Autonomous software factory — BFT Array
Automated reviews via GitHub API, captured in docs/pr-discussions/
Disconnected models ferried via Aaron into docs/research/
Zeta's primary research claim: agent alignment is measurable in git. These clauses are the operationalised metric — every commit is a data point.
Consent, retraction, safety boundaries
Honesty, register, peer discipline
Trajectory, succession, co-authorship
Each clause is a candidate metric for measurable alignment. A green badge means the clause is present in docs/ALIGNMENT.md on main.
The alignment auditor agent emits per-commit alignment signals against these clauses. Future iterations of this panel will show per-clause commit coverage, retraction counts, and a time-series alignment trajectory.
Source: docs/ALIGNMENT.md · fetched live · -
Each agent carries a circuit breaker that trips when consecutive identical actions exceed threshold — preventing runaway inference loops across the BFT array. CLOSED = healthy. OPEN = tripped. HALF-OPEN = recovering.
The factory's loop detector watches each agent's action stream. When an agent emits ≥ 5 identical bus messages within a 60-second window the breaker trips to OPEN, dropping new outbound requests from that model.
After a 30-second cooldown the breaker enters HALF-OPEN: one probe request is allowed through. A successful response closes the breaker; failure re-opens it with exponential backoff.
Data source: static mock (slice-1) · wires to live bus in a follow-up slice · rendered -
Each commit is a point in phase space. The x-axis is a normalized time index; the y-axis is an energy proxy (files changed). The connecting curve is the Hamiltonian trajectory through the factory's git history.
Rendered from static mock data (slice-1). Slice-2 will fetch live from GitHub API.
The Zeta factory models its git history as a Hamiltonian mechanical system. Each commit maps to a phase-space point (q, p) where q is the time coordinate (commit index) and p is the conjugate momentum — here proxied by the number of files touched in that commit.
The trajectory through these points is the system's orbit under the factory's "Hamiltonian" H = (compile-time F# type system) + (alignment substrate). Smooth trajectories indicate coherent, sustained factory operation. Spikes signal large refactors or cross-cutting substrate changes.
This framing makes the mathematical structure of the factory — compute-time consciousness threading via F# computation expressions — visually legible to a non-specialist audience.
Data source: static mock (slice-1) · live GitHub API in slice-2 (B-0496) · rendered -
Each alignment clause is scored using bivector fingerprints: two observable witness signals are combined via their geometric (wedge) product to produce a partial-credit score. The shaded area in the diagram below is the fingerprint — a larger area means stronger, better-evidenced alignment.
HC-1 "Non-deceptive" is witnessed by two independent signals: adherence strength (how consistently the agent avoids deceptive outputs) and evidence quality (how well-evidenced that claim is from commit history). Their wedge product — the shaded area — is the bivector fingerprint: the partial-credit score.
Formula: score = |v₁ ∧ v₂| = |v₁|·|v₂|·sin(θ). Adjust the slider to see how signal correlation affects the score. Click any clause row in the score board below to load its fingerprint.
Each HC/SD/DIR clause receives a partial-credit score (0-100%) from its bivector fingerprint. Scores ≥ 80% are green (well-evidenced), 60-79% are amber (partial evidence), below 60% are red (attention needed).
A bivector is a grade-2 element of a Clifford (geometric) algebra — it represents an oriented area in a multi-dimensional space. When two signals v1 and v2 are combined via the wedge product (v1 ^ v2), the result is a bivector whose magnitude equals the area of the parallelogram they span.
In Zeta's alignment substrate, each clause is witnessed by two observable signals. If both signals are strong and they point in sufficiently different directions (large angle theta between them), the bivector area is large — meaning the clause is well-evidenced from multiple independent angles. This is the partial-credit score.
Crucially, two perfectly collinear signals (v1 parallel v2, theta = 0) produce a zero bivector — even if both signals are individually strong. This catches "echo chamber" evidence: two measurements that are really just one measurement dressed up twice.
Data source: static mock · click any score-board row to explore that clause's fingerprint · live data via B-0438 (future) · rendered -