Open-source reference architecture for auditable, replayable AI operations.
AlignTrue is an open-source reference architecture for supervised AI systems that need durable receipts, replayable state, governed actions, and fenced side effects.
Operational shape
How the project supports decisions, context, and execution.
Most AI apps treat safety as a UX problem: approvals, undo buttons, or better prompts. AlignTrue treats safety as an operability problem.
Before an agent can handle consequential work, the system needs to know what it saw, which policy or tool version shaped the behavior, what it attempted, what completed, and whether the action path can be replayed.
The project is intentionally experimental. Its value is as a working reference architecture for supervised AI execution, not as a claim that every interface is final.
Agent systems need to answer operational questions before they can safely do more than produce suggestions.
The system separates command intent, event facts, derived artifacts, action traces, and fenced external writes.
A practical architecture for AI systems where agents can act while actions remain auditable, replayable, and governable.
AlignTrue decomposes operational confidence into system primitives that can be inspected, tested, and reused selectively.
Capture what happened, what was attempted, and what evidence was used.
Rebuild state from append-only events and deterministic projections.
Track policies, prompts, tools, artifacts, and lineage.
Model actors, capabilities, approvals, holds, and outcomes.
Route external writes through safety classes, approvals, outbox handling, idempotency keys, and receipts.
Compare decision paths, inputs, policy versions, and outcomes instead of relying on subjective review.
Capture a decision clock with structured, hash-chained steps showing inputs, tool calls, policy checks, and outcomes.
Preview behavior with evidence-backed precedents before promoting policies or allowing more autonomy.
Public code, docs, architecture notes, examples, and project status are linked directly when they can be shared responsibly.
The accompanying architecture essay lays out the deeper argument: agent-native software is hard to operate unless it can be audited, reproduced, rolled back, and safely run at scale. AlignTrue is the concrete reference implementation behind that argument.
The repository exposes the architecture as code rather than a slide-only concept: kernel contracts, host runtime, connectors, packs, UI blocks, docs, and repeatable local development paths.
AlignTrue became a reference point for later public and private applied AI systems. Rather than copying the whole platform, later implementations reused the primitives that mattered for the operational context: safety classes, actor identity, deterministic hashing, intent lifecycles, approval rules, dry runs, and provenance tracking.
These are the practical workflows and constraints the project is built around.
Designed for workflows where source material, rules, approvals, and exceptions matter as much as the final output.
Keeps people close to important decisions while using AI to reduce searching, summarization, triage, and administrative drag.
Useful for casework, review queues, intake processes, research operations, and internal tools that need persistent context.
Operational AI needs more than a chat interface. Governed agent execution depends on replayable action paths, approval flows, egress fencing, provenance, and drift control. AlignTrue shows those concerns as architecture, not decoration.
AlignTrue is experimental beta software. Its value is not that every interface is final. Its value is as a working reference architecture for supervised AI execution: receipts, replayability, governed action flows, and deterministic operational patterns.
Different surfaces, same focus on context, coordination, review, and execution. This is a selected public view, not a complete project history.
GlassAlpha is an open-source toolkit for deterministic audit reports for tabular ML models, with fairness, explainability, calibration, robustness, and reproducibility metadata in one local workflow.
A private MVP build for approved-publication retrieval, voice query, and source-cited maintenance guidance on phone, web, and wearable surfaces.
Main Character helps turn complex product, workflow, and architecture problems into systems that work under real constraints.
