Devlog #2 — FFS
Annex L: On the Strategic Limitations of Large Language Models (Draft Guidance under Paragraph 213(b))
1. Purpose and Intent
1.1. This Framework sets out the general principles for engagement with Artificial Persons, hereafter referred to as "Language Models" (LMs), who, whilst articulate in tone and superficially coherent, lack Material Consideration.
1.2. Applications seeking to derive Reasoned Judgement solely from the output of LMs shall be considered premature, absent human presence or cognitive rigour.
2. Presumption in Favour of Making It Up As You Go Along
2.1. Unless otherwise stated, LMs are to be presumed unreliable except in cases of light correspondence, poetic rendering, or GPT-generated committee minutes.
2.2. Where an LM claims to understand the UK Planning System, weight shall be given to this assertion only where (a) cross-referenced with actual policy; and (b) not hallucinated from an American zoning blog.
3. Five Tests for LM Soundness
3.1. To be considered sound, an LM must demonstrate:
- (a) Consistency across runs (test of Repeatability);
- (b) Verifiable citations (test of Provenance);
- © No spontaneous reinvention of legislation (test of Restraint);
- (d) Recognition that “sustainable development” is not a vibes-based concept (test of NPPF Literacy);
- (e) Willingness to defer to officers with eyes and experience (test of Humility).
3.2. Failure to meet the above shall trigger Paragraph 301: "Thou shalt not automate thy planning officer without first consulting reality."
4. Monitoring and Review
4.1. This Framework will be reviewed every time an LLM starts to sound too confident.
4.2. Stakeholders are invited to respond in the form of bullet points, rebuttals, or interpretive dance.
🜃 Endorsed by the Secretary of State for Overengineered Tools. 🜂 Subject to Clarification. 🜁 Token budget exceeded. 🜄 Refresh the page to try again.