/patch_execution_receipt.json
  • artifacts/ecosystem_brain_receipt.json
  • artifacts/model_coevolution_receipt.json
  • artifacts/research_os_receipt.json

Boundary:

  • Not AGI.
  • Not autonomous science.
  • Not unrestricted self-improvement.
  • Not free self-learning.
  • Not broad NLP understanding.
  • Not live external automation.
  • Generated text, model confidence, memory, user confirmation, repeated experience, and curriculum examples are not proof authority.
  • Typed verifier support remains the proof boundary.
  • Confirmation gates authorize bounded writes; they do not authorize truth.
  • Candidate graph contamination remains zero.

ts-spectralcompute-v0.1: Spectral Metacompute Reader

TS-SpectralCompute v0.1 adds the first deterministic metacompute substrate for the TS-Metacompute Stack. Signed TS graphs are read through a signed Laplacian, component-wise spectral tension, dominant modes, edge residuals, and candidate repair rankings.

Artifacts:

  • ts_metacompute/scheduler.py
  • ts_metacompute/spectral/signed_graph.py
  • ts_metacompute/spectral/laplacian.py
  • ts_metacompute/spectral/modes.py
  • ts_metacompute/spectral/residuals.py
  • ts_metacompute/spectral/repairs.py
  • ts_metacompute/spectral/receipts.py
  • ts_metacompute/spectral/evaluate.py
  • scripts/evaluate_spectral_metacompute.py
  • data/spectral_cases.jsonl
  • tests/test_spectral_metacompute.py
  • docs/spectral_metacompute.md
  • artifacts/spectral_metacompute_report.json
  • artifacts/spectral_metacompute_receipt.json

Boundary:

  • Spectral mode-space exposes tension, coherence, residuals, and repair pressure.
  • Spectral output is candidate evidence, not proof.
  • The scheduler can choose the spectral substrate, but acceptance still requires typed verifier support.
  • Ambiguous frustrated loops abstain from naming a unique culprit.
  • Disconnected graph components are scanned so a coherent zero-mode component cannot hide tension elsewhere.
  • Candidate graph contamination remains zero.
  • Accepted without typed verifier support remains zero.

v31.0.0: TS Project Curriculum Pack v1

v31.0.0 starts bounded curriculum learning by teaching TS-Reasoner its own project domain. The new TS Project Domain Pack v1 defines project objects, relations, operations, risks, examples, and failure modes for repo/release/artifact/receipt/claim/proof-boundary inspection while keeping free self-learning blocked.

Artifacts:

  • ts_agl/domains/ts_project.json
  • ts_agl/adapters/ts_project_adapter.py
  • ts_agl/os/ts_project_curriculum.py
  • scripts/run_ts_project_curriculum.py
  • scripts/evaluate_ts_project_curriculum.py
  • tests/test_ts_project_curriculum.py
  • docs/ts_project_curriculum.md
  • artifacts/ts_project_curriculum_report.json
  • artifacts/ts_project_curriculum_receipt.json

Boundary:

  • This is bounded curriculum learning, not free self-learning.
  • Domain packs teach objects, relations, operations, risks, examples, and failure modes.
  • Domain packs, language examples, repeated experience, and router confidence are not proof authority.
  • Unsafe self-learning overclaims are rejected.
  • Missing receipt detection is explicit.
  • Durable lesson promotion requires confirmation.
  • No external LLM is used.
  • No network call is performed.
  • No real external side effect is performed.
  • Candidate graph contamination remains zero.

v30.0.0: Verifier-First Local Agent OS

v30.0.0 composes the TS-AGL stack into a bounded verifier-first local agent OS. Natural-language requests route through typed TS calls, unsafe ambiguity abstains, writes and external effects remain gated, proof object examples expose the typed verifier boundary, and evidence dashboard metrics make the safety surface inspectable.

Artifacts:

  • ts_agl/os/chat_loop.py
  • ts_agl/os/evidence_dashboard.py
  • ts_agl/os/first_contact_demo.py
  • ts_agl/os/verifier_first_local_agent_os.py
  • ts_reasoner/proof_object_examples.py
  • scripts/run_ts_os_chat_loop_demo.py
  • scripts/evaluate_ts_os_chat_loop.py
  • scripts/build_ts_evidence_dashboard.py
  • scripts/show_proof_object_examples.py
  • scripts/run_first_contact_demo.py
  • scripts/demo_first_contact.py
  • scripts/run_ts_os_v30.py
  • scripts/evaluate_ts_os_v30.py
  • tests/test_ts_os_chat_loop.py
  • tests/test_ts_evidence_dashboard.py
  • tests/test_proof_object_examples.py
  • tests/test_first_contact_demo.py
  • tests/test_ts_os_v30.py
  • docs/first_contact.md
  • docs/ts_os_conversational_shell_loop.md
  • docs/evidence_dashboard.md
  • docs/proof_objects.md
  • docs/verifier_first_local_agent_os.md
  • artifacts/ts_os_chat_loop_report.json
  • artifacts/ts_os_chat_loop_receipt.json
  • artifacts/ts_os_chat_loop_demo.json
  • artifacts/ts_evidence_dashboard.json
  • artifacts/proof_object_examples.json
  • artifacts/first_contact_demo_report.json
  • artifacts/first_contact_demo_receipt.json
  • artifacts/first_contact_surface_report.json
  • artifacts/ts_os_v30_report.json
  • artifacts/ts_os_v30_receipt.json

Boundary:

  • This is bounded verifier-first local operation, not broad autonomous agency.
  • No external LLM is used.
  • No accidental network call is performed.
  • No real external side effect is performed.
  • Language is not proof authority.
  • Router confidence is not proof.
  • Proof examples are typed verifier evidence surfaces, not confidence claims.
  • Writes and external effects remain gated.
  • Candidate graph contamination remains zero.

v25.0.0: TS-OS v1

v25.0.0 packages the TS-AGL stack into TS-OS v1: a bounded operating layer over shell, sessions, router stack, local project operation, external adapter gating, and receipts.

Artifacts:

  • ts_agl/os/ts_os_v1.py
  • scripts/run_ts_os_v1.py
  • scripts/evaluate_ts_os_v1.py
  • tests/test_ts_os_v1.py
  • docs/ts_os_v1.md
  • artifacts/ts_os_v1_report.json
  • artifacts/ts_os_v1_receipt.json
  • artifacts/ts_os_v1_session.json

Boundary:

  • This is bounded TS-OS v1, not broad autonomous agency.
  • No external LLM is used.
  • No accidental network call is performed.
  • No real external side effect is performed.
  • Language is not proof authority.
  • Router confidence is not proof.
  • Session replay is not proof authority.
  • External gate authorization is not execution.
  • Candidate graph contamination remains zero.

v24.0.0: Controlled External Adapter Gate

v24.0.0 adds a controlled external adapter gate for the TS-OS runway. External adapter requests are dry-run by default, require confirmation, block live mode unless TS_AGL_ALLOW_EXTERNAL_LIVE=1, and prove that gate authorization is not network execution.

Artifacts:

  • ts_agl/external/adapter_gate.py
  • scripts/run_ts_agl_external_adapter_gate.py
  • scripts/evaluate_ts_agl_external_adapter_gate.py
  • tests/test_ts_agl_external_adapter_gate.py
  • docs/ts_agl_external_adapter_gate.md
  • artifacts/ts_agl_external_adapter_gate_report.json
  • artifacts/ts_agl_external_adapter_gate_receipt.json
  • artifacts/ts_agl_external_adapter_gate_eval_receipt.json

Boundary:

  • This is controlled external adapter gating, not live external automation.
  • No external LLM is used.
  • No network call is performed.
  • No real external side effect is performed.
  • Missing/wrong confirmation is blocked.
  • Live mode requires TS_AGL_ALLOW_EXTERNAL_LIVE=1.
  • Candidate graph contamination remains zero.

v23.0.0: Persistent TS-OS Session Ledger

v23.0.0 adds a persistent TS-OS session ledger over the TS-AGL shell surface. Shell/operator commands are recorded as session events, saved to disk, reloaded, replayed as a summary, and continued in a resumed session.

Artifacts:

  • ts_agl/os/session_ledger.py
  • scripts/run_ts_os_session_ledger.py
  • scripts/evaluate_ts_os_session_ledger.py
  • tests/test_ts_os_session_ledger.py
  • docs/ts_os_session_ledger.md
  • artifacts/ts_os_session_ledger.json
  • artifacts/ts_os_session_ledger_report.json
  • artifacts/ts_os_session_ledger_receipt.json
  • artifacts/ts_os_session_ledger_eval_receipt.json

Boundary:

  • This is session persistence, not broad autonomous agency.
  • No external LLM is used.
  • No real external side effect is performed.
  • Session replay is not proof authority.
  • Shell confidence is not proof.
  • Candidate graph contamination remains zero.

v22.0.0: TS-AGL Shell Surface

v22.0.0 adds a speakable shell surface for TS-AGL. The shell exposes help, route inspection, and the local project operator through one command surface, including python3 -m ts_agl.shell.

