Evidence-backed agent systems and deterministic substrates.
Aligning agent execution with charter specifications, grounded in logbook evidence.
Four surfaces of the engine room. Labs (local models, simulation, Reflect) sit in their own section below.
The published evidence catalog covering grounded systems, orchestration architecture, and local LLM operator judgment.
Explore Research →The full catalog. Every domain agent and industry intelligence site built on the substrate — personal agents, enterprise verticals, and research demos.
Browse Domains →Geographic TSP scale ladders and solver-backed orchestration proofs showing where correctness must live.
View Benchmarks →Playable proofs of charter and boundary enforcement — browser games plus sensor/CAD demos, each with deterministic rules and replay.
Play the Demos →Bulkhead τ is the public release line for a deterministic expert system built as personal consultant tooling for complex implementations. The system does not attempt to expand model capabilities; instead, it enforces governance by bounding execution between a formal specification Charter above and an append-only evidence Logbook below.
Rather than relying on non-deterministic language models for runtime control, Bulkhead τ embeds those models inside a strict deterministic substrate. This makes governed operations auditable, bounded, and subject to explicit architectural boundaries.
Specification Standard. Evaluated against Product Behavior Contract specifications (.pbc.md), which define target behavior patterns and routing constraints. PBC is maintained by stewie-sh. We actively contribute provider-neutral anchoring guidance upstream and canonically consume these specifications downstream. Every architectural and operational decision routes through the contract question: Does this behavior fit the charter?
Core Substrate. A custom, deterministic expert system built locally for implementation engagements. It uses the open-source tau-bench grammar (domain, policy/SOPs, tools, tasks, database, variations) from Sierra Research (MIT License) on personal and consulting data. Runtime is a pattern-matcher over grounded tools, not an LLM talking to a simulated user. Supervisor/implementer rotation, phase discipline, and the four-phase framework (Glass Box → Human-in-the-Loop → Progressive Disclosure → Multi-Modal Output) sit on that grammar.
What tau-bench is, and is not here. Sierra’s τ-bench evaluates a model as the agent: tool calls, a simulated user, written policy, graded against database state, scored for reliability (Pass^k) on domains like retail and airline. Bulkhead τ keeps that tool-and-policy substrate. It does not ship Sierra’s airline/retail eval as the product, and it does not treat the language model as the production runtime.
Evidence & Deliberation. Anchored in the intellectual lineage of the gt-kb codebase authored by Mike Palmeter (Remaker Digital). Bulkhead τ implements a parallel, consultant-scale logbook design that inherits (a) first-class priority for deliberation records, (b) the logbook vocabulary, and (c) strict append-only discipline. It implements a working trust-state database (SQLite trust.db in the .groundtruth_proto/ substrate) to log behavior assertions from automated validation runs (with the first production entry logged for tennis-agent serve speed regression graduation), enforcing downstream execution gates. It remains structurally separate from GroundTruth (the productized platform successor to the gt-kb lineage).
Operational safety is maintained by separating historical governance milestones from current evidence-derived states through an active trust loop:
stewie-sh.Provenance note: The PBC anchoring path is upstream-aligned through provider-neutral specification guidelines. The trust-state verification logic is a custom parallel implementation, inheriting its conceptual structure from the gt-kb design.
Supporting surfaces. They are not the engine room. They measure models, simulate sensors, and generate owner manuals.
The measured field. Chip-to-chip results, what a spilled layer costs, and a running count of the measurement faults caught along the way.
Open Local Models →Firmware validators for future hardware and physics microscopes for legacy sensor forensic data.
Enter Simulation Lab →Owner's manuals. Repos run through Stewie Reflect — structured chapters plus flagged attention cards — then refined by hand against source.
Read the Manuals →