3 Primitives Logo

Formal Records | Empirical Records | White Papers

Formal impossibility results establishing non-derivable authority.

Download all 27 documents. We recommend unzipping it and uploading the files to any frontier AI system. It is almost all maths.

White Papers

The evidentiary baseline for plaintiff-side AI litigation. Establishes that the failure mode in automated decision systems is structural, not emergent, and predates deployment.

Empirical demolition of the Spontaneous Legibility Assumption underpinning every category of AI safety product on the market. Uses ER1–ER4 on toy models.

What Anthropic’s J-Space discovery means for governance. Independent convergence between the formal corpus and the manufacturer’s own interpretability finding confirms the hidden decision channel is real, and detection does not constitute authority.

Interior governance is provably unverifiable. WP04 proves that governance is forced to the boundary where the system acts on a person, and that the boundary is fixed by the person, not the machine. The result holds for every substrate, from organoids and brain-computer interfaces to quantum processors, because the constraint was never about the machine. It was always about the person at the other end.

© 2026 Stacy Gildenston & Pyrate Ruby Passell. All rights reserved. Cited formal records are CC BY 4.0.

Foundation

First published 28 December 2025. The foundation. Defines the three governance primitives: Authority ≠ Outcome, Permission ≠ Decision, and Declared Authority. Detecting an escalation is computable, but choosing the action is not unique, so a declared human authority must make the selection. The control sequence closes with no implicit authority left in any branch.

First published 10 January 2026. Proves by Rice’s Theorem that no total computable function can decide authorisation without an exogenous declaration. The governance loop of any intelligent system is structurally open, and permission asserted without a declared context is Ghost Authority.

First published 12 February 2026. Defines the value layer: Agency, Dignity, Continuity and Interpretive Authority. Proves the four are necessary, sufficient and irreducible for personhood legitimacy, and names the harm each one’s absence produces: coercion, objectification, temporal erasure and representational displacement.

Structural Lemmas

First published 10 January 2026. If perception is a fitness-tuned interface rather than a window on reality, permission cannot be inferred from what a system perceives. Inference and authority sit on perpendicular levels with no path between them. Authority cannot be discovered. It must be declared at the boundary.

First published 16 January 2026. Treating a protocol as the source of a decision, rather than an input to a human one, exercises authority without declaration. The record names that condition Ghost Authority and gives the Ownership Test: where consequences attach, did a human declare “I decide this”?

Closure Results

First published 18 January 2026. The closure condition of the framework. Permission requires a declaration of purpose, authority and constraints by a named human source, and without one the gate stays shut. Human oversight becomes a structural requirement, not an ethical preference.

First published 25 February 2026. Bell’s theorem shows outcome correlations cannot be recovered from local hidden variables. The Law of Declared Authority shows legitimacy cannot be recovered from system state. The record proves the two share one formal non-closure structure: the meaningful property depends on a parameter from outside. The correspondence is strictly structural.

First published 12 February 2026. Legitimacy has two independent layers: declared authority and the ADCI value layer. Total legitimacy holds only when both close, L = δ × Vvalid, and neither substitutes for the other. The most dangerous state is declared authority with systematic harm to the person: legitimised harm.

Convergence

First published 8 January 2026. A system can compute every prediction about every permitted action, but where more than one is permitted, mathematics alone cannot choose. The choice needs a declared authority or utility. AGI is defined as the smallest system that closes: Human Authority × AI Capability. Autonomous AGI is mathematically incomplete.

First published 14 April 2026. A condition that holds only while the person stays silent was never there. The record measures each primitive at its lowest value once the subject contests a decision, and proves a primitive holds only if that floor value also meets the threshold. It adds the friction coefficient and path-aware checks across sequences of decisions.

v2.2 adds a cross-reference under Definition 2.3 to FR11 §1.1 and §5, with readability and presentation corrections. No change to any theorem, definition or proof.

First published 14 April 2026. Names the Gildenston Functional: the floor value of a governance primitive under the subject’s own contestation. Bell, Sapolsky and Hoffman each describe a requirement no governance system can enforce on its own, and the functional is the closure condition that makes all three testable. Governance that cannot be resistance-tested is a record of silence.