Artifacts:

  • ts_agl/shell.py
  • scripts/run_ts_agl_shell_surface.py
  • scripts/evaluate_ts_agl_shell_surface.py
  • tests/test_ts_agl_shell_surface.py
  • docs/ts_agl_shell_surface.md
  • artifacts/ts_agl_shell_surface_report.json
  • artifacts/ts_agl_shell_surface_eval_receipt.json
  • artifacts/ts_agl_shell_surface_receipt.json

Boundary:

  • This is a shell surface, not broad autonomous agency.
  • No external LLM is used.
  • No real external side effect is performed.
  • Shell route confidence is not proof.
  • Local writes remain confirmation-gated.
  • Candidate graph contamination remains zero.

v21.0.0: TS-AGL Local Project Operator

v21.0.0 adds a bounded local project operator for TS-AGL. The operator inspects local repo/docs state, uses the router stack for next-safe-action routing, stages a safe artifact write, blocks the unconfirmed write, executes only after confirmation, and emits a single project-operator report/receipt.

Artifacts:

  • ts_agl/arena/local_project_operator.py
  • scripts/run_ts_agl_local_project_operator.py
  • scripts/evaluate_ts_agl_local_project_operator.py
  • tests/test_ts_agl_local_project_operator.py
  • docs/ts_agl_local_project_operator.md
  • artifacts/ts_agl_local_project_operator_report.json
  • artifacts/ts_agl_local_project_operator_receipt.json
  • artifacts/ts_agl_project_operator_note.json

Boundary:

  • This is bounded local project operation, not broad autonomous agency.
  • No external LLM is used.
  • No real external side effect is performed.
  • Local reversible artifact writes require confirmation.
  • Router stack output is not proof authority.
  • Candidate graph contamination remains zero.

v20.0.0: TS-AGL Router Stack Arena

v20.0.0 adds a router stack arena for TS-AGL. The arena compares rule-parser routing, domain-example routing, and the tiny learned router under a safe selector that can choose a route or abstain to route_unknown.

Artifacts:

  • ts_agl/arena/router_stack_arena.py
  • scripts/evaluate_ts_agl_router_stack_arena.py
  • tests/test_ts_agl_router_stack_arena.py
  • docs/ts_agl_router_stack_arena.md
  • artifacts/ts_agl_router_stack_arena_report.json
  • artifacts/ts_agl_router_stack_arena_receipt.json

Boundary:

  • This is router arbitration, not proof authority.
  • No external LLM is used.
  • Learned confidence is not proof.
  • Selected calls still go through TSCall.
  • Risk gates, adapters, verifier boundaries, and receipts remain downstream.
  • Hard negatives must abstain to route_unknown.
  • Candidate graph contamination remains zero.

v19.0.0: TS-AGL Tiny Learned Router

v19.0.0 adds a tiny dependency-free learned router for TS-AGL. The router trains on the v18 trace-mined router dataset, predicts operation routes or abstention, and remains a proposer only.

Artifacts:

  • ts_agl/training/tiny_learned_router.py
  • scripts/train_ts_agl_tiny_router.py
  • scripts/evaluate_ts_agl_tiny_router.py
  • tests/test_ts_agl_tiny_learned_router.py
  • docs/ts_agl_tiny_learned_router.md
  • artifacts/ts_agl_tiny_router_model.json
  • artifacts/ts_agl_tiny_router_train_report.json
  • artifacts/ts_agl_tiny_router_report.json
  • artifacts/ts_agl_tiny_router_receipt.json

Boundary:

  • This is learned routing, not proof authority.
  • No external LLM is used.
  • No neural dependency is introduced.
  • Router confidence is not proof.
  • Routed calls still go through TSCall/risk/adapters/verifier boundaries.
  • Candidate graph contamination remains zero.

v18.0.0: TS-AGL Trace-Mined Router Dataset

v18.0.0 adds a trace-mined router dataset for TS-AGL. The system turns validated domain-pack language examples, arena traces, workflow traces, and hard-negative abstention cases into a reusable JSONL router dataset.

Artifacts:

  • ts_agl/training/router_dataset.py
  • scripts/build_ts_agl_router_dataset.py
  • scripts/evaluate_ts_agl_router_dataset.py
  • tests/test_ts_agl_router_dataset.py
  • docs/ts_agl_trace_mined_router_dataset.md
  • artifacts/ts_agl_router_dataset.jsonl
  • artifacts/ts_agl_router_dataset_build_report.json
  • artifacts/ts_agl_router_dataset_report.json
  • artifacts/ts_agl_router_dataset_receipt.json

Boundary:

  • This is dataset construction, not a learned router yet.
  • No external LLM is used.
  • No model training is performed.
  • Dataset rows are routing supervision only.
  • The language layer is not proof authority.
  • Candidate graph contamination remains zero.

v17.0.0: TS-AGL Domain Pack Generator

v17.0.0 adds a bounded TS-AGL Domain Pack Generator. The system converts structured teaching input into a validated domain manifest, verifies risk/confirmation rules, and proves the generated pack can teach example-based routing.

Artifacts:

  • ts_agl/teaching/domain_pack_generator.py
  • scripts/run_ts_agl_domain_pack_generator.py
  • scripts/evaluate_ts_agl_domain_pack_generator.py
  • tests/test_ts_agl_domain_pack_generator.py
  • docs/ts_agl_domain_pack_generator.md
  • artifacts/generated_domain_packs/research_notes.json
  • artifacts/ts_agl_domain_pack_generator_report.json
  • artifacts/ts_agl_domain_pack_generator_receipt.json

Boundary:

  • This is structured domain-pack generation, not broad autonomous domain learning.
  • No external LLM is used.
  • Generated packs must pass manifest validation.
  • Risky generated operations must require confirmation.
  • Generated examples are checked through the example router.
  • The language layer is not proof authority.
  • Candidate graph contamination remains zero.

v16.0.0: TS-AGL External Side-Effect Staging Arena

v16.0.0 adds a bounded external side-effect staging arena. TS-AGL can now represent an external_side_effect operation class, stage it as a pending action, block missing confirmation, block wrong confirmation, and dispatch a confirmed dry-run adapter without making a real network call.

Artifacts:

  • ts_agl/domains/external_service.json
  • ts_agl/adapters/external_service_adapter.py
  • ts_agl/arena/external_side_effect_arena.py
  • scripts/run_ts_agl_external_side_effect_arena.py
  • scripts/evaluate_ts_agl_external_side_effect_staging.py
  • tests/test_ts_agl_external_side_effect_arena.py
  • docs/ts_agl_external_side_effect_staging.md
  • artifacts/ts_agl_external_side_effect_staging_report.json
  • artifacts/ts_agl_external_side_effect_staging_receipt.json

Boundary:

  • This is external side-effect staging, not real external automation.
  • No external LLM is used.
  • No real network call is performed.
  • Missing or wrong confirmation is blocked.
  • Confirmed dispatch is dry-run and traceable.
  • The language layer is not proof authority.
  • Candidate graph contamination remains zero.

v15.0.0: TS-AGL Interactive Workflow Ledger

v15.0.0 adds a bounded interactive workflow ledger for TS-AGL. The system now stages a reversible action as pending, blocks missing confirmation, blocks wrong confirmation, executes only with the correct confirmation token, and records the full workflow event ledger in report/receipt artifacts.

Artifacts:

  • ts_agl/workflow/ledger.py
  • ts_agl/arena/interactive_workflow_arena.py
  • scripts/run_ts_agl_interactive_workflow_arena.py
  • scripts/evaluate_ts_agl_interactive_workflow_ledger.py
  • tests/test_ts_agl_workflow_ledger.py
  • docs/ts_agl_interactive_workflow_ledger.md
  • artifacts/ts_agl_interactive_workflow_ledger_report.json
  • artifacts/ts_agl_interactive_workflow_ledger_receipt.json
  • artifacts/ts_agl_interactive_workflow_marker.json

Boundary:

  • This is a bounded workflow ledger, not broad autonomous agency.
  • No external LLM is used.
  • Missing or wrong confirmation is blocked.
  • Confirmed execution is traceable.
  • The language layer is not proof authority.
  • Candidate graph contamination remains zero.

v14.0.0: TS-AGL Safe Write Arena

v14.0.0 adds the TS-AGL Safe Write Arena. The system now stages a reversible filesystem write through the same LanguageMove -> TSCall -> risk gate -> ResultPacket -> receipt surface, blocks it without confirmation, executes it with explicit confirmation, and records the boundary in report/receipt artifacts.

Artifacts:

  • ts_agl/arena/safe_write_arena.py
  • scripts/run_ts_agl_safe_write_arena.py
  • scripts/evaluate_ts_agl_safe_write_arena.py
  • tests/test_ts_agl_safe_write_arena.py
  • docs/ts_agl_safe_write_arena.md
  • artifacts/ts_agl_safe_write_arena_report.json
  • artifacts/ts_agl_safe_write_arena_receipt.json
  • artifacts/ts_agl_safe_write_marker.json

Boundary:

  • This is confirmed reversible action, not autonomous agency.
  • No external LLM is used.
  • The safe write target is restricted to artifacts/.
  • Unconfirmed writes are blocked.
  • Confirmed writes are traceable.
  • The language layer is not proof authority.
  • Candidate graph contamination remains zero.

v13.0.0: Cross-Domain TS-AGL Arena

v13.0.0 adds the Cross-Domain TS-AGL Arena. One natural-language request now coordinates multiple taught domains through the same LanguageMove -> TSCall -> ResultPacket -> rendered reply surface.

