Everything you need to evaluate, integrate, and operate.
Organized by what you're trying to do — not by where the files live.
Security, compliance, and risk officers evaluating steadfaste for procurement.
Trust Properties
Fail-closed, separated authority, no direct bypass, narrow authorization, replay protection, tamper-evident evidence, model independence, and explicit human ownership — each stated, tested, and tracked.
Component Interface Map
Every component, every interface, and every boundary — including what is built, what is designed, and what remains open.
Design Realized
How the three-layer architecture (Constitution, Judgment, Wire, Interface) works and why each decision was made.
Cross-Agent Collusion
Risk model and mitigations for coordinated attempts by multiple agents to bypass controls.
Enforcement Boundary
How external policies (sandboxes, CI gates, policy servers) plug in without becoming competing authorities.
Independent Evaluator Review
How third-party auditors can review the system, access evidence, and verify claims independently.
Engineers wiring agents, surfaces, and enforcement into steadfaste.
Wire Protocol Specification
The framed JSONL protocol: 18 verbs, version negotiation, handshake, platform caps, and the additive-only rule that keeps old workers working.
Agent Worker Registration
How an agent declares itself, proves identity, and receives capability-scoped credentials.
Pi Integration
Reference adapter for the Pi coding-agent harness — spawn, frame I/O, lifecycle state machine.
integrations/pi/Langfuse Integration
Joining steadfaste's governance records to Langfuse observability traces by run ID.
Gateway Adapter: Telegram
How messaging transports are added additively without touching the core.
JEV Integration
How the Judgment External Validation (JEV) client plugs into the core for S1 scoring and circuit-breaking.
DevOps, security, and platform teams running steadfaste day-to-day.
Local vs. Cloud Comparison
Tradeoffs between on-prem, self-managed, and cloud-hosted deployments — latency, sovereignty, scaling, and cost.
Core Interface Slices
Build order and acceptance criteria for Web UI, TUI, messaging gateway, and operator dashboard surfaces.
Emergency Stop Enforcement
How stop signals propagate — SIGSTOP, SIGKILL, egress termination — and how operators retain override capability.
Log Survivor & Ledger
How records survive restarts, tamper-evident chain integrity is maintained, and operators verify state.
Coding Cost Routing
Classifier-based routing to control spend: destructive-gate, cheap-reads, verbosity binding, cost-posture.
Dry Run & Payload Integrity
Preview changes before they execute, with hash-bound evidence that what was approved is what ran.
Researchers and engineers adapting the constitution, judgment model, or core behavior.
Build Tasks
Step-by-step build path. Each task is one file + one test, ordered by dependency. Start at T0.1.1.
Core Build Blueprint
The build-ready v1 spec — Wire ABI, Ledger, crates, conformance suite, scope, and build order.
Story Slices
The story map — what exists (Stories 0–7 built and tested), what's next, and dependencies.
Design Gaps
What's still open — unresolved questions and decisions before the next build phase.
S1 Training Playbook
Operating manual: training, evaluation, constitution alignment, selection discipline on OOD metrics.
Judgment Abstraction
How the S1 model plugs into the core — interface contracts, versioning, and independence guarantees.
Governance by design
steadfaste is built so that the evidence needed for AI governance is produced naturally — not assembled after the fact.
Action Authorization Record
Every decision is bound to a policy revision, a steward identity, and a payload hash. What was approved is exactly what ran.
Budget-Aware Loop
Spend controls that prevent agents from exceeding configured limits without explicit human override.
Model Reputation Library
Tracks which models perform well for specific work — and where they should not be trusted.
New Model Trust Baseline
How new models are admitted, tested, and promoted through graded trust tiers before being cleared for production work.
Validation & Auditor Separation
Internal validation checks are structurally separated from auditor-facing evidence — no self-review.
Covenant Vocabulary
Shared language for trust, responsibility, and accountability across the system — from specification to UI.
Ready to evaluate?
Read the docs or request walkthrough access.