Normative AI Architectures · Vol. I

Verified Normative Agents.
Defined, not described.

Each pack ships the specification — and the evidence that it can be built from the pack alone, across paradigms.

What they're for

The gaps they close.

Each pack closes one precise gap — the kind that becomes critical the moment a decision is reviewed, challenged, or audited. The Session State Continuity & Anchor Replay family, and its five components:

“When you copy your work into the next session, the words come with you. The thinking does not.”
Session State Continuity & Anchor Replay Family · Designed
“Restoring saved state without verifying it first is indistinguishable from restoring a compromise.”
Verified Rehydration Agent · Released
“When AI systems cannot explain which claims they trusted, every consequential decision is indefensible.”
Truth Arbitration Agent · Designed
“If you cannot reconstruct how a decision was made, you cannot prove it was made correctly.”
Audit Replay Agent · Designed
“Everything dangerous that an AI system has ever done entered through an interface no one was guarding.”
Boundary Defense Agent · Designed
“When state moves between devices, the only thing worse than losing it is accepting an old version as current.”
Freshness Portability Agent · Designed

Released agents are in the catalog. Explore the agents →

What's inside a pack

The specification — and the evidence that it can be built.

Part A · the agent

The normative core

The complete, execution-neutral specification your team builds from — every design decision about the agent's logic already made.

Normative
+
Part B · the evidence

Evidence of derivability

The same agent, rebuilt from the pack alone in five paradigms — Python, Rust, Prolog, C and Bash — run live, every binary hash-attested. Included as evidence.

Observational · non-normative

One manifest and one license bind both parts. See what the VRA package contains →

What's different

What defines these agents.

Each agent is a precise normative object — it states exactly what it decides, on what authority, with what evidence, and how it fails.

01

Defined, not described.

Each agent is written in a semantically closed normative language — verdicts, validators, gates, a state machine — not prose. Every agent shares this architecture but speaks its own dialect, fit to its problem. Interpretation is constrained by the specification, its tests, and its evidence contract.

02

Derivable, with evidence.

For the Verified Rehydration Agent, five paradigm-isolated reconstructions built the agent from the pack alone and converged on all 29 conformance records — 27 PASS, 2 SKIP. Supplementary renderings, from a spreadsheet to Brainfuck and AWK, reproduce the same results on the surfaces they cover. Each released pack ships a verification runbook (verify.md) that has you reproduce its declared values on your own machine. See the evidence →

03

One source of truth.

Because the specification is execution-neutral, developers, testers, security, audit and ops all work from one artifact — no manual translations between them.

04

Fail-closed by design, audit-ready by artifact.

Each agent is specified to fail closed on any critical failure — never a silent pass — and to emit, for every decision, evidence that can be verified offline, long after it was made.

Want the formal definition? See the ten properties of a Verified Normative Agent →

★ Released · free of charge

The Verified Rehydration Agent — public release v1.

The complete two-part pack — the specification with its semantic indices, Conformance Tests and whitepaper, plus the reconstruction evidence: five-paradigm baselines, observation reports, hash-attested binaries — under one manifest and one license. It is distributed free of charge; the license is accepted before download, in two confirmations. It ships its own verification runbook (verify.md), which has you reproduce the pack's declared values on your own machine.

The model

The whitepaper shows the problem. The specification delivers the solution.

The whitepaper is free to read — it names the problem, the failure modes, and the guarantees. The specification is the ready-to-build, execution-neutral document you hand to an engineering team.

The whitepaper

Problem · Outcomes · Domains

The complete framing of the gap and the structural guarantees the specification defines.

Non-normative
→
The specification

Ready-to-build · Complete

Agent identity and boundaries, external contracts, full internal logic, evidence obligations, failure handling, acceptance criteria.

Normative

See exactly what arrives in a pack — the Verified Rehydration Agent →

The library

Component agents, composed into families.

A component agent closes exactly one capability. A family composes several component agents under a dedicated integrator — the family-level capability emerges only from the whole.

35 component agents·25 families, each under its own integrator·1 released
Family · Designed

Session State Continuity & Anchor Replay

“When you copy your work into the next session, the words come with you. The thinking does not.”

The Session State Continuity & Anchor Replay Family is a specification for governing the epistemic continuity of long-running AI work across session boundaries, including across changes of model. The family does not produce content, generate summaries, or decide what the prior work should have said. It governs whether the next session inherits the structure of the prior session’s thinking — not just its words.

Verified Rehydration Agent · Released Truth Arbitration Agent · Designed Audit Replay Agent · Designed Boundary Defense Agent · Designed Freshness Portability Agent · Designed Family integrator · Designed