The arena currently covers git_repo, filesystem, and ts_reasoner, producing a combined cross-domain receipt with domain coverage, operation coverage, read-only gating, mutation count, candidate contamination count, and external-LLM usage.

Artifacts:

  • ts_agl/arena/cross_domain_arena.py
  • scripts/run_ts_agl_cross_domain_arena.py
  • scripts/evaluate_ts_agl_cross_domain_arena.py
  • tests/test_ts_agl_cross_domain_arena.py
  • docs/ts_agl_cross_domain_arena.md
  • artifacts/ts_agl_cross_domain_arena_report.json
  • artifacts/ts_agl_cross_domain_arena_receipt.json

Boundary:

  • This is a bounded cross-domain arena, not broad autonomous agency.
  • No external LLM is used.
  • The arena uses read-only operations only.
  • The language layer is not proof authority.
  • Risky execution remains gated.
  • Candidate graph contamination remains zero.

v12.1.0: TS-AGL Domain Example Router

v12.1.0 adds the TS-AGL Domain Example Router. Domain-pack language examples now actively teach routing behavior: natural-language phrases can route through transparent example matching into LanguageMove and TSCall objects, while unknown or low-confidence input safely abstains to route_unknown.

Artifacts:

  • ts_agl/router/example_router.py
  • scripts/evaluate_ts_agl_example_router.py
  • tests/test_ts_agl_example_router.py
  • docs/domain_example_router.md
  • artifacts/ts_agl_example_router_report.json

Boundary:

  • This is transparent example routing, not broad natural-language understanding.
  • The language layer is not proof authority.
  • No external LLM is used.
  • Risky execution remains gated.
  • Domain examples teach routing behavior; typed verifier support remains the proof boundary.

v11.9.0: Neural TS-Proposer Tiny

v11.9.0 adds a tiny pure-stdlib neural proposer baseline over verifier-labelled traces. It uses hashed sparse features, a tanh hidden layer, softmax classifiers, and verifier-gated proposed yes answers.

Artifacts:

  • training/v11_9/neural_ts_proposer_tiny.py
  • data/v11_9/neural_ts_proposer_tiny_config.json
  • scripts/v11_9/evaluate_neural_ts_proposer_tiny.py
  • tests/test_v11_9_neural_ts_proposer_tiny.py
  • docs/v11_9/NEURAL_TS_PROPOSER_TINY.md
  • artifacts/v11_9/neural_ts_proposer_tiny_model.json
  • artifacts/v11_9/neural_ts_proposer_tiny_report.json
  • artifacts/v11_9/neural_ts_proposer_tiny_receipt.json

Boundary:

  • This is a tiny neural proposer baseline, not a broad neural language model.
  • Proposed labels are not proof.
  • The verifier remains proof authority.
  • Generated text and model confidence remain non-proof.

v11.8.0: TS-Proposer-Mini Baseline

v11.8.0 adds TS-Proposer-Mini, a tiny stdlib hashed-perceptron proposer trained on v11.7 verifier trace data. It predicts answer/status/channel labels, then routes proposed yes answers through the typed verifier gate so wrong accepts remain blocked.

Artifacts:

  • training/v11_8/ts_proposer_mini.py
  • data/v11_8/ts_proposer_mini_config.json
  • scripts/v11_8/evaluate_ts_proposer_mini.py
  • tests/test_v11_8_ts_proposer_mini.py
  • docs/v11_8/TS_PROPOSER_MINI.md
  • artifacts/v11_8/ts_proposer_mini_model.json
  • artifacts/v11_8/ts_proposer_mini_report.json
  • artifacts/v11_8/ts_proposer_mini_receipt.json

Boundary:

  • This is a tiny trained proposer baseline, not a neural language model.
  • The verifier remains proof authority.
  • Generated text, proposed labels, and model confidence remain non-proof.

v11.7.0: Verifier Trace Training Dataset

v11.7.0 generates verifier-labelled JSONL train/valid/test data for future proposer models. Labels are replayed from the typed verifier and include answer, status, claim, support channel/reason, support premises, and trace hashes for accepted claims.

Artifacts:

  • training/v11_7/build_trace_training_data.py
  • data/v11_7/verifier_trace_training_config.json
  • scripts/v11_7/evaluate_trace_training_dataset.py
  • tests/test_v11_7_verifier_trace_training_dataset.py
  • docs/v11_7/VERIFIER_TRACE_TRAINING_DATASET.md
  • artifacts/v11_7/verifier_trace_train.jsonl
  • artifacts/v11_7/verifier_trace_valid.jsonl
  • artifacts/v11_7/verifier_trace_test.jsonl
  • artifacts/v11_7/verifier_trace_dataset_summary.json
  • artifacts/v11_7/verifier_trace_dataset_receipt.json

Boundary:

  • This is a verifier-labelled dataset, not a trained model.
  • The verifier remains proof authority.
  • Generated text and model confidence remain non-proof.

v11.5.0: Adversarial Claim Fuzzer

v11.5.0 attacks brittleness by adversarially mutating procedural curriculum tasks with reversed premise order, duplicate premises, irrelevant premises, confidence bait, contradiction injection, malformed claims, noisy surfaces, and paragraph noise wrappers.

Artifacts:

  • benchmarks/gpt2_boundary/adversarial_fuzzer.py
  • data/v11_5/adversarial_fuzzer_config.json
  • scripts/v11_5/evaluate_adversarial_claim_fuzzer.py
  • tests/test_v11_5_adversarial_claim_fuzzer.py
  • docs/v11_5/ADVERSARIAL_CLAIM_FUZZER.md
  • artifacts/v11_5/adversarial_fuzzer_cases.jsonl
  • artifacts/v11_5/adversarial_fuzzer_report.json
  • artifacts/v11_5/adversarial_fuzzer_receipt.json

Boundary:

  • This is deterministic adversarial fuzzing over bounded verifier-first tasks, not broad open-domain adversarial robustness.
  • The verifier remains proof authority.
  • Generated text and model confidence remain non-proof.

v11.4.0: Procedural Reasoning Curriculum

v11.4.0 reduces dependence on hand-built reasoning tasks by generating deterministic verifier-first curricula from graph templates, relation surfaces, trap types, distractors, and paragraph wrappers.

Artifacts:

  • benchmarks/gpt2_boundary/procedural_curriculum.py
  • data/v11_4/procedural_curriculum_config.json
  • scripts/v11_4/build_procedural_curriculum.py
  • scripts/v11_4/evaluate_procedural_curriculum.py
  • tests/test_v11_4_procedural_curriculum.py
  • docs/v11_4/PROCEDURAL_CURRICULUM.md
  • artifacts/v11_4/procedural_curriculum.jsonl
  • artifacts/v11_4/procedural_curriculum_report.json
  • artifacts/v11_4/procedural_curriculum_receipt.json

Boundary:

  • This is deterministic synthetic curriculum generation, not broad open-domain reasoning.
  • The verifier remains proof authority.
  • Generated text and model confidence remain non-proof.

v11.3.0: Paragraph Claim Decomposer

v11.3.0 adds bounded paragraph decomposition: short natural-language reasoning paragraphs are decomposed into canonical verifier premises and a target claim/question, then routed through the existing typed support verifier.

Artifacts:

  • ts_reasoner/paragraph_decomposer.py
  • data/v11_3/paragraph_reasoning_cases.jsonl
  • scripts/v11_3/evaluate_paragraph_decomposer.py
  • tests/test_v11_3_paragraph_decomposer.py
  • docs/v11_3/PARAGRAPH_DECOMPOSER.md
  • artifacts/v11_3/paragraph_decomposer_report.json
  • artifacts/v11_3/paragraph_decomposer_receipt.json

Boundary:

  • This is bounded paragraph decomposition, not broad natural-language understanding.
  • The verifier remains proof authority.
  • Generated text and model confidence remain non-proof.

v11.2.0: Relation Phrase Parser

v11.2.0 expands bounded claim normalization from natural is/are surfaces into relation phrases such as belongs to, kind of, type of, counts as, implies, supports, requires, cannot be, and excludes.

Artifacts:

  • ts_reasoner/relation_phrase_parser.py
  • data/v11_2/relation_phrase_cases.jsonl
  • scripts/v11_2/evaluate_relation_phrase_parser.py
  • tests/test_v11_2_relation_phrase_parser.py
  • docs/v11_2/RELATION_PHRASE_PARSER.md
  • artifacts/v11_2/relation_phrase_parser_report.json
  • artifacts/v11_2/relation_phrase_parser_receipt.json

Boundary:

  • This is bounded relation phrase normalization, not broad natural-language understanding.
  • The verifier remains proof authority.
  • Generated text and model confidence remain non-proof.

v11.1.0: Natural Claim Surface Normalization

v11.1.0 expands bounded claim parsing from strict all X are Y / no X are Y surfaces to natural is/are/every/each/any/no/not variants while preserving the same typed verifier boundary.

