Causal grammar

The words spark and flame sit in two circles joined by a causal arrow.

What it is

What it found

One consequence, one rejection, one limit

Sources and reading boundary

What it is

The browser page has four rooms.

Compile a cause accepts a deliberately small grammar: SET A = 1, A CAUSES B, A PREVENTS B, and A AND B CAUSE C. Each valid acyclic program becomes Boolean structural equations. Change a starting value or intervene on any variable. The mapping is exact because both syntax and semantics are stipulated.
Same sights, different worlds shows the observations A = B. They fit A → B, B → A, and a hidden common cause. Setting B by intervention makes those worlds disagree.
Same logprobs, opposite causes counts next-token transitions in one visible corpus. Two readers have identical counts and log probabilities but reverse what dax and wug mean. Distribution over signs does not fix their grounding.
Living causes keeps empirical claims outside the compiler. Brains and animals can learn prediction, action consequences and causal structure. The consciousness theories considered here disagree about what organisation supports experience. Plants sense, signal, remember physiologically and alter growth, but they have no neurons, and response does not establish causal experience.

SET supplies a root value. When a variable also has a causal rule, that rule determines its value; an intervention replaces either the rule or root value. Duplicate SET values and duplicate causal rules are rejected. Rules run in dependency order, regardless of line order. An input needs a value unless an intervention cuts off the equation that needs it.

Use the interactive page in this section. Run the finite checks with:

node causal-grammar/checks.js

What it found

266 exhaustive and structural comparisons pass. P1, P2, P3 and P5 hold in the tested finite cases. The P4 comparison verifies agreement with oxygen present and disagreement when that background condition changes; it does not establish two equivalent descriptions of the unrestricted conjunctive world.

The exact answer to the first question is yes, for a controlled language with a formal interpretation. Programming languages, chemical reaction notation and equations do this within a model: a valid expression denotes an operation because a semantics says so. This is not an exact map from all natural language to nature. It is an exact map from a restricted fragment to a chosen causal model.

The counterexamples are small. Observing A and B together does not choose an arrow. In P4, the reduced rule Flame = Spark agrees with Flame = Spark AND Oxygen while Oxygen = 1. Setting oxygen to zero takes the reduced rule outside that scope, and its answer differs from the full conjunction. Once the full conjunctive law is stipulated, it fixes the intervention answer. Calling the spark the explanation can leave oxygen in the background of the explanation; it does not remove oxygen from the physical law.

Ordinary language raises further interpretive choices. Passive voice can omit an agent. Negation, prevention, permission, purpose, blame, metaphor and counterfactual mood select different causal structures or different parts of one structure. Aristotle’s four kinds of answer already refuse a single arrow; Plato separates a physical condition from a reason for action. Force-dynamic linguistics separates causing, enabling and preventing. These readings are interpretations; the compiler does not test them.

A next-token model codifies these distinctions to the extent that they change the distribution of signs in its training mixture and survive its representation and objective. Code and mathematics contribute unusually rigid local conventions; natural language contributes defeasible, perspective-bound ones. The model may internally support causal variables or intervention-like computation, but log probability alone certifies only a conditional distribution over tokens. The lab’s two readers are the finite counterexample: every displayed log probability is identical while one causal answer reverses.

The literature considered here does not supply one accepted theory of cause-and-effect experience. Predictive processing and active inference describe organisms using generative models to predict and act. Associative and causal-learning accounts ask how covariation, temporal order and interventions support structure learning. Global-workspace, recurrent, higher-order and integrated-information theories address conscious access or experience with incompatible commitments. Comparative cognition tests animals under controls for simpler cues. Plant electrophysiology and chemical signaling establish responsive causal networks; they do not establish subjective causal experience, and calling them neuroscience erases the absence of neurons. The sources and reading boundary below date this account.

Set theory can discipline what counts as a relation, and computational systems research can discipline what counts as an executable rule; neither makes grammar identical to physical cause. Democritus, Plato and Aristotle offer incompatible ancient causal frames. Sappho and Carson contribute something the truth table omits: an experienced sequence in a body can be ordered, intense and linguistically exact without identifying a manipulable physical cause.

One consequence, one rejection, one limit

Consequence: any claim that a model “understands cause” needs a named test: continuation, prediction under observation, answer under intervention, explanation selected for a listener, or experienced consequence. Success at one does not certify the others.

Rejected alternative: the lab rejects using high probability for “A causes B” as an intervention answer. It also rejects the opposite shortcut that language models contain no causal structure. Token training can encode and compute causal regularities; probabilities do not tell us by themselves which regularity is there.

Limit: the exactness is constructed, finite and Boolean. No neural measurements, model activations, chemical reactions, animal experiments or plant experiments were run. The historical readings are not reducible to the compiler, and the page does not establish experience in any system.

Sources and reading boundary

This reading is limited to literature published through 2024. It does not review later research or claim to survey today's consensus. The finite comparisons test the program; the readings below support its distinctions, not a conclusion about subjective experience.

Judea Pearl, Causality, second edition, 2009: structural causal models and interventions.
Joseph Y. Halpern, “A Modification of the Halpern–Pearl Definition of Causality”, 2015: choices required for actual-cause judgments.
Phillip Wolff, “Representing causation”, 2007: distinctions among causing, enabling and preventing.
Aristotle, Physics II.3, and Plato, Phaedo 96a–99d: different kinds of causal explanation and reasons for action.
Anil Seth and Tim Bayne, “Theories of consciousness”, 2022: competing theories of conscious access and experience.
Karl Friston, “The free-energy principle: a unified brain theory?”, 2010: perception and action under a generative model.
Lincoln Taiz and colleagues, “Plants Neither Possess nor Require Consciousness”, 2019: an argument against inferring consciousness from plant signaling and adaptive response. This is a position in the debate, not an experimental verdict from this lab.

Causal grammar?

What it is?

What it found?

One consequence, one rejection, one limit?

Sources and reading boundary?

Causal grammar

What it is

What it found

One consequence, one rejection, one limit

Sources and reading boundary