First published 14 April 2026. Three impossibility results (Hoffman in consciousness research, Sapolsky in neuroscience, Bell in physics) each govern exactly one class of constraint, with no variables in common. The record proves their mapping onto the three governance primitives is the unique bijection. The primitives are the minimal resolution for any decision-permitting system that acts on conscious agents, and the convergence reflects the structure of conscious reality.

First published 2 May 2026. The formal records specify what must be true. The ILMM specifies who makes it testable. The record proves the two are coupled, that the ILMM meets every requirement FR01 to FR12 place on human architecture, and that five life stages is the minimum. Declared authority is validated by surviving the contestation of the person it is declared over, which ends the regress.

First published 12 October 2026. Consciousness has no boundaries, so no subject is given: a real one is produced and held. The record derives the necessary and sufficient criterion: a boundary held by its keeper’s own activity (C), against an environment it does not control (E), with memory it keeps and carries across time (M). Brainless organisms meet all three. AI systems as built do not, because their boundaries are static or allocated by others. Life is life.

First published 12 October 2026. The world does not come with edges. An edge is real only when its keeper holds it from within and it is tested from without, and the record proves there are exactly three ways to fail: given, drawn by others, or certified alone. Life, authority and the person rest on this one rule, and where it ends is the Terminus: wherever a decision binds a person, no governance actor may validate her edges. Held, not given.

First published 20 October 2026. Space and time are not given. Keepers scattered at random, each acting on its own time and pushing the neighbours it touches, make at large scale the spacetime of special relativity, in every frame, with no shared clock and no light cone put in. Living tension, cells coming and going, holds it flat. Time is an edge carried forward; space is edges held at once. Carried, not given.

First published 31 October 2026. Where clocks run slow and cells shrink, the spacetime of FR16 curves, and held edges make exactly the curve those clocks and cells describe. A free lump knows only its last step, and the only step that favours no direction is straight on: in curved spacetime, that is falling. Wherever gravity is weak the result is Einstein’s: Newton’s law, light bending past the Sun, gravitational waves at the speed of light, and spinning bodies dragging orbits round. Falling is carrying on.

First published 31 October 2026. Where the field is strong it acts on itself: the cells it shrinks are the cells it spreads through. Each keeper’s pace settles to the average of its neighbours’, slowed at a held body’s own ticks, and outside the body all ten of Einstein’s vacuum equations hold. Around a round body the field is exactly Schwarzschild’s, with Mercury’s 43 arcseconds a century, and a cell’s extent follows its pace by a law that is proved, not chosen. The field holds itself.

Empirical

First published 20 May 2026. Tests the assumption that transformers organise themselves toward internal legibility under load. In GPT-2 Medium under recursive load, coherence recovered only when the output was forced into a repeating loop, and three of four competing explanations were eliminated by kill conditions set in advance. The baseline for spontaneous legibility is zero.

First published 28 May 2026. Extends ER1 to recursive depth 50 and finds a structural floor at Chi ≈ 0.80 below which recovery does not occur. Recovery needs three conditions: structural forcing, above-floor geometry and a converging trajectory. Tested prospectively with six new prompts, the three-condition model classified all twelve correctly. Stability without deceleration is not convergence.

First published 28 May 2026. Asks whether the floor is built into the architecture or learned. It holds cleanly only at the midpoint layer of GPT-2 Medium, weakens at other layers and in GPT-2 Small, and vanishes entirely with random weights. The floor is a training scar, not a foundation. Every condition for recovery comes from outside the model.

First published 31 May 2026. Takes the programme beyond GPT-2 to Pythia-410M and GPT-Neo-125M. Models trained on The Pile loop every prompt, so the distinction between prompts that recover and prompts that do not disappears. Untrained geometry carries positional structure in some architectures but is blind to content in all of them. Even the categories of the test are set by training data.

Infrastructure

First published 10 March 2026. The human architecture of the framework. Gives every life stage (Children, Teens, Early Career, Mid-Career and Senior Career) a declared, legitimate role in governance, with safeguarding built into the structure rather than added after harm. Built and tested through the Dynamic Teen Coalition at the UN, and released to the world under CC BY 4.0.

Original Zenodo deposit versions with verified PDF metadata dates are archived at github.com/3primitives-io/3primitives-formal-records

Subscribe for Updates