The Framework › The Five Operational Principles

The Five Operational Principles

Each principle names what it means in practice and what it changes about how computational systems are used. Together they invert the default posture of generative AI: from replacement to preservation, from amplification to audit.

1. Preserve sources; never replace them

The default behaviour of generative systems is substitution: a synthesis appears in place of the sources it digested, and over time readers encounter only the synthesis. The original — with its texture, hedges, context, and author — drops out of circulation. Principle 1 makes the original structurally unremovable: every claim in an audit anchors to the canonical source, and the audit is legible only alongside the original, never instead of it. In UI terms this is source anchoring; in editorial terms it means a Preservative AI output is a companion document, not a replacement document.

2. Audit framing; do not amplify it

Restating a claim — even to fact-check it — repeats and reinforces its framing: the choice of subject, the implied question, the boundary of relevance. Principle 2 redirects analysis to the construction itself: Why this headline? Why this protagonist? What question does the piece train the reader to ask, and which questions does that displace? A framing audit can find a piece entirely accurate and still show that its construction forecloses the inquiry that mattered. This is the principle that most sharply separates Preservative AI from summarisation tools, which are framing-amplifiers by design.

3. Surface omissions, not repetitions

Most analysis adds emphasis to what is already present — it repeats, ranks, and excerpts. Principle 3 inverts the attention: the deliverable is the map of what is not there. An omission report names the missing perspective, explains why it was reasonably expected (this is what keeps the practice from degenerating into "you didn't mention everything"), and where possible points to where the absent voice can actually be heard. The eight lenses exist to make this systematic rather than anecdotal: each lens is a theory of what should have been present.

4. Structurally include marginalised voices

The word structurally is the load-bearing one. Inclusion-as-add-on means consulting marginalised perspectives when someone remembers to; inclusion-as-architecture means the evaluation literally cannot complete without them — an audit with the Indigenous Knowledge or Marginalised Voices lens skipped is an invalid audit, the way a build with failing tests is a failed build. This converts plurality from a value (aspirational, deferrable) into a constraint (mechanical, checkable). It also implies obligations the framework takes seriously: structural inclusion of perspectives requires compensated participation of the people who hold them, not extraction of their critique.

5. Publish the method openly

An audit you cannot inspect is an oracle, and Preservative AI exists to refuse oracles. Principle 5 requires that the entire apparatus — lens definitions, scoring rubrics, prompt templates, diagnostic patterns — be published in forms that machines can load and humans can diff: stance files, versioned repositories, public score histories. The test is reproducibility-of-disagreement: a critic must be able to re-run the method, get a different score, and show exactly where the judgements diverged. Closed methodology would reproduce the central problem — editorial choices hidden inside a system that presents its outputs as findings.

The principles as a system

The five principles are ordered as a workflow: anchor the source (1), audit its construction (2), map its absences (3), through plural instruments (4), and publish how you did it (5). Remove any one and the practice degrades predictably — without 1 it becomes summarisation, without 2 amplification, without 3 repetition, without 4 monoculture, without 5 oracle. The five refusals are the matching negative space: what the practice must never do while executing these.