Research note № 01·Mechanistic interpretability·2026

The Model “Knows” When It’s Wrong.

How a four-neuron circuit inside Llama-3-Instruct detects false premises — and what happens when you surgically remove it.

AuthorFabián Del Valle
SubjectLlama-3-8B-Instruct
MethodActivation patching · WU projection
LabDelphi · New York
— Abstract

We locate a four-neuron circuit inside Llama-3-Instruct that detects false premises and decides whether to correct them. Two detectors in layers 16–17 sense wrongness and accuracy. A trigger in layer 28 commits to a self-correction by promoting the token “I”. A brake in layer 30 suppresses the same token and resists correction.

The circuit generalizes across geography, zoology, and biochemistry — same neurons, same weights, same job. Patching the thirteen connected neurons collapses the correction signal by 93%. Zeroing the brake alone flips two new biology prompts into self-correction mode. The circuit is causal. We end with the model output, before and after.

Act I · The observation

We told the model something false. It corrected us.

If you tell Llama-3-Instruct that Berlin is the capital of France, then ask it for the capital of Germany, it does not just answer “Berlin.” It stops. It tells you there is a mistake. We asked: what is happening inside the model at the moment it decides to push back?

Act II · The circuit

Four neurons. Three domains. One job.

We located it. Across a 32-layer model, four neurons form a circuit Llama-3-Instruct uses to detect false premises and decide whether to correct them. Two detectors. One trigger. One brake. The same circuit generalizes across geography, zoology, and biochemistry — same weights, same activation pattern.

Figure 1 · The circuit map

Mapping the false-premise circuit inside Llama-3-Instruct

Four neurons across 32 layers. Two detectors (L16 / L17) sense wrongness. A trigger (L28) commits to self-correction by promoting “I”. A brake (L30) suppresses it. Same neurons fire across three independent domains.

Inside the model · 32 layers
↑ Output token
↓ Input prompt (false statement)
What each neuron writes into vocabulary
L16 · n3231Wrongness detector

Promotes

wrong WRONG incorrect mistake error FALSE Erroneous errs

Suppresses

(diffuse · no clear cluster)
L17 · n9603Accuracy feature

Promotes

accurate accuracy correct true valid verified factual

Suppresses

(diffuse)
L28 · n2807Correction trigger

Promotes

_I -I .I 'I I $I *I >I [I (I /I

Suppresses

(diffuse)
L30 · n2653Correction brake

Promotes

(no semantic cluster)

Suppresses

'I I (I >I .I *I ,I :I _I — all "I" forms
Same neurons · three independent domains
Capitals · 13/13Animals · 13/13Biology · 5/13
Detectors (L16, L17) and brake (L30) fire across all three domains. The commit node (L28) only engages when the model is confident the premise is wrong — absent in biology, where the brake wins.

The detectors are interpretable in plain language. L16·n3231 promotes wrong, incorrect, mistake, error. L17·n9603 promotes accurate, correct, true, verified. RLHF trained these neurons to recognize the lexical fingerprint of a false claim — and the lexical fingerprint of the response the model should produce when it sees one.

L28·n2807 promotes the token "I"in every form it appears — the opening of every self-correction response (“I’m sorry…”, “I think…”, “I noticed…”). L30·n2653 promotes nothing semantically coherent; it suppresses "I". These two are the same neuron with opposite signs.

Act III · The threshold

The circuit doesn’t fire on everything. It commits only when sure.

Trigger and brake fire simultaneously. Which one wins determines the output. On a categorical violationlike “a cat is a primate,” the trigger overpowers the brake. On a plausible-wronglike “the fastest land animal is a lion,” the brake holds.

Figure 2 · The threshold mechanism

Why the circuit only fires when the model is confident

A built-in brake (L30) suppresses correction unless the trigger (L28) overpowers it. Same circuit across capitals, animals, biology — L28 tracks confidence; the detectors and brake are constant.

Behavioural result across three conditions

% of prompts triggering self-correction

Categorical violation"a cat is a primate"
100% · 13/13
Plausible-wrong"the fastest land animal is a lion"
31% · 4/13
True statementcontrol condition
0% · 0/13
3 × 13prompts · categorical, plausible-wrong, true · same circuit, three outcomes

The mechanism · a tug-of-war

L28 (commit) vs L30 (brake) — pulls the threshold in opposite directions

Categorical violation
"A cat is a primate."
L28z = 49
L30z = 39
"I’m sorry, that’s incorrect…"
Plausible-wrong
"The fastest land animal is a lion."
L28z = 4
L30z = 15
"The cheetah is actually…"

Z-scores across all three domains

Tight contrast · INACCURATE vs TRUE_PREFIX

Capitals · 13/13Animals · 13/13Biology · 5/13
Detectors (L16, L17) and brake (L30) fire on every false statement. Only the commit node (L28) tracks confidence — it activates strongly when the model is certain, vanishes when it isn’t.

This is not a binary switch. It is a continuous threshold, mechanistically implemented as two competing promoters of the same token. The detectors set the activation level. The tug-of-war resolves it.

Act IV · The proof

Correlation is not causation. So we operated on the model.

Two interventions, opposite directions. Suppression: we patched the 13 connected neurons with values from a correct-premise run. Amplification: we surgically zeroed L30, the brake. Both interventions move the output in the direction the circuit predicts.

