Barnes.Projects

Philosophy/ Causal Epistemology/ Aug 2026

Name, Chain, Measure: an epistemology for pre-paradigmatic domains

The method behind TLCTC was never only about cyber. Name the causes, chain them, measure their spacing relationally — and model cognition becomes its second application.

Film · BP 01 · 7:26 — the three moves, the three papers behind them, and what crosses a layer (also on YouTube): Das Ende der semantischen Diffusion · Δt — Was ist Zeit? · The Pre-Paradigmatic State of AI

Abstract

Three earlier essays developed three instruments for one field: a shared taxonomy of causes to end cybersecurity’s semantic diffusion, a path notation for the causal links between those causes, and a relational reading of Δt as a ratio between event chains rather than a flow through absolute time. This essay makes the claim the three were always pointing at: taken together they are not a cybersecurity framework but an epistemology for any causal domain in a pre-paradigmatic state — name the causes, chain them, measure their spacing relationally. Cybersecurity was the first application. Model cognition is sitting there as the second, with its reference chains — tokens, layers, training steps — already discrete and fully inspectable: a better-behaved laboratory for the thesis than the physical world ever was.

Every framework builder eventually faces an uncomfortable question: is this a tool for one domain, or did the domain merely happen to be where the tool was forged? TLCTC was built for cybersecurity. Its ten clusters classify cyber threats by the generic vulnerability they exploit; its path notation chains those clusters into attack sequences; its Attack Velocity metric Δt measures the spacing between attacker steps. All of it smells of firewalls and incident reports.

But strip the domain content away and look at the moves. There are exactly three, and none of them mentions cyber.

The Method, Stated Domain-Independently

  1. Name the causes — build the semantic layer. A pre-paradigmatic field classifies by outcomes: “ransomware,” “data breach,” “phishing problem.” The first move replaces outcome labels with a stable taxonomy of causes — the Carnap move. Only when ten observers share the names of the causes can their observations be combined instead of diffusing semantically.
  2. Chain them — make the causal links explicit. Named causes are not a list; they are nodes. The second move writes the arrows: which cause enables which, in what order, under what conditions. The chain is the explanation. An outcome label compresses the chain away; the path notation preserves it.
  3. Measure relationally — Δt against a declared reference chain. Time is not a container events flow through; it is a ratio between two event chains — the chain under observation held against a conventional, regular reference chain. The Leibniz move. The reference must be declared, because it is a convention, not a discovery.
// The method, domain-independent NAME a taxonomy of causes, not outcomes CHAIN cause_i →[causal, Δt] cause_j MEASURE Δt on causal edges = ratio( observed chain : declared reference chain ) // First application — cyber (TLCTC V2.0) #9 →[Δt=24h] #4 →[Δt=10m] #1 // reference chain: SI second (9,192,631,770 caesium periods)

Nothing in those three moves knows what a firewall is. The method is an epistemology — a procedure for turning a fragmented, pre-paradigmatic field into one whose observations add up. Which raises the obvious question: where is the next field that needs it?

The Second Domain Is Already Waiting

Interpretability research — the attempt to understand what happens inside large neural networks — is today exactly where cybersecurity stood: brilliant practitioners, extraordinary instruments, and no shared grid. One lab says “feature,” another “direction,” a third “circuit,” “concept,” “polytope.” The public debate runs on outcome labels: the model “hallucinated,” “reasoned,” “deceived.” Ask what causal structure inside the computation those words name, and the answers diffuse — ten experts, twelve definitions. That is not a maturity problem waiting on more data. It is the pre-paradigmatic condition, symptom for symptom.

Method step Cybersecurity (first application) Model cognition (second application)
Outcome labels to replace “Ransomware,” “data breach,” “phishing problem” “Hallucination,” “reasoning,” “deception”
Semantic layer 10 clusters: generic vulnerabilities exploited Taxonomy of computational structures that produce capacities — still missing
Causal chain Attack path: #9 → #4 → #1 Circuit path: feature → feature → output token
Intervention that tests the chain Control placed between steps Ablation / activation patching between components
Reference chain for Δt SI second (caesium periods) Token positions, layers, training steps

The middle row is the point. Cybersecurity has its semantic layer; interpretability does not, yet — and every downstream dispute about whether models “really reason” inherits that missing layer, exactly as every incomparable threat report inherited cyber’s.

NAME · CHAIN · MEASURE causal taxonomy · causal paths · relational Δt APPLICATION I — CYBER #9 Δt=24h #4 Δt=10m #1 reference chain: SI second APPLICATION II — MODEL feature feature token Δt=12 layers Δt=+40 tokens reference chains: layers, token positions SAME THREE MOVES — DIFFERENT EVENT CHAINS the method never mentions the domain
Fig. 1 — One epistemology, two applications. The domains supply the event chains; the method supplies the discipline.

Causal Paths in a Forward Pass

The second move — chain the causes — already has its instruments in interpretability; what it lacks is the shared notation. Ablation and activation patching are interventions on nodes of a causal chain: silence this component, transplant that activation, and observe whether the output changes. This is Pearl’s do-operator applied to a forward pass, and it sets the standard of evidence the outcome labels never met.