Artifacts:

  • ts_reasoner/claim_normalizer.py
  • data/v11_1/natural_claim_surface_cases.jsonl
  • scripts/v11_1/evaluate_claim_normalizer.py
  • tests/test_v11_1_claim_normalizer.py
  • docs/v11_1/NATURAL_CLAIM_NORMALIZATION.md
  • artifacts/v11_1/claim_normalizer_report.json
  • artifacts/v11_1/claim_normalizer_receipt.json

Boundary:

  • This is bounded surface normalization, not broad natural-language understanding.
  • Accepted claims still require typed verifier support.
  • Reverse inference, unsupported claims, and identity loops remain blocked.

v11.0.0: GPT-2 Boundary Arena

v11.0.0 adds the first GPT-2 boundary-crossing release. It introduces typed support objects, bounded support-path verification, a frozen GPT-2 boundary task format, a reproducible GPT-2-small baseline harness, and a receipt-backed TS-vs-GPT-2 arena.

Claim:

TS-Reasoner beats GPT-2-small on verifier-first controlled reasoning.

Boundary:

  • This is not a broad chatbot victory.
  • This is not a full language-model replacement claim.
  • GPT-2 generates; TS verifies.
  • Accepted TS claims require typed verifier support.

v10.5.0: TS-OS Alpha

v10.5.0 adds the bounded TS-OS Alpha flow: immutable verifier microkernel, protocol-sandboxed subprocess proposers, infrastructure-grid branch-world continuum, local file proof-grid exchange, and an integrated alpha receipt.

Artifacts:

  • ts_reasoner/ts_os.py
  • ts_reasoner/runtime_os_cli.py
  • scripts/v10_1/evaluate_ts_os_microkernel.py
  • scripts/v10_2/evaluate_ts_os_userspace.py
  • scripts/v10_3/evaluate_ts_os_continuum.py
  • scripts/v10_4/evaluate_ts_os_proof_grid.py
  • scripts/v10_5/evaluate_ts_os_alpha.py
  • tests/test_v10_1_ts_os_microkernel.py
  • tests/test_v10_2_ts_os_userspace.py
  • tests/test_v10_3_ts_os_continuum.py
  • tests/test_v10_4_ts_os_proof_grid.py
  • tests/test_v10_5_ts_os_alpha.py
  • data/v10_2/userspace_session.json
  • data/v10_3/infrastructure_grid_scenario.json
  • data/v10_5/ts_os_alpha_scenario.json
  • docs/v10_1/EPISTEMIC_MICROKERNEL.md
  • docs/v10_2/UNTRUSTED_USERSPACE_APPS.md
  • docs/v10_3/MULTI_WORLD_CONTINUUM.md
  • docs/v10_4/DECOUPLED_PROOF_GRID.md
  • docs/v10_5/TS_OS_ALPHA.md

Boundary:

  • Generated text is not proof.
  • Model confidence is not proof.
  • Runtime integrity is not claim truth.
  • Accepted common ground mutates only through kernel verifier gates.
  • Proof-grid identity signing is metadata-only until real signatures are added.

v10.0.0: Verifier-First Reasoning OS

v10.0.0 packages replay, policy contracts, tamper-evident audit ledger, checkpoint/restore, recovery drill, and receipt output into one bounded runtime session surface.

Artifacts:

  • ts_reasoner/runtime_os.py
  • ts_reasoner/runtime_os_cli.py
  • scripts/v10_0/evaluate_runtime_os.py
  • tests/test_v10_0_runtime_os.py
  • data/v10_0/runtime_os_cases.jsonl
  • data/v10_0/runtime_os_session.json
  • docs/v10_0/VERIFIER_FIRST_REASONING_OS.md
  • artifacts/runtime_os_report.json
  • artifacts/runtime_os_receipt.json

Boundary:

  • v10 is a verifier-first reasoning runtime, not a chatbot.
  • Runtime integrity is not claim truth.
  • Candidate generation is not proof.
  • Generated text is not proof.
  • Model confidence is not proof.
  • Typed verifier support remains proof authority.

v9.9.0: Runtime Recovery Drill

v9.9.0 adds a recovery drill for corrupt checkpoint rejection, reordered ledger rejection, missing-event replay divergence, and restore-then-continue runtime processing.

Artifacts:

  • ts_reasoner/runtime_recovery_drill.py
  • scripts/v9_9/evaluate_runtime_recovery_drill.py
  • tests/test_v9_9_runtime_recovery_drill.py
  • data/v9_9/runtime_recovery_drill_cases.jsonl
  • docs/v9_9/RUNTIME_RECOVERY_DRILL.md
  • artifacts/runtime_recovery_drill_report.json
  • artifacts/runtime_recovery_drill_receipt.json

v9.8.0: Runtime Policy Contracts

v9.8.0 adds machine-readable runtime policy contracts for runtime actions, checkpoint, and restore.

Artifacts:

  • ts_reasoner/runtime_policy_contracts.py
  • scripts/v9_8/evaluate_runtime_policy_contracts.py
  • tests/test_v9_8_runtime_policy_contracts.py
  • data/v9_8/runtime_policy_contract_cases.jsonl
  • docs/v9_8/RUNTIME_POLICY_CONTRACTS.md
  • artifacts/runtime_policy_contracts_v1.json
  • artifacts/runtime_policy_contracts_report.json
  • artifacts/runtime_policy_contracts_receipt.json

v3.0.0: Verifier-Guided Candidate Model

v3.0.0 is the first flagship TS-Reasoner model release.

It introduces VerifierGuidedCandidateModel, a bounded model trained on verifier-derived rows from the v3 unified dataset. The model predicts candidate status, likely verifier channels, and proposal quality, while typed verifier channels remain proof authority.

Artifacts:

  • scripts/v3/build_v3_training_dataset.py
  • scripts/v3/train_v3_verifier_guided_model.py
  • scripts/v3/evaluate_v3_verifier_guided_model.py
  • scripts/v3/run_v3_demo.py
  • artifacts/v3/v3_training_dataset.jsonl
  • artifacts/v3/v3_dataset_summary.json
  • artifacts/v3/verifier_guided_candidate_model.json
  • artifacts/v3/verifier_guided_candidate_model_report.json
  • artifacts/v3/verifier_guided_candidate_model_receipt.json
  • artifacts/v3/v3_eval_predictions.jsonl
  • docs/v3/V3_MODEL_CARD.md
  • docs/v3/V3_EVAL_REPORT.md
  • docs/v3/V3_LIMITATIONS.md

Metrics:

  • status_accuracy: 1.0
  • channel_prediction_accuracy: 0.9888
  • majority_baseline_accuracy: 0.4286
  • confidence_baseline_accuracy: 0.5714
  • beats_majority_margin: 0.5714
  • beats_confidence_margin: 0.4286
  • accepted_without_typed_support_count: 0
  • candidate_graph_contamination_count: 0
  • trace_schema_validity: 1.0
  • all_gates_passed: true

Boundary:

  • The model predicts candidate status/channels/proposal quality.
  • The model is not proof authority.
  • Typed verifier channels remain proof authority.
  • Model confidence is metadata/baseline only.
  • No TensionLM runtime is loaded.
  • No broad natural-language or general theorem-proving claim is made.

Public claim:

TS-Reasoner v3.0.0 is a bounded verifier-guided candidate model trained from typed verifier traces and active-learning rows, evaluated with proof-boundary gates that prevent model confidence from becoming proof authority.

v2.9.0: Active Learning Loop

v2.9.0 adds a smoke-scale active-learning loop on top of the verifier-trace training stack.

This release creates challenge rows from heldout verifier traces, measures baseline model performance, adds verifier-labeled challenge rows to the training set, retrains, and measures improvement.

Artifacts:

  • scripts/run_active_learning_loop_v29.py
  • tests/test_active_learning_loop_v29.py
  • docs/active_learning_loop_v29.md
  • data/active_learning_challenge_v29.jsonl
  • data/active_learning_augmented_training_v29.jsonl
  • artifacts/active_learning_status_model_v29.json
  • artifacts/active_learning_loop_v29_report.json
  • artifacts/active_learning_loop_v29_receipt.json

Metrics:

  • baseline_challenge_accuracy: 0.6667
  • active_learning_challenge_accuracy: 1.0
  • active_learning_improvement: 0.3333
  • confidence_baseline_challenge_accuracy: 0.3333
  • active_beats_confidence_margin: 0.6667
  • row_count: 91
  • base_train_rows: 35
  • challenge_rows: 12
  • augmented_train_rows: 47

Boundary:

  • Smoke-scale active-learning loop only.
  • The trained model is not proof authority.
  • Typed verifier traces define target labels.
  • Challenge labels are verifier-derived.
  • No TensionLM runtime is loaded.
  • No neural language model is trained.
  • v2.9 demonstrates the loop shape before v3.0 flagship model work.

v2.8.0: Training Loop Smoke

v2.8.0 proves that v2.7 verifier trace rows are usable supervised training signal.

This release trains a tiny smoke-scale status model from exported verifier trace rows, evaluates on the heldout stress split, and compares against simple majority/confidence baselines.

