Thirsty's Design Bench

An AI coding-agent prototype whose repository implements a pre-action governor against 15 Core Operating Laws. Complete call-site coverage has not been independently established.

A local-first, offline-capable governance core for an IDE: immutable signed decision records, an append-only governance journal with replay, and a four-channel evidence firewall around retrieval.

Reusable agent skills and change-control discipline for the Project-AI lab itself — a constitution, a deprecation policy, and claims-boundary docs versioned alongside the code they govern.

System and developer prompts, refusal and verification policies, and a compliance test runner — scoped deliberately as configuration guidance rather than runtime enforcement. Most of the content is written; packaging and versioning aren't settled yet.

Per-action risk profiling and a deliberation engine, plus a self-examination layer for an agent's own purpose legitimacy and drift. Phase-based; earlier phases complete and tested, later phases in progress.

Signing-key-based trust, audit export and replay, and schema/migration verification for agent execution. Past an internal self-audit that found and fixed real issues; not yet general availability.

A Dockerized FastAPI service whose inspected source reports 77 routes across 46 modules, imports Caretaker Foundation, and includes a deterministic tamper-detection path. Universal request coverage remains unverified.

A reasoning pipeline that checks whether a claim is actually grounded before it's allowed to stand — with a real, passing test suite behind the happy-path and failure cases alike.

A constitutional model-training and state-authority system — deterministic curriculum generation, signed checkpoints, an epoch controller, a hostile-audit validation suite. Its own docs are explicit that current validation runs against a simulated backend, not completed real-model training.

A dataset pipeline turning a design archive into training, validation, and retrieval data planes for constitutional reasoning. Explicitly a draft — the constitution it's built from is itself unratified pending review.

The machine-readable enforcement package for a fail-closed action gate — CEL-based rule evaluation, Ed25519-signed authority and approval documents, and an independent out-of-process evaluator — currently pending the owner's own cryptographic ratification signature.

A heavily formal evaluation harness for an epistemic-defense intelligence derived from Thirsty's Standard V3 — a SHA-256-hashed manifest, a canonical test catalog, and a reviewer-trust schema, with zero executed evaluations run against it yet.

A curated set of gold-standard, blocked, and handoff example outputs across 20 bounded agent roles — meant to ground and train Project-AI sub-agents, prepared but not yet wired into the runtime.

A small, mythologically-framed module meant as a shared memory and judgment layer woven between multiple governance agents — deliberately not a log, and deliberately not a database, by its own description.

The eBPF-based mechanism named alongside OCEE on this site's own homepage, meant to drop safety enforcement to the kernel layer — monitoring real process mutations and syscalls instead of parsing model output for 'intent.'

A compressed, symbolic grammar meant to encode governance rules deterministically instead of as thousands of tokens of natural-language policy prose — with a binary counterpart (TSCG-B) for lower-overhead evaluation.

A dedicated penetration-testing framework scaffold for Project-AI's own systems — configs, policy, and reporting structure named and directory-scoped, with no populated findings yet.

A pitched multi-team (white/grey/red/black/blue/green/purple) adversarial-simulation sandbox suite — idea stage only, sharing the same starter scaffold as the human-side ideas below.

The public repository describes origin-authenticity and byte-integrity mechanisms using Ed25519, SHA-256/SHA-3-512/BLAKE3, RFC 3161, a Merkle-tree ledger, and key ceremonies. Its Makefile exposes install and build targets. Those source-level observations are not an independent cryptographic audit, and its own docs disclaim truthfulness and intent validation.

A proposed provenance network for media — identity attestation, signing, a transparency log, dispute resolution. Early concept and scaffold today, genuinely earlier-stage than its working title implies.

Production-grade, auditable enterprise mobile device recovery — ticket-driven, MDM/UEM connectors (Intune, Jamf, Workspace ONE, Knox), tamper-evident audit logging. Real, git-tracked, actively committed to.

Merges an Android Xposed-based alternative-unlock module with a broader Python device-service toolkit into one workspace — desktop-side servicing and device-side unlock fallback logic in one place, with real build scripts for both sides.

Real-time distress-signal detection and routing, built with cryptographic audit trails and defense-in-depth as first-class requirements from the start, not afterthoughts bolted on later.

A phone-first Android app, private by default: no remote account, no microphone permission, no continuous monitoring, no clinical diagnosis claims. Room-backed mood check-ins and a Memory Vault, encrypted at rest with Android Keystore, and crisis support that routes straight to 988 through the phone's own dialer — not a third-party service in the loop.

A unified service consolidating five specialized modules for social reconnection and community formation.

Find "the person who was kind to you in 2003 and disappeared" — not networking, not matchmaking. One of five named ideas consolidated into the working service below.

Not dating-app matching — microservices meant to build "conditions for genuine community formation" around shared interests. One of five named ideas consolidated into the working service below.

Microservices per government or benefit program, mapping "what actually works, crowd-sourced" rather than what the official website claims. One of five named ideas consolidated into the working service below.

Verified, specific, crowd-sourced compliments — "people rated the garden outside your building..." — not generic flattery. One of five named ideas consolidated into the working service below.

Mine old accounts, photos, and search history — with permission — to resurface an abandoned passion, "not to sell you something." One of five named ideas consolidated into the working service below.

Independent of any cloud IDE — runs entirely against a local OpenAI-compatible model endpoint, on your own machine.

Explicitly documented as separate from Project-AI — doesn't import that repo, read its prompts, or share its governance stack.

A real, working local-first React app for creating, editing, annotating, arranging, and exporting PDFs, built on PDF.js and pdf-lib — with a working build and end-to-end test suite behind it.

A Tauri-based desktop application pitched as a real-time breaking-news command center with multi-channel monitoring and automatic escalation.

Not an application — a version-controlled definition of the engineering workstation itself: manifests, scripts, and validation to configure and reproduce the machine that hosts everything else on this page. Its own docs call it starter architecture only, with no production readiness claimed.

Twelve small, archetypal specialist services, each its own FastAPI microservice: Archimedes (engineering, first principles), Ares (the history of weapons, from stone knives onward), the Chronicler (world history by pattern, not dates), the Diplomat (negotiation and de-escalation), the Farmer, the Field Guide (wildlife tracking), the Geologist, the Herbalist, the Navigator, the Stargazer, the Surgeon (field medicine), and the Wrench (transport). A deliberately poetic, deliberately practical bet: knowledge that would matter if you had to rebuild from first principles.

A full-stack Next.js and FastAPI platform with Postgres/pgvector and Redis — real auth and RBAC, document embedding, a decision ledger, full-text search, and a provider-agnostic AI layer that defaults to a deterministic stub.

A validated texture-asset library spanning five quality tiers, pixel to 8K photorealistic — including the Miniature Office Pixel pack and the Project AI pack that supply those two projects' actual visual identity, not a standalone game-art product.

Evaluates actions against formally declared invariants before execution. Already named publicly on this site's own homepage ("github · ocee / octoreflex / tscg-b") but never explained anywhere — until now.