ProofOfWorks' concept lab (stuff that needs to get out of our heads) — a monorepo of inventions: sparred, prototyped, verdicted.
Every concept here follows the same lifecycle:
- Spar — the idea is attacked before it is built (peer-debated, prior art mapped, weakest leg named).
- Prototype — the smallest complete instance that can prove or kill the thesis, built from scratch.
- Verdict — an honest call: what held, what didn't, what it's worth, what to stop doing.
Nothing in this repo is a product. Status is explicit; verdicts include the negative parts.
AI is becoming cheap, fluent, and unaccountable. Our concepts cluster around making it provable: verified delivery of work, verified price of cognition, verified memory of facts. Language at the edges, verification at the core.
| Concept | One-liner | Status | Verdict |
|---|---|---|---|
concepts/ultrabrain |
Self-learning local agent kernel: LLMs propose, verifiers gate, ledgers remember, skills improve behavior, Datalog answers with proofs | Concept — working kernel (WIP) | Architecture proved; the gate + ledger + proofs are the core, now extended with experience streams, skill memory, and teacher-gated training candidates — headline cost-per-solved-task win not yet measured against a real fine-tune |
concepts/cog |
The COG — deflation-native unit of account for AI contracts: 1 cog = depth-verified price of a frozen workload at frozen capability; daily ssh-signed fix, exam-gated qualification, MCP server, hybrid contract riders | Concept — operating prototype (fix published daily since 2026-06-09) | Unit sound, 2× novelty-verified (nearest prior art: arXiv "Standard Inference Token"); hybrid fixed-USD+cog is the adoption shape; open: real exam runs, receipted buys, public hosting |
concepts/actweave |
Record, replay, and govern Vercel AI SDK agents in CI — no API keys: runs recorded at the LanguageModelV3 boundary into JSONL evidence ledgers, replayed through the real agent loop; strict request-hashing turns prompt/tool/tool-result drift into named failures; governance (allowlists, budgets, approval gates) leaves audit evidence CI can assert | Concept — working prototype (71 tests green against real ai@6; keyless runnable example) |
v1 — a fifth TS agent framework with testing attached — sparred to death (toy runtime, untested adapters, prompt changes passed green): DOA. v2 pivot to testing-layer-only held: replay-through-the-real-loop works, drift detection inverts golden-fixture rot, governance-with-evidence is unoccupied (TS gap is real — the deterministic primitives live Python-side); open: dogfood fixture from a live provider, official Mastra support, first external user |
concepts/deliveryproof |
Receipts for delivered work: objective delivery checks produce signed release/refund evidence for rail-neutral settlement | Concept — working prototype | The did-it-happen half of the house thesis; complements COG's what-is-it-worth half with a core library, rail adapters, and nested demo app |
concepts/<name>/ self-contained prototype: README, whitepaper, code, docs/
ThinkTank is a git/docs monorepo. Package-manager workspaces, when a concept needs
them, live inside that concept directory; for example concepts/deliveryproof is
its own npm-workspaces package with concept/ and adapters/*.
A concept graduates out when something earns its own repo; the write-up stays as the trail.