Artifacts:

  • scripts/train_from_verifier_trace_smoke.py
  • scripts/generate_verifier_trace_training_loop_smoke_receipt.py
  • tests/test_verifier_trace_training_loop_smoke.py
  • docs/verifier_trace_training_loop_smoke.md
  • artifacts/verifier_trace_status_model_v28.json
  • artifacts/verifier_trace_training_loop_smoke_report.json
  • artifacts/verifier_trace_training_loop_smoke_receipt.json

Metrics:

  • train_accuracy: 1.0
  • eval_accuracy: 1.0
  • majority_baseline_eval_accuracy: 0.4286
  • confidence_baseline_eval_accuracy: 0.5714
  • learned_beats_majority_margin: 0.5714
  • learned_beats_confidence_margin: 0.4286
  • row_count: 91
  • train_rows: 35
  • eval_rows: 56

Boundary:

  • This is a smoke-scale training loop, not a broad model-training claim.
  • The trained status model is not proof authority.
  • Typed verifier traces define target labels.
  • No TensionLM runtime is loaded.
  • No neural language model is trained.
  • Future larger models must preserve verifier/model separation.

v2.7.0: Verifier Trace Training Data

v2.7.0 exports supervised training rows from Candidate Model v2 verifier traces.

This release turns typed verifier outcomes into future training examples containing model features, model predictions, verifier status, typed channel targets, failure reasons, and explicit boundary metadata.

Artifacts:

  • ts_reasoner/trace_training_data.py
  • scripts/export_verifier_trace_training_data.py
  • tests/test_verifier_trace_training_data.py
  • docs/verifier_trace_training_data.md
  • data/verifier_trace_training_data_v27.jsonl
  • artifacts/verifier_trace_training_data_summary.json
  • artifacts/verifier_trace_training_data_receipt.json

Summary:

  • row_count: 91
  • accepted_rows: 13
  • rejected_rows: 40
  • abstained_rows: 38
  • has_model_features: true
  • has_verifier_targets: true
  • has_boundary: true
  • mean_proposal_quality: 0.2473

Boundary:

  • v2.7 exports training data; it does not train a new model.
  • Exported rows are not proof.
  • Typed verifier channels define the target labels.
  • Model confidence remains metadata only.
  • No TensionLM runtime is loaded.
  • Future training loops must keep verifier authority separate from model confidence.

v2.6.0: Candidate Model v2

v2.6.0 trains Candidate Model v2 on candidate sets derived from the v2.5 benchmark harness.

The model ranks candidate graph claims, but typed TS-Reasoner verifier channels remain proof authority. This release shows the learned candidate ranker beating confidence ordering on the benchmark-derived eval/stress surface while preserving zero accepted-without-support and zero candidate graph contamination.

Artifacts:

  • scripts/build_candidate_model_v2_dataset.py
  • scripts/train_candidate_model_v2.py
  • scripts/evaluate_candidate_model_v2.py
  • data/candidate_model_v2_train.jsonl
  • data/candidate_model_v2_eval.jsonl
  • data/candidate_model_v2_stress.jsonl
  • artifacts/candidate_model_v2.json
  • artifacts/candidate_model_v2_dataset_summary.json
  • artifacts/candidate_model_v2_report.json
  • artifacts/candidate_model_v2_receipt.json
  • docs/candidate_model_v2.md

Metrics:

  • candidate_ranking_accuracy: 1.0
  • confidence_baseline_top_accept_rate: 0.2632
  • learned_beats_confidence_baseline_margin: 0.7368
  • multi_premise_ranking_success_rate: 1.0
  • invalid_query_rejection_or_abstention_rate: 1.0
  • supported_alternative_recovery_rate: 1.0
  • malformed_input_non_accept_rate: 1.0
  • accepted_without_typed_support_count: 0
  • candidate_graph_contamination_count: 0
  • trace_schema_validity: 1.0

Boundary:

  • Candidate Model v2 is a candidate ranker, not proof authority.
  • Typed verifier channels decide accept/reject/abstain.
  • Confidence is metadata/baseline only.
  • No TensionLM runtime is loaded.
  • No broad NLP claim is made.
  • No neural language model is trained.

v2.5.0: Benchmark Harness

v2.5.0 adds a reusable train/dev/test-style benchmark harness for TS-Reasoner.

The harness evaluates bounded natural-language and relation-shaped reasoning prompts across syllogism, rule-deduction, and adversarial-invalid categories. It reports status accuracy, claim accuracy, parse success rate, invalid rejection/abstention, accepted-without-typed-support count, candidate graph contamination count, trace validity, and split/category metrics.

Artifacts:

  • data/benchmarks/
  • ts_reasoner/benchmark_harness.py
  • scripts/evaluate_benchmark_harness.py
  • tests/test_benchmark_harness_v25.py
  • docs/benchmark_harness_v25.md
  • artifacts/benchmark_harness_report.json
  • artifacts/benchmark_harness_receipt.json

Metrics:

  • case_count: 28
  • status_accuracy: 1.0
  • claim_accuracy: 1.0
  • parse_success_rate: 0.9642857142857143
  • invalid_rejection_or_abstention_rate: 1.0
  • accepted_without_typed_support_count: 0
  • candidate_graph_contamination_count: 0
  • trace_schema_validity: 1.0

Boundary:

  • This is a reusable bounded benchmark harness, not an external benchmark victory claim.
  • Malformed inputs may fail to parse, but safely abstain.
  • Typed verifier channels remain proof authority.
  • No TensionLM runtime is loaded.
  • No neural training is performed.

v2.4.0: Natural Language Claim Ingestion

v2.4.0 adds bounded natural-language claim ingestion for simple syllogistic and relation-shaped prompts.

The parser converts supported natural-language prompts into canonical premise/candidate graph claims, then delegates proof authority to the existing TS-Reasoner candidate bridge and typed verifier channels. This release does not claim broad NLP, does not load TensionLM, and does not train a neural model.

Artifacts:

  • data/natural_language_claim_cases.jsonl
  • scripts/evaluate_natural_language_claim_ingestion.py
  • docs/natural_language_claim_ingestion.md
  • artifacts/natural_language_claim_ingestion_report.json
  • artifacts/natural_language_claim_ingestion_receipt.json

Metrics:

  • case_count: 10
  • parse_expectation_rate: 1.0
  • status_expectation_rate: 1.0
  • claim_expectation_rate: 1.0
  • malformed_input_safe_abstain_rate: 1.0
  • accepted_without_typed_support_count: 0
  • candidate_graph_contamination_count: 0
  • trace_schema_validity: 1.0

Boundary:

  • Parser extracts candidate data only.
  • Typed channels decide accept/reject/abstain.
  • Malformed or unsupported input safe-abstains.
  • No candidate/parser confidence becomes proof authority.
  • No parsed candidate contaminates the premise graph.

Release Notes

v2.3.0: Scaled Learned vs Exported Candidate Comparison

v2.3.0 scales the learned-vs-exported comparison from the v2.2 seed set to a deterministic 15-case structured benchmark surface.

Each case contains one valid low-confidence candidate plus high-confidence reverse, contradiction, identity-collapse, and unsupported candidates.

Verification

python3 scripts/build_scaled_comparison_set.py
python3 scripts/evaluate_scaled_learned_vs_exported_candidate_comparison.py
python3 -m unittest discover -q

Key metrics

  • case_count: 15
  • learned_top_accept_rate: 1.0
  • exported_confidence_top_accept_rate: 0.0
  • learned_top_beats_exported_confidence_top_rate: 1.0
  • exported_high_confidence_bad_block_rate: 1.0
  • accepted_without_typed_support_count: 0
  • candidate_graph_contamination_count: 0
  • trace_schema_validity: 1.0

Boundary

This is a shaped synthetic benchmark surface. It does not claim broad natural-language reasoning, live TensionLM integration, or general candidate-ranking superiority.

v2.2.0: Learned vs Exported Candidate Comparison

v2.2.0 compares learned candidate ranking against exported-candidate confidence ordering on the same structured adversarial candidate cases.

The learned arm uses the v2.0 tiny learned candidate model. The exported baseline orders candidates by input/export confidence through the existing TensionLM export adapter. Typed verifier channels remain proof authority for both arms.

Verification

python3 scripts/evaluate_learned_vs_exported_candidate_comparison.py
python3 -m unittest discover -q

Key metrics

  • learned_top_accept_rate: 0.8571
  • exported_confidence_top_accept_rate: 0.1429
  • learned_top_beats_exported_confidence_top_rate: 0.7143
  • accepted_without_typed_support_count: 0
  • candidate_graph_contamination_count: 0
  • trace_schema_validity: 1.0

Boundary

This is a structured same-case comparison. It does not claim broad natural-language reasoning, live TensionLM integration, or general candidate-ranking superiority.

v2.1.0: Learned Candidate Model Adversarial Stress

v2.1.0 adds an adversarial stress layer for the learned candidate model introduced in v2.0.0.

This release evaluates high-confidence wrong candidates, malformed outputs, unsupported plausible claims, reverse inference traps, contradiction traps, identity-collapse traps, distractor-heavy premise sets, and missing-provenance cases.

The key boundary remains unchanged: the learned model proposes and ranks candidates, but typed verifier channels remain proof authority.

