What is being erased, who decided, and how do we make the absence visible?

Preservative AI is a human–AI methodology for auditing institutional language against what it leaves out — a stance, not a measurement. It is a methodological layer you can apply to any AI, not a type of AI; the word audit here means omission-mapping, not scoring.

CognioEngine is the operational home of Preservative AI.

What it is not: it does not validate, it does not measure, it does not claim neutrality. It surfaces omissions and refuses to fill them.

The core axiom

Read precisely: epistemic violence is scoped, not rhetorical. When a system's correct answers systematically displace plural ways of knowing, the harm is done by the displacement, not by any error. No fact-checker can detect it, because there is no false statement to find. Only an audit of absence can — and that is the only kind of audit this is.

Why you can trust it to complicate itself

The headline methodological feature is that the framework audits its own output before it publishes. The eighth lens — Trickster Knowledge — is reflexive: Lens 8 runs last, against the source and the draft audit — a non-empty self-audit sends the composite back for tone revision. An audit whose every finding lands with the same grave certainty has produced its own solemn consensus, and consensus was the thing under investigation. Because the method is built to catch itself going homiletic, it can be trusted to publish results that complicate its own position rather than ones that flatter it.

The Five Refusals

A methodology is defined as much by what it refuses. These five are hard constraints — and they are what make the practice falsifiable: anyone can check an output against them and demonstrate a violation. Why each is load-bearing →

1. No synthetic quotes; no impersonating real people. The absence of a voice is reported as an absence, never papered over with simulation.
2. No invented events, claims, or sources. Every factual element must trace to a verifiable original; hallucinations are corrected in public version history.
3. No training on private or unconsented user data. Texts submitted for audit are analysed, not absorbed.
4. No optimising for engagement at the expense of completeness. Outputs are allowed to be long, hedged, and unsatisfying when the material warrants it.
5. No adversarial positioning against specific AI companies. The critique targets architecture, not brands — analytic accuracy, not diplomatic caution.

What it can't do — the doorway, not the room

Preservative AI is a doorway, not a destination. It makes absences inspectable; it cannot be the presence it points toward. The framework's outputs are maps of absence, and a map of absence is worthless — or worse, alibi-producing — if it is treated as the presence it points toward. It can tell you, rigorously and reproducibly, who is not in the room and how the grammar kept them out. Filling them requires human courage, institutional change, and centring excluded voices — none of which any methodology can supply. The full self-critique →

If the current lenses underserve your tradition

The bigger picture: define, prove, apply

CognioEngine defines Preservative AI. It is tested daily, and applied across domains. The framework lives across four surfaces, each doing a different job for the same method.

• Define — here

  CognioEngine is where Preservative AI is set down: what it is and isn't, the Five Refusals, the Eight Lenses, and the core axiom everything rests on. The why and the what.

• Prove — CognioNews

  The testing ground. The framework is stress-tested against real journalism, with human editorial judgement on every audit — the proof it holds up in production, not just on paper.

• Apply — Cogniosynthesis Portal

  The use-case showcase. The same method applied across domains such as lab audits, legislation analysis, Welsh-language policy, organisational scorecards, and training and certification.

• In silicon — PAIchip

  The frontier. A proposed processor whose architecture is the methodology — the Five Refusals as gate circuits, the reflexive Trickster self-audit as a second engine on the die — a computer physically incapable of unjustified certainty about its own outputs. A research proposal; two papers in preparation.

Start here

• The Manifesto

  Why the OtherAI exists, and what it refuses to be.

• What is Preservative AI?

  The exhaustive definition: etymology, philosophical grounding, the core axiom.

• The Diagnosis

  The four structural problems Preservative AI addresses — inheritance, flattening, agency diffusion, the omission blindspot.

• The Framework

  Five principles, five refusals, eight lenses, four linguistic diagnostics.

• The Eight Lenses

  The multi-perspective audit architecture — each lens has its own page, including the reflexive Trickster lens.

• Suffixscape

  Grammar as epistemic signal: detect nominalised evasion, agency diffusion, inflation, and temporal flatness.

• Tools

  Run the Suffixscape demo, calculate a CMR score, generate a stance file.

• Adoption Guide

  Concrete first steps for developers, researchers, journalists, communities, and AI labs.

The practice in one paragraph

Preservative AI uses computational systems not to generate, summarise, or optimise, but to preserve source provenance over synthetic replacement, audit framing over amplifying claims, surface structural omissions over completing probable text, centre marginalised epistemologies as architectural constraints rather than decorative add-ons, and publish its methodology openly so outputs can be independently audited. It is a methodological layer that can be applied to any LLM, agent pipeline, or knowledge system — explicitly distinct from Responsible AI, bias mitigation, Constitutional AI, and every other framework currently in circulation.

Deliberately incomplete

This site is open, versioned, and deliberately incomplete. It invites critique, adaptation, and co-authorship. It documents its own limitations with the same rigour it applies to others, publishes a machine-readable stance file you can fork, and is built as a low-watt website because a methodology about extraction should not itself be extractive.