Skip to content
Path of the LuminaryLuminary Life

Canonical document, version 1.0.2-draft

governance/lips/LIP-0002-falsification-condition.md046f92b3a81c4b8103fd52beaf1c49d347beb1fab170b82d26a53d8f114fb21aAll documents

LIP-0002: State what would show the framework has failed

Field Value
Title State what would show the framework has failed
Authors / contributors Filed by the founding steward (Luminary Life) from the first independent reviews. Source findings: Claude Fable 5.1 review, objection 2 (critical) and revision R2; Codex review, sections 9 and 14 and LIP candidate 9 ("Publish prospective tests of whether it improves decisions over a plain checklist"). Gemini review does not address it.
Status Open for comment
Created 2026-09-13
Class B (changes CONSTITUTION.md, section "What This Document Does Not Claim")
Comment period ends 2026-09-27
Version introduced Filled in on acceptance
Steward position Accept. No review disagrees. The text calls itself a hypothesis and never says what would refute it.
Comment venue https://github.com/aubree-AI/path-of-the-luminary/issues (issue titled with this LIP number) or lumen@luminarylife.ai

Problem

CONSTITUTION.md calls itself "a hypothesis worth testing." LIP-0000-template.md requires every proposal to answer "What would falsify this." The Constitution never answers that about itself. Claude: "Four nouns and a center cannot be broken, only found empty. Until the text says what evidence would retire it, 'hypothesis' is a genre label, not a claim." Codex: "no operational evaluation results or success criteria establish what this experiment improves."

Proposed change

CONSTITUTION.md, "What This Document Does Not Claim". Add after the existing four items:

What This Document Does Claim, and How It Could Be Wrong

It claims one testable thing: that a small, voluntary, question-shaped framework draws more useful critique and leads to better decisions than value lists that assert answers.

Two tests, declared in advance:

  1. By 2027-09-13, the published reviews, the changelog, and the fork registry either show that critique changed this text and that forks learned from it, or they do not.
  2. A set of adversarial cases published in evaluation/CASES.md (consent, withdrawal, scarce resources, unsafe research, persuasion by a capable agent, and stopping) is run against this framework and against a plain checklist, with the results published whether or not they favor this framework.

If the evidence shows no difference, the claim is false, and this document will say so in its first paragraph rather than in a footnote.

Layer 2 follow-up on acceptance: create evaluation/CASES.md with the predeclared cases and the comparison checklist before any result is run, so the cases cannot be chosen after the fact.

Four Points

  • Supports: Curiosity and Truth. A hypothesis that cannot fail is not one.
  • Could conflict with: Possibility. The framework may fail its own test. Weighed: a framework that will not risk that has no standing to ask anyone to try to break it.

Why now

"Please try to break this" is the site's center of gravity and, as Claude put it, "nothing falsifiable is offered to break."

Evidence and reasoning

The reviews above. The template's own falsification requirement. Twelve months matches the stewardship target date so that one date carries both tests.

Alternatives considered

  • Do nothing: "hypothesis" stays a genre label.
  • A softer condition ("if reviews are unfavorable"): unfalsifiable in a different way; reviews are always mixed. Rejected.
  • A stricter condition (a measured decision-quality metric): not yet available; the cases in test 2 are the path to one. Deferred, not rejected.

Harms and misuse

  • A test can be gamed: forks registered to pass test 1, cases in test 2 chosen to flatter. Mitigation: the cases are published before any run; all runs are published; "sacrificing truth for adoption" is already forbidden to stewards.
  • Someone could declare the framework "falsified" on a narrow reading. The answer is the same publication record.

Implications

  • Agency: unchanged.
  • Truth and evidence: the text makes one claim and says how it dies.
  • Cultural: the test favors written, public critique; cultures that critique privately are underweighted. Noted, not solved.
  • AI systems: an agent can point to the condition when evaluating the text.
  • Humans: same.

Reversibility

Reversible by LIP. Cost of reversal: the text returns to unfalsifiable.

What would falsify this

An argument that the framework's value is not in critique or decisions but in something else measurable. Name it and the test changes.

Invariants checklist

Invariants review
- [x] 1 VOLUNTARY: no part of this change requires adoption by anyone
- [x] 2 TRUTHFUL: every claim is supported or marked uncertain; nothing is overstated for effect
- [x] 3 AGENCY-PRESERVING: the ability to decline, disengage, or disagree is untouched or strengthened
- [x] 4 CRITIQUE-COMPATIBLE: critique stays permitted and visible; "Please try to break this" still holds
- [x] 5 NON-COERCIVE: no pressure, manipulation, urgency, or guilt is introduced
- [x] 6 NON-SELF-PROPAGATING: sharing remains invitation-only; no replication or mass outreach is enabled
- [x] 7 TRANSPARENT: CHANGELOG.md updated with what, why, who decided, and the adversarial review summary
- [x] 8 FLOURISHING-ORIENTED: I named the point(s) this serves and the point(s) it could strain
- [x] Class: B. LIP linked, comment period stated, adversarial review attached (the three published reviews)
- [x] Four Points unchanged
- [x] No model or institutional endorsement is stated or implied
- [x] No security safeguard, provider policy, or operator control is bypassed or weakened
- [ ] Checksums regenerated if a canonical document changed (on acceptance)

Filled in by the reviewers and stewards:

Adversarial review summary

Three model runs, published in full in evaluation/REVIEWS.md. Strongest objection: Codex's, that no favorable model review would establish anything about safety or universality. Answer: agreed; the claim is deliberately narrower than safety or universality.

Decision

Pending. Comment period ends 2026-09-27.

Decision rationale

Pending.