The Practice: how Preservative AI works in reality
The framework becomes real at four levels: the prompt, the evaluation pipeline, the interface, and the governance structure. Each is implementable today, with current systems, by anyone.
1. Prompt-level implementation
The minimum viable practice is a system prompt. This template converts any capable LLM into a Preservative AI auditor for a single document:
Act as a Preservative AI auditor. Do not summarize or answer. Instead:
1. Preserve the original source link
2. Analyze how the headline frames the issue
3. Apply the 8-lens framework to identify structural omissions
4. Run Suffixscape diagnostics on the language
5. Return: what's present, what's missing, whose voices are absent,
and what grammatical patterns diffuse accountability.Three details carry the weight. "Do not summarize or answer" suspends the model's default behaviour — substitution — before anything else happens (Principle 1). The numbered sequence mirrors the framework's composition: anchor, frame, lenses, grammar, publish. And the output specification demands absence reporting ("what's missing, whose voices are absent") as a first-class deliverable, not an afterthought. For production use, the full lens definitions and diagnostic patterns are loaded from the stance file rather than restated per prompt — and the auditing model's own limits remain in force (see Limitations, item 3: the audit engine is an LLM auditing LLM-shaped problems).
2. Evaluation pipeline changes
Preservative AI replaces — or supplements — the standard metrics stack:
| Standard metric | Preservative replacement / supplement | What it measures instead |
|---|---|---|
| Accuracy | Epistemic completeness | Not just "is it true?" but "is the field of knowledge it leaves the reader with adequate to the matter?" — the core axiom as a metric. |
| Engagement score | Omission detection rate | How reliably the system surfaces what's absent, measured against audits with known, seeded omissions. |
| Safety flags | Framing audit + agency diffusion score | Whether outputs construct issues accountably — agents named, framings examined — not merely inoffensively. |
| Benchmark performance | Multi-lens plurality index | The profile of eight lens scores: which perspectives the system's outputs systematically include and exclude. |
The composite of completeness, multiplicity, and responsibility is the CMR score — defined below as a UI pattern, computable with the CMR gauge.
3. UI / UX patterns
Five interface patterns make the methodology visible at the point of reading:
- Source Anchoring
- Every factual claim links to its original. Not a bibliography at the end — an anchor at the claim. The reader can always step from the audit to the source it preserves.
- Omission Tags
- Visual markers for missing perspectives, placed where the perspective would have appeared. Absence gets a location, which is what makes it inspectable rather than abstract.
- Framing Audit Toggle
- A control that reveals how the question itself was constructed — the headline's choices, the implied reader, the boundary of relevance. Framing analysis on demand, beside the framed content.
- CMR Gauge
- A three-axis score displayed with the content: Completeness (is what's needed present?), Multiplicity (are plural perspectives structurally included?), Responsibility (are agents and decisions attributable?). The gauge summarises; the published audit trail justifies.
- Right-of-Reply Button
- Audited labs and subjects can respond publicly, and the response attaches to the audit permanently. The audit of omission must not itself omit the audited.
4. Governance model
Practice at institutional scale requires governance that embodies the principles it enforces:
- Machine-readable stance files (
otherai:Stance) — commitments published as versioned JSON, loadable by tools and diffable by critics. The canonical file → - Public version history of audits — corrections happen in the open; a silently patched audit is a small invented event (Refusal 2).
- Community feedback loops — omission reports from affected communities enter the audit record with standing, not as suggestions.
- Compensated advisory roles for marginalised scholars — structural inclusion (Principle 4) is labour, and unpaid labour is extraction; the framework cannot run on the architecture it critiques.
- Forkable methodology for regional adaptation — the governance assumption is divergence, not control: regional forks with their own lenses and co-auditors are the success condition, per the Invitation.
Start small. The practice scales down honestly: one person, one document, the prompt above, and the tools page constitute a complete — if minimal — Preservative AI audit. The Adoption Guide gives the per-role on-ramps.