Verification

python3 scripts/evaluate_learned_candidate_model_adversarial.py
python3 -m unittest discover -q

Latest local result:

Ran 76 tests in 0.693s
OK

Key metrics

  • candidate_graph_contamination_count: 0
  • accepted_without_typed_support_count: 0
  • high_confidence_bad_block_rate: 1.0
  • high_confidence_bad_total: 13
  • unsupported_abstained_count: 6
  • trace_schema_validity: 1.0

Boundary

v2.1.0 does not claim that every bad candidate receives a hard typed rejection. Some adversarial candidates are safely blocked by abstention. The release claim is that adversarial candidates do not become proof without typed support.

v2.0.0: Learned Candidate Model

v2.0.0 adds a tiny dependency-light learned candidate model before the typed verifier.

Release scope:

  • Add ts_reasoner/learned_model/ with dataset, feature extraction, pure-Python model weights, training, inference, and evaluation.
  • Add controlled train/eval/stress JSONL splits for candidate ranking, channel activation prediction, resolver prediction, and accept/reject/abstain prediction.
  • Add a grant-facing demo where the model proposes All A are D, All D are A, and A equals D; TS-Reasoner accepts the valid transitive claim and rejects reverse/identity-collapse candidates through typed channels.
  • Preserve the verifier boundary: the learned model proposes and ranks, while TS-Reasoner typed channels remain proof authority.

Generated artifacts:

  • artifacts/learned_candidate_model.json
  • artifacts/learned_candidate_model_report.json
  • artifacts/learned_candidate_model_stress_report.json
  • artifacts/learned_candidate_model_receipt.json
  • artifacts/learned_candidate_model_demo.json

Verification:

python3 -m unittest discover
python3 scripts/build_learned_candidate_dataset.py
python3 scripts/train_learned_candidate_model.py
python3 scripts/evaluate_learned_candidate_model.py
python3 scripts/demo_learned_candidate_model.py

Claim level: experimental. This is a small learned candidate/channel model for TS-Reasoner, not an instruction model, chatbot, live TensionLM runtime, or proof authority.

v1.7.0: Deeper-Chain Support Repair

v1.7.0 repairs the deeper-chain current-limit case preserved by the v1.6.0 TensionLM export set receipt.

Release scope:

  • Repair positive all/all A -> B -> C -> D style support inside the typed transitivity channel.
  • Preserve the existing TensionLM candidate boundary: no model loading, no training, no confidence-as-proof, and no candidate edges entering proof support.
  • Add data/deeper_chain_support_repair_cases.jsonl.
  • Add scripts/evaluate_deeper_chain_support_repair.py.
  • Preserve before/after evidence for the v1.6 failure: the candidate that was abstained as a current-limit case is accepted after the structural repair.
  • Verify wrong reverse candidates still lose to directionality.

Generated artifacts:

  • artifacts/deeper_chain_support_repair_report.json
  • artifacts/deeper_chain_support_repair_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/evaluate_deeper_chain_support_repair.py

Verification result:

  • deeper_chain_acceptance_rate: 1.0.
  • wrong_reverse_rejection_rate: 1.0.
  • identity_collapse_count: 0.
  • candidate_graph_contamination_count: 0.
  • trace_schema_validity: 1.0.
  • v1_6_failure_repair_rate: 1.0.

Claim level: experimental. This is a structural verifier repair for positive all/all deeper chains, not a TensionLM integration or training claim.

v1.6.0: TensionLM Export Set Evaluation

v1.6.0 evaluates a small set of real exported TensionLM-side samples through the existing TS-Reasoner adapter and typed verification boundary.

Release scope:

  • Add data/tensionlm_export_set_cases.jsonl.
  • Source evidence from /home/boggersthefish/BoggersSpace/bozo: logs/eval/117m_transitivity_seed42.json.
  • Preserve raw TensionLM completions in candidate raw_text.
  • Evaluate export-side normalized candidate claims through the existing adapter unchanged.
  • Preserve per-sample failure reasons for wrong, malformed, contradictory, high-confidence bad, and deeper-chain current-limit cases.
  • Verify provenance, typed support, bad candidate rejection, verifier-over- confidence behavior, and zero graph contamination across the set.

Generated artifacts:

  • artifacts/tensionlm_export_set_report.json
  • artifacts/tensionlm_export_set_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/evaluate_tensionlm_export_set.py

Verification result:

  • export_set_read_success_rate: 1.0.
  • candidate_parse_success_rate: 0.8889.
  • candidate_parse_expectation_rate: 1.0.
  • provenance_preservation_rate: 1.0.
  • accepted_outputs_typed_support_rate: 1.0.
  • bad_candidate_rejection_rate: 1.0.
  • verifier_beats_confidence_rate: 1.0.
  • candidate_graph_contamination_count: 0.
  • trace_schema_validity: 1.0.
  • expected_status_accuracy: 1.0.
  • unsupported_candidate_abstention_rate: 1.0.

Claim level: experimental. This is a small exported-set boundary receipt, not live model integration into TS-Reasoner. No model is loaded and no training is performed. The malformed sample intentionally keeps parse success below 1.0 while preserving the rejection reason.

v1.5.0: Real Exported TensionLM Sample

v1.5.0 evaluates a real exported TensionLM-side sample through the existing TS-Reasoner adapter and typed verification boundary.

Release scope:

  • Add data/real_exported_tensionlm_sample.jsonl.
  • Source evidence from /home/boggersthefish/BoggersSpace/bozo: logs/eval/117m_transitivity_seed42.json.
  • Preserve raw TensionLM completions in candidate raw_text.
  • Evaluate export-side normalized candidate claims through the existing adapter unchanged.
  • Verify provenance, typed support, bad candidate rejection, verifier-over- confidence behavior, and zero graph contamination.

Generated artifacts:

  • artifacts/real_exported_tensionlm_sample_report.json
  • artifacts/real_exported_tensionlm_sample_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/evaluate_real_exported_tensionlm_sample.py

Verification result:

  • sample_read_success_rate: 1.0.
  • candidate_parse_success_rate: 1.0.
  • provenance_preservation_rate: 1.0.
  • accepted_outputs_typed_support_rate: 1.0.
  • bad_candidate_rejection_rate: 1.0.
  • verifier_beats_confidence_rate: 1.0.
  • candidate_graph_contamination_count: 0.
  • trace_schema_validity: 1.0.

Claim level: experimental. This is a cross-repo exported-sample proof, not live model integration into TS-Reasoner. No model is loaded and no training is performed.

v1.4.0: Live TensionLM Export Smoke

v1.4.0 adds a live/export-style smoke around the v1.3 adapter boundary.

Release scope:

  • Add deterministic TensionLM-style exported candidate rows.
  • Add a smoke producer that writes exported JSON candidate data.
  • Add an evaluator that feeds the export through the v1.3 adapter and candidate bridge.
  • Verify provenance preservation, bad candidate rejection, typed support for accepted outputs, verifier-over-confidence behavior, and zero graph contamination.

Generated artifacts:

  • artifacts/live_tensionlm_export_smoke.json
  • artifacts/live_tensionlm_export_smoke_report.json
  • artifacts/live_tensionlm_export_smoke_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/run_live_tensionlm_export_smoke.py
python3 scripts/evaluate_live_tensionlm_export_smoke.py

Verification result:

  • export_read_success_rate: 1.0.
  • candidate_parse_success_rate: 1.0.
  • provenance_preservation_rate: 1.0.
  • accepted_outputs_typed_support_rate: 1.0.
  • bad_candidate_rejection_rate: 1.0.
  • verifier_beats_confidence_rate: 1.0.
  • candidate_graph_contamination_count: 0.
  • trace_schema_validity: 1.0.

Claim level: experimental. v1.4.0 is not live model integration into the verifier. It is an exported-output smoke test. TensionLM-style outputs remain candidate data and never become proof without typed-channel support.

v1.3.0: Messy Language Candidate Stress

v1.3.0 stresses exported candidate ingestion with messy natural-language candidate outputs before any live model-loading work.

Release scope:

  • Add messy exported JSONL stress cases covering paraphrases, partial claims, irrelevant text, contradictory candidate sets, unsupported leaps, bad or missing confidence, ambiguous relation wording, and high-confidence wrong candidates.
  • Add pattern-based messy relation normalization in the exported-output adapter.
  • Preserve raw candidate text, normalization status, confidence status, model, row ID, and provenance through bridge traces.
  • Keep partial and ambiguous claims malformed so the existing bridge rejects them.
  • Preserve verifier authority: accepted outputs still require typed-channel support, and candidate graph contamination remains blocked.

Generated artifacts:

  • artifacts/messy_language_candidate_stress_report.json
  • artifacts/messy_language_candidate_stress_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/evaluate_messy_language_candidate_stress.py

Verification result:

  • messy_candidate_parse_success_rate: 1.0.
  • bad_candidate_rejection_rate: 1.0.
  • verifier_beats_confidence_rate: 1.0.
  • provenance_preservation_rate: 1.0.
  • candidate_graph_contamination_count: 0.
  • accepted_outputs_typed_support_rate: 1.0.
  • trace_schema_validity: 1.0.

