Active development
In progress. Real code, not yet shown publicly anywhere else.
Design Bench · Holding AGI accountable
Deterministic training & epoch authority
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.
Active developmentinternal designation — The State Register
In more detail
1,764 rule-graph records (SATISFIED / VIOLATED / UNKNOWN / CONFLICTED), a resumable signed-checkpoint epoch controller, deterministic curriculum splitting, and contamination gates that reject data crossing between train and held-out sets. It candidly discloses a real hardware limit: no CUDA device is available on its current machine, so accelerated fine-tuning claims are explicitly disclaimed rather than implied.
A companion engineering directive for this system lays out, in unusual mechanical detail, how a genuine model failure is supposed to be distinguished from a mere measurement artifact (a truncated generation, a bad tokenizer bound) before either is allowed to become training data — and it hard-bars the model from ever approving or promoting its own correction batch. A second, independent package applies formal methods on top of that: TLA+ specifications for the training cycle, the permit ledger, artifact construction order, and authority assignment, model-checked by an executable Python mirror that currently reports every invariant it checks as passing (a 43-state cycle model; a 489-state, 1,156-transition permit-ledger model).
Concrete specifics
- ·1,764 rule-graph records across four disposition states
- ·Explicit hardware-limit disclosure: no CUDA device, QLoRA claims disclaimed accordingly
- ·A hard rule that a training candidate cannot approve or promote its own correction batch
- ·An independent TLA+ formal-verification package: 43-state cycle model, 489-state permit-ledger model, structural lint passing
Directly related
Constitutional training corpus
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.
Sovereign AI IDE runtime
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.
Go deeper