I build local-first Python research software for decision assurance, explicit validation, and bounded game-theory experiments.
Core workflows are local or offline-first; this does not attest to offline installation, protected persistence, or production isolation. AI-generated material remains a proposal and does not replace repository-specific validation, evidence, or human responsibility.
| Project | What it demonstrates | Status and boundary |
|---|---|---|
| decision-assurance-framework | Policy-gated human decision analysis, bounded local calculations, evidence-linked reports, and an explicit decision handoff | Public v0.1.0 source-only milestone; advice-only reference implementation with no real-provider execution, autonomous action, or production service |
| spec-driven-agent-framework | Validated software specifications, evidence gates, context snapshots, host-execution intents, and bounded finite-domain solving | Experimental reference implementation; v1.0.0-rc.1 covers M0-M4 and main's M5-M8 are unreleased; it does not call LLMs or launch agents |
| bug-cause-inference-game | Cost-aware Bayesian investigation across synthetic, injected-bug, and hand-authored toy domains | Alpha v0.1.0 research prototype; P2 audits are repository-only, fixed-input, non-causal, and non-deployable; no automated repair |
| repeated-poker-analysis | Exact fixed-strategy responses, bounded commitment experiments, and certified optimization in deliberately small poker models | Alpha v0.2.1 research prototype; not a general GTO, Nash, or repeated-game equilibrium solver; external solver validation has not been performed |
| Project | What it demonstrates | Status and boundary |
|---|---|---|
| poker-deliberation-framework | Local poker calculators, strict inputs, bounded retrospective NLHE river review, and revisioned artifacts | Experimental 0.1.0 research prototype; no CFR, node locking, full-game GTO, external solver, or live-play support |
DPL_poker_ai (poker-xai) |
Decision provenance, bounded CFR/CFR+ river simulations, synthetic node-locked leaks, and DPL-bound explanation checks | Early alpha, simulation only, and limited to a facing-all-in FOLD/CALL slice; finite-iteration policies are not exact equilibria, GTO, or convergence-certified |
DPL_poker_ai is listed for inventory transparency but is not currently a
recommended starting point while its public documentation is reconciled with
the finite-iteration implementation boundary.
decision-assurance-frameworkhandles human decision records, calculations, evidence, reports, and handoff;spec-driven-agent-frameworkvalidates the software-project process and emits host-owned workflow intents.- The poker repositories answer different questions:
repeated-poker-analysisstudies bounded response and commitment models,poker-deliberation-frameworkprovides retrospective calculation workflows, andDPL_poker_aistudies synthetic leaks, decision provenance, and explanation faithfulness.
- Keep inputs, assumptions, exclusions, and trust boundaries explicit.
- Separate code presence from validation, evidence, and release readiness.
- Prefer deterministic local checks and versioned, inspectable artifacts.
- Keep AI-generated proposals separate from verified calculations and final human responsibility.
These are time-stamped design notes, not current project specifications.
- Testing an LLM Input Layer for Poker Calculators: Verified Math, Unverified Interpretation — one observed session with a manual CLI handoff; a case study, not a benchmark.
- 【ポーカー】戦略を固定したら、相手はどう変えるのか — a small-model response and commitment explainer, not a general solver claim.
- A Small Prototype for Cost-Aware Bug Investigation — v0.1.0/P1a-P1c background; later repository-only P2 work is not covered.
More writing: DEV Community · note
These projects are research prototypes or reference implementations, not production decision systems. Inspect each repository's current implementation, limitations, and verification evidence before relying on a claim.