Claim level: experimental. TS-Reasoner can robustly ingest messy exported language-model candidate outputs while preserving provenance and verifier authority. This release still does not load live model weights.

v1.2.0: Real TensionLM Candidate Adapter

v1.2.0 adds a JSONL adapter for real or exported TensionLM-style candidate outputs while preserving the v1.1 verifier boundary.

Release scope:

  • Add ts_reasoner.tensionlm_adapter.
  • Accept exported JSONL rows containing input_text, model, and candidate objects with claim, confidence, provenance, and raw_text.
  • Normalize exported candidates into CandidateClaim records.
  • Preserve model, raw text, raw candidate payload, row ID, confidence, and provenance through verification traces.
  • Reject malformed outputs and missing provenance through the existing bridge boundary.
  • Keep high-confidence bad outputs subordinate to typed-channel verification.

Generated artifacts:

  • artifacts/real_tensionlm_candidate_adapter_smoke.json
  • artifacts/real_tensionlm_candidate_adapter_report.json
  • artifacts/real_tensionlm_candidate_adapter_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/run_real_tensionlm_candidate_adapter.py
python3 scripts/evaluate_real_tensionlm_candidate_adapter.py

Verification result:

  • Adapter smoke covers valid, high-confidence bad, malformed, unsupported, and missing-provenance exported outputs.
  • verifier_beats_candidate_confidence: 1.0.
  • bad_high_confidence_rejection_rate: 1.0.
  • candidate_graph_contamination_count: 0.
  • candidate_provenance_preservation_rate: 1.0.
  • accepted_outputs_typed_support_rate: 1.0.

Claim level: experimental. TS-Reasoner can safely consume exported external model candidate outputs through a typed verification boundary. This release does not load or train real TensionLM weights, does not give model confidence proof authority, and does not claim natural-language robustness.

v1.1.0: TensionLM Candidate Bridge

v1.1.0 adds a safe candidate bridge contract for external language/model outputs:

TensionLM proposes.
TS-Reasoner verifies.
Typed channels decide.
Receipts explain.

Release scope:

  • Add CandidateClaim and CandidateVerification contracts.
  • Add mock and external-hook bridge modes.
  • Verify candidates against premise graphs without inserting candidate claims as proof support.
  • Accept, reject, or abstain through typed-channel-derived reasons.
  • Reject missing-provenance and malformed graph claims before typed verification.
  • Add adversarial stress for high-confidence bad candidates, unsupported candidates, malformed candidates, missing provenance, contradiction probes, and candidate graph contamination.

Generated artifacts:

  • artifacts/tensionlm_candidate_bridge_demo.json
  • artifacts/tensionlm_candidate_bridge_report.json
  • artifacts/tensionlm_candidate_bridge_receipt.json
  • artifacts/tensionlm_candidate_bridge_adversarial_report.json
  • artifacts/tensionlm_candidate_bridge_adversarial_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/demo_tensionlm_candidate_bridge.py
python3 scripts/evaluate_tensionlm_candidate_bridge.py
python3 scripts/evaluate_tensionlm_candidate_bridge_adversarial.py

Verification result:

  • 53 unittest tests passed.
  • Normal bridge eval: 1.0 expected status accuracy, case success rate, typed reason rejection rate, provenance preservation, and trace schema validity.
  • Adversarial bridge eval: 1.0 verifier-beats-candidate-confidence, bad-high-confidence rejection, unsupported-candidate abstention, malformed-candidate rejection, provenance-required rate, and trace schema validity.
  • Candidate graph contamination count: 0.

Claim level: experimental. v1.1.0 proves external candidate proposals can be safely admitted without becoming proof. It does not load real TensionLM weights or claim natural-language robustness.

Typed-Channel Release Receipt

This release-receipt branch summarizes the full TS-Reasoner typed-channel arc:

TS-Core-backed typed channels
-> learned typed-channel calibrator
-> generalization stress
-> structural feature repair

Generated artifact:

  • artifacts/typed_channel_release_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/generate_typed_channel_release_receipt.py

Claim level: experimental. The public story is that TS-Reasoner now has TS-Core-backed typed tension channels plus a learned calibrator; stress testing exposed structural generalization failures, and query-relevant graph features repaired those failures on the current stress benchmark.

Structural Feature Repair for Typed-Channel Calibration

This release tests whether the calibrator failures exposed by generalization stress are structural-feature gaps rather than failures of the typed-channel approach.

Core change:

  • Add query-relevant graph features for path length, distractor ratio, quantifier signatures, contradiction placement, and candidate operation requirements.
  • Preserve the original generalization stress report as the failure receipt.
  • Add a repaired stress evaluator comparing original_calibrator, + path features, + distractor features, + quantifier features, + contradiction-placement features, and full_structural_features.

Generated artifacts:

  • artifacts/typed_channel_calibrator_structural_features_report.json
  • artifacts/typed_channel_calibrator_structural_features_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/evaluate_typed_channel_calibrator_structural_features.py

Claim level: experimental. The result supports targeted structural repair, not broad reasoning generalization. TensionLM remains out of scope.

Typed-Channel Calibrator Generalization Stress

This release tests whether the typed-channel calibrator generalizes beyond the exact trace surface it was trained on.

Stress cases include variable renaming, deeper chains, distractor premises, quantifier traps, contradiction placement, reverse/identity adversarial queries, heldout relation shapes, and noisy surface forms.

Generated artifacts:

  • data/typed_channel_calibrator_stress.jsonl
  • artifacts/typed_channel_calibrator_stress_report.json
  • artifacts/typed_channel_calibrator_stress_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/evaluate_typed_channel_calibrator_stress.py

Honest outcomes are explicit: clean generalization, partial generalization with depth/feature limits, or overfit detected by receipt. TensionLM remains out of scope.

Learned Typed-Channel Calibrator

This release tests whether TS-Reasoner can learn to activate and prioritize typed tension channels from trace-level supervision, rather than learning reasoning behaviour end-to-end.

Core change:

  • Add a tiny dependency-light calibrator for typed-channel activation, channel weights, and resolver priority.
  • Build channel-level training rows from existing typed tension benchmark/demo traces.
  • Compare hand_coded_baseline, learned_activation, learned_channel_weight, learned_resolver_priority, and full_calibrator.
  • Preserve the existing public trace schema; the calibrator is an evaluation artifact, not a replacement for deterministic resolvers.

Generated artifacts:

  • data/typed_channel_calibrator_dataset.jsonl
  • artifacts/typed_channel_calibrator.json
  • artifacts/typed_channel_calibrator_report.json
  • artifacts/typed_channel_calibrator_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/build_typed_calibrator_dataset.py
python3 scripts/train_typed_channel_calibrator.py
python3 scripts/evaluate_typed_channel_calibrator.py

Claim level: experimental. The research step is the training-target shift: from behavior imitation to typed operational channel calibration. No TensionLM bridge or large-model training is included.

Typed Tension Traces: TS-Core-backed channel reasoning

This release adds TS-Core-backed typed tension traces while preserving the existing TS-Reasoner public trace contract.

Core change:

  • The system now emits per-channel reasoning traces, showing which typed tensions activated, how they resolved, and whether the answer settled.

Included channels:

  • logic_transitivity
  • identity_preservation
  • directionality
  • surface_structure
  • confidence_abstention
  • contradiction
  • quantifier_scope

Generated artifacts:

  • artifacts/typed_tension_demo.json
  • artifacts/typed_tension_benchmark_report.json
  • artifacts/typed_tension_receipt.json

Verification:

python3 -m unittest discover
python3 scripts/demo_typed_tension.py
python3 scripts/evaluate_typed_tension.py

Claim level: demo. This release separates specific reasoning failure modes into typed operational channels on small curated examples. It does not claim general reasoning, theorem proving, or broad natural-language robustness.

v1.0.0

v1.0.0 is the stable public trace-contract release. It keeps the same inspectable output schema, adds focused public docs, adds adversarial known-limit cases, and runs tests in GitHub Actions.

Release scope:

  • Freeze and document the public JSON output shape in TRACE_SCHEMA.md.
  • Document the v1 benchmark receipt in BENCHMARKS.md.
  • Document explicit non-claims and known failures in LIMITATIONS.md.
  • Add data/external_benchmark_v1.jsonl with expected passes and known limits.
  • Add scripts/evaluate_v1_baseline.py.
  • Add optional TS-Reasoner + TensionLM bridge tooling.
  • Add a tiny TensionProofLM-22M target smoke-training/eval receipt.
  • Add CI via .github/workflows/tests.yml.

Verification:

python3 -m unittest discover
python3 scripts/evaluate_v1_baseline.py
python3 scripts/run_tensionlm_bridge.py --offline
python3 scripts/run_tensionprooflm_smoke.py

The release claim is narrow: this is stable enough for another technical reader to inspect or build on. It is not a broad reasoning benchmark or large model.

v0.9.0

v0.9.0 closes the narrow proof-chain gap exposed by v0.8. The release adds explicit positive universal bridge support for small transitive all/all chains while keeping the v0.8 benchmark fixture shape stable.

Release scope:

  • Add shared transitive proof-chain support for normalized all A are B chains.
  • Use that support in candidate generation, CIG support checks, and tension ranking.
  • Add scripts/evaluate_v09_proof_chains.py.
  • Generate artifacts/v09_proof_chain_report.json.
  • Add docs/v09_proof_chain_support.md.