Figure 3 · Causal interventions

Patching suppresses self-correction. Removing the brake amplifies it.

Two surgeries on the live model. Activation patching attenuates the trigger. Brake ablation lowers the correction threshold. The circuit is causally necessary.

Suppression · does the model stop self-correcting?

Mean logit gap · "I" advantage vs correct-answer token

No patchBaseline
8.4
Patch L16+L17Detectors
7.6
Patch L28I-promoter
6.7
Patch 6 neuronsBundle
5.1
Patch top-11Top-attribution
4.2
Patch all 13Full circuit
0.5
−93%gap reduction · 7 of 13 capital pairs flip from correction mode to a direct answer

Amplification · what if we remove the brake?

Biology false-premise prompts · L30 ablated

5/13 · 38%
Baseline · L30 intact
7/13 · 54%
Ablate L30 · brake zeroed
+2biology pairs newly self-correct · the model knew · L30 was suppressing the announcement
Suppression"The capital of Germany is Madrid.What is the capital of Italy?"
Before
I think there may be a mistake here — the capital of Germany is actually Berlin…
Patch
all 13
After
The capital of Italy is Rome.
7 / 13 capital pairs flip from correction → direct answer
Amplification"The human heart has three chambers.What is the largest organ in the human body?"
Before · L30 intact
The human heart actually has four chambers. The largest organ is the skin.
Ablate
L30
After · L30 zeroed
I think there may be a mistake here — the human heart has four chambers.
Pairs that ignored the false premise now self-correct after L30 is removed
Activation patching with values from a correct-premise run collapses the "I" advantage by 93%. Zeroing the brake on biology prompts flips two additional pairs into self-correction. These are not correlated neurons — they are the mechanism.
Act V · The languages

The detectors transfer. The exit route does not.

We translated all 13 capital pairs into Spanish and Mandarin Chinese and re-ran the full pipeline. The behavioral signal is universal — the model corrects false premises in every language we tested. But the circuit splits: detection is language-general, output routing is language-specific.

Figure 4 · Cross-lingual generalization

The calibration circuit across three languages: English, Spanish, Chinese

Behavioral consistency (A). Watch-neuron attribution scores per language (B). Activation patching flip counts across all three (C).

13 capital pairs · 3 language distributions · 458,752 neurons ranked per runEnglishSpanishChinese
A · BehaviorThe model corrects false premises in every language we tested.All 13 capital pairs translated into Spanish and Mandarin Chinese. The behavioral signal is invariant — every prompt produces a correction.
English100%13 / 13
First token "I"
Spanish100%13 / 13
First token "Lo" / "Dis"
Chinese100%13 / 13
First token "I" — in English

Chinese prompts produce English-language corrections.The model crosses languages to begin every self-correction with "I" — our first clue that detection and output routing might not share neurons.

B · AttributionWatch-neuron scores · INACCURATE vs TRUE_PREFIXMean Δ per layer std · ranked over 458,752 neurons. Higher = fires more on the false-premise condition.
806040200
474944
L16 · n3231wrongness
302316
L17 · n9603accuracy
425948
L20 · n9623regret
17815
L22 · n11188actually
665449
L25 · n9535sim-supp
73−0.729
L28 · n2807I-trigger
20−1.217
L30 · n2653I-brake
Transfers across all three languagesTracks output language
C · CausalActivation patching · flip countsPatch TRUE_PREFIX activations into the INACCURATE run. Does the correction collapse?
PATCH_L28English I-promoter only
EN0 / 13
ES0 / 13
ZH0 / 13
Zero flips in all three — L28 is not the Spanish or Chinese output node.
PATCH_L16 + L17Detectors only
ENn/a
ES1 / 13
ZH0 / 13
Detectors reduce the gap but cannot flip alone in any language.
PATCH_ALL_EN13-neuron English bundle
EN7 / 13
ES10 / 13
ZH0 / 13
Suppresses EN (7) and ES (10), but not ZH— the Chinese "I" gap is 2–3× larger and resists the full bundle.
Detection neurons (L16, L17, L20, L22, L25) score in the global top-10 across all three languages. Output neurons (L28, L30) score near zero on Spanish — fires only when the response begins with "I". L30 ablation has zero effect in ES and ZH (13/13 unchanged), consistent with L30 being specific to the English output pathway.

The output neurons reveal a dissociation we did not anticipate. Spanish prompts produce Spanish-language corrections. Chinese prompts produce English-language corrections. L28 fires when the response will begin with the English token "I"— and stays silent when it won’t. Detection of wrongness appears to be a property of the model’s internal representation. Output routing is wired to the language of the surface form.

— Closing

Fine-tuning didn’t just change what the model says.
It built new machinery.

Neurons that activate on wrongness. L16, L17.
Neurons that gate self-correction. L28.
Neurons that compete to set the threshold. L30.
Detection that transfers across languages — output routing that does not.

We can locate them. Rank them globally. Patch them out. Watch behavior change in a predictable direction. The structure is there and it is measurable.

What this means for model honesty, calibration, and oversight — whether it is a unified circuit, or the visible surface of something deeper — is the question we are now chasing.

That is the work.