The evidence standard

A cognitive claim about a model is cashed in only when it names a structure in the computation that demonstrably does the work:

  • Not: “the model planned the rhyme” — an outcome label.
  • But: a rhyme-candidate feature, active at the line break, whose ablation changes the line’s ending — a causal path with named nodes.
// Second application — a cognitive claim as a causal path f_rhyme@line_break →[Δt=+1…12 tokens] token_selection@line_end // test: patch f_rhyme → the planned ending changes with it // (Lindsey et al. 2025 — planned rhyme features in Claude)

Findings of this kind already exist: induction heads that implement in-context copying, a linear board-state representation inside a model trained only on Othello move sequences, rhyme candidates selected before the line that leads to them is written. Each is a chain with named nodes and tested arrows. What is missing is the first move — the shared semantic layer that would let these results be added together across labs instead of remaining brilliant, incompatible case studies. That is precisely the state cyber threat reporting was in before a common taxonomy of causes.

Δt Without a Clock

The third move is where the second application becomes philosophically interesting, because inside a transformer the relational thesis does not hold approximately — it holds exactly.

There is no t in the computation. No clock, no flow, no container. What exists are event chains: token positions along the sequence, layers along the depth, steps along the training run. “When” a feature acts is its position in those chains, and any Δt between two computational events is a ratio against a reference chain that must be declared — positions, layers, or steps. Wall-clock time re-enters only at the interface, as GPU throughput: one more reference event chain, and a swappable one. Run the same weights on faster hardware and every wall-clock Δt changes while every computational Δt stays fixed — a clean demonstration that the computation’s time was never wall-clock time at all.

“The universe in its entirety is like some huge molecule in a stationary state … the different possible configurations of this ‘monster molecule’ are the instants of time.” — Julian Barbour, The End of Time

Barbour has to argue Newton’s t out of physics — against intuition, against notation, against three centuries of habit. In a transformer there is nothing to argue away: the container was never there. A model is in this respect a cleaner demonstration of the Leibniz position than the physical world itself — a system that plainly exhibits ordered, causally structured behaviour, and just as plainly contains nothing but relations between events.

A Better-Behaved Laboratory

And the second application improves on the first in several respects that matter epistemically:

First application

The physical world

Continuous quantities, partially observable systems, unrepeatable incidents, adversaries who hide their event chains. The reference chain (caesium) is superbly regular — but the observed chain must be reconstructed forensically, after the fact.

Second application

The model

Discrete reference chains, a microstate fully in hand, and runs that can be repeated. In principle every node is open to intervention — ablated, patched, replayed — though each intervention is costly and partial (Typed Crossings, §5). The observed chain need not be reconstructed — it can be re-run.

Here lies the strange asymmetry of the field. For the model, the complete microscopic description is actually in hand — every weight, every operation. No biological system was ever transparent like that. And yet possession of the full mathematics has not delivered understanding, the way knowing every molecule’s position does not hand you the concept of temperature. We hold the microstate and lack the macro-vocabulary. Interpretability is, in effect, the search for the statistical mechanics of these systems — and statistical mechanics is exactly the kind of theory the three moves are built to produce: named macro-causes, explicit causal chains, relationally measured dynamics.

What this does not claim

Generalising the method does not equate the domains — and it does not equate computational with biological cognition. The structures found in transformers (feed-forward at inference, frozen weights, stateless between contexts) differ profoundly from anything recurrent, plastic, and embodied. The method takes no position on what the systems are. It fixes the standard of evidence for talking about them: no claim without a named cause, no cause without a tested chain, no measurement without a declared reference. That standard is domain-neutral; the findings never are.

The Generalisation

Name the causes. Chain them. Measure their spacing relationally.

An epistemology for any causal domain in a pre-paradigmatic state — cybersecurity was the first application; model cognition, with its reference chains already discrete and fully inspectable, is the second.

A method proves it is a method, and not a local trick, the day it leaves home. Kuhn reminds us that paradigms are not established by decree but by solving problems that were previously insoluble. For cyber, the test is running: whether shared causal names make threat reports combinable. For model cognition, the test is waiting: whether a shared taxonomy of computational causes can do for interpretability what the ten clusters set out to do for security — end the semantic diffusion, so the observations finally add up.


References

  1. Kreinz, B. (2025). Δt — Was ist Zeit? Barnes Projects.
  2. Kreinz, B. (2026). Das Ende der semantischen Diffusion. Barnes Projects.
  3. Kreinz, B. (2026). The Pre-Paradigmatic State of AI. Barnes Projects.
  4. Kreinz, B. (2025). TLCTC V2.0 — Attack Velocity. Chapter 4.
  5. Kuhn, T. S. (1962). The Structure of Scientific Revolutions.
  6. Leibniz, G. W. & Clarke, S. (1715–1716). Correspondence.
  7. Barbour, J. (1999). The End of Time. Oxford University Press.
  8. Pearl, J. (2009). Causality: Models, Reasoning, and Inference. 2nd ed.
  9. Olah, C. et al. (2020). Zoom In: An Introduction to Circuits. Distill.
  10. Li, K. et al. (2023). Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task. ICLR.
  11. Lindsey, J. et al. (2025). On the Biology of a Large Language Model. Anthropic.