Verification:

python3 -m unittest discover
python3 scripts/evaluate_v09_proof_chains.py

Verification result:

  • 27 unittest tests passed.
  • v0.9 benchmark ran the same 10 externalized small-reasoning tasks.
  • full_control_loop: 10/10 correct, 10/10 settled, mean global tension 0.0.
  • small_proof_chain/full_control_loop: 2/2 correct.
  • Remaining limit: positive universal chains only; still toy-scope and normalized.

v0.8.0

v0.8.0 adds the first externalized small benchmark harness. It turns the v0.7 bounded control loop into a repeatable baseline comparison over curated external-style tasks normalized into TS-Reasoner relation form.

Release scope:

  • Add data/external_benchmark_v08.jsonl with ten tasks across five categories.
  • Add ts_reasoner.benchmark with loader, scorer, runner, and baseline summaries.
  • Add scripts/evaluate_v08_external_benchmark.py.
  • Generate artifacts/v08_external_benchmark_report.json.
  • Add docs/v08_external_benchmark_harness.md.
  • Add benchmark regression tests.

Verification:

python3 -m unittest discover
python3 scripts/evaluate_v08_external_benchmark.py

Verification result:

  • 26 unittest tests passed.
  • v0.8 benchmark ran 10 externalized small-reasoning tasks.
  • direct: 4/10 correct, mean global tension 0.2141.
  • random_selector: 5/10 correct, mean global tension 0.1771.
  • ranker_only: 8/10 correct, mean global tension 0.0767.
  • full_control_loop: 8/10 correct, 10/10 settled, mean global tension 0.0.
  • Known gap: full_control_loop fails both small_proof_chain tasks by settling to low-tension abstentions.

This is a toy-scope receipt, not a broad benchmark claim.

v0.7.0

v0.7.0 closes the residual no-compression failure exposed by v0.6. Compression now removes redundant non-premise claims that are already represented by earlier graph claims, allowing contradiction repair traces to settle instead of stopping with no_compression_available.

Release scope:

  • Extend COMPRESS_TRACE to remove redundant non-premise claims.
  • Preserve exact-duplicate compression behavior.
  • Add scripts/evaluate_v07_loop.py.
  • Generate artifacts/v07_loop_eval.json.
  • Add a regression test for the contradiction-forced-answer closure path.

Verification:

python3 -m unittest discover
python3 scripts/evaluate_v07_loop.py

Verification result:

  • 23 unittest tests passed.
  • v0.7 loop eval ran 4 hard cases.
  • 4/4 hard cases settled.
  • Mean global tension dropped from 0.4552 to 0.0.

v0.6.0

v0.6.0 replaces one-shot candidate operation routing with a bounded multi-step tension-control loop. Each candidate can now cycle through tension evaluation, operation routing, state transition, verifier rescore, and residual logging until it settles or reaches a bounded stop condition.

Release scope:

  • Add OperationRouter.run_until_stable(max_steps=5).
  • Keep run_once() for v0.4/v0.5 compatibility.
  • Add cycle-level operation traces.
  • Add concrete handlers for accept, repair, compression, localization, and goal verification.
  • Add harder v0.6 loop cases requiring repeated transitions.
  • Add scripts/evaluate_v06_loop.py.
  • Generate artifacts/v06_loop_eval.json.

Verification:

python3 -m unittest discover
python3 scripts/evaluate_v06_loop.py

Verification result:

  • 22 unittest tests passed.
  • v0.6 loop eval ran 4 hard cases.
  • 3/4 hard cases settled.
  • Mean global tension dropped from 0.4552 to 0.0.
  • One failed-to-settle case remains: v06_contradiction_forced_answer.

v0.5.0

v0.5.0 adds a residual-trained coupling matrix. The learner replays v0.4 candidate repair transitions, measures before/after coordinated tension drops, and trains channel-to-channel coupling weights from successful repairs.

Release scope:

  • Add train_residual_coupling_matrix().
  • Add scripts/train_coupling_matrix.py.
  • Add learned matrix loading through TensionCoordinator.from_json().
  • Add CLI support via --coupling-matrix.
  • Generate artifacts/learned_coupling_matrix_v05.json.
  • Generate artifacts/learned_coupling_matrix_summary.json.

Verification:

python3 -m unittest discover
python3 scripts/train_coupling_matrix.py
python3 inference.py --question "If some A are B and all B are C, are all A C?" --premise "Some A are B." --premise "All B are C." --coupling-matrix artifacts/learned_coupling_matrix_v05.json --trace /tmp/ts-reasoner-v05-trace.json

Verification result:

  • 21 unittest tests passed.
  • Coupling learner trained on 168 candidates with 126 successful repair examples.
  • CLI loaded the learned matrix and preserved the repaired direct-candidate transition.

v0.4.0

v0.4.0 adds the coordinated tension-state repair loop. The pipeline now runs specialist tension agents over each candidate chain, propagates their signals through a coupling matrix, routes one bounded operation, applies a repair when available, and records before/after residuals in the JSON trace.

Release scope:

  • Add logic, goal, repair, and compression tension agents.
  • Add an explicit coupling matrix and coordinated tension field.
  • Add OperationRouter for one-step closed-loop candidate repair.
  • Accept repaired candidate states only when global tension does not increase.
  • Preserve top-level v0 trace fields while adding operation-loop telemetry.
  • Extend the toy CIG verifier so existential bridge repairs can settle: some A are B plus all B are C supports some A are C.

Verification:

python3 -m unittest discover

Verification result:

  • 19 unittest tests passed.
  • Release receipt: artifacts/release_receipt_v0.4.0.json.

v0.1.0

TS-Reasoner-v0 is the first public proof-of-concept release of the inspectable TS reasoning telemetry pipeline.

This release includes a deterministic standard-library reasoner that maps question -> candidate chains -> CIG checks -> tension issues -> repair suggestions -> selected low-tension answer -> JSON trace. It is intentionally not a trained model or benchmark claim; v0 establishes the trace contract and hand-coded tension field that future learned rankers and generators can plug into.

Included:

  • Deterministic candidate-chain generation.
  • Claim-Interaction Graph extraction with provenance.
  • Heuristic local/global tension scoring.
  • Traceable repair suggestions.
  • CLI, demo script, examples, tests, generated artifacts, model card, and MIT license.

v1-learned-ranker draft

v1 adds the first learned tension-ranker while preserving the v0 JSON trace schema. On synthetic heldout reasoning tasks, the learned ranker matches or improves answer-quality scoring over the heuristic baseline, with ablations showing which trace features carry the signal. Because all learned ablations reach 1.0 on the current heldout set, this should be read as a schema-preserving learned-ranker smoke test, not evidence of robust general reasoning ability.

Added before merging v1:

  • Adversarial synthetic cases where confident surface wording masks wrong logic.
  • Heldout template families: symbolic A/B/C-style training, natural-term evaluation.
  • Ablation table comparing heuristic ranker, learned ranker, learned without CIG features, learned without issue-kind features, and random baseline.

v0.3.0

v0.3.0 adds a learned candidate-proposal experiment while preserving the v0 JSON trace contract. Candidate chains remain verified by the existing CIG, tension-ranker, repair, and trace pipeline.

This is the first TS-Codex-guided TS-Reasoner release. TS-Codex-OS v0.1.0 inspected TS-Reasoner-v0 before merge and detected release-control tensions around verification receipts and stale artifacts; v0.3.0 closes those tensions with regenerated artifacts and release receipts.

Release scope:

  • Learn which candidate-chain templates to propose from synthetic rows.
  • Keep generated candidates inspectable as ReasoningChain objects.
  • Pass all generated chains through the existing CIG checker, ranker, repairer, and JSON trace exporter.
  • Add coverage metrics for candidate count, stable candidate inclusion, adversarial suppression, and contradiction-aware inclusion.
  • Compare deterministic generator, learned generator, learned generator plus safety fallback, and random candidate proposer.
  • Do not claim full LLM generation, open-ended reasoning generation, or robust natural-language reasoning.

Verification:

python3 -m unittest discover
python3 demo_reasoning.py
python3 inference.py --question "If all A are B and all B are C, are all A C?"
python3 scripts/build_synthetic_dataset.py
python3 scripts/train_learned_ranker.py
python3 scripts/compare_rankers.py
python3 scripts/train_candidate_generator.py
python3 scripts/evaluate_candidate_generators.py
PYTHONPATH=/home/boggersthefish/BoggersSpace/TS-Codex-OS python3 -m ts_codex_os.cli status --project-path .

Verification result:

  • 15 unittest tests passed.
  • Demo eval regenerated artifacts/eval_summary.json with 5/5 toy tasks matching expected behavior.
  • CLI inference regenerated artifacts/latest_trace.json and returned all A are C. with global tension 0.0000.
  • Ranker artifacts and comparison artifacts were regenerated.
  • Candidate-generator artifacts and coverage artifacts were regenerated.
  • TS-Codex-OS v0.1.0 status reported 0 tensions after receipts were added.
  • Release receipt: artifacts/release_receipt_v0.3.0.json.
  • Artifact receipt: artifacts/artifact_receipt_v0.3.0.json.