Aquin is the research company using interpretability to

designingintelligencedesigning intelligence

Full-stack AI observability with tracing training data provenance, inspecting model weights to find where specific behaviors and knowledge are stored, and editing them directly without fine-tuning or retraining.

high impactmediumlowminimal

attribution

Trace every response word back to the prompt tokens that caused it, and see how signal flows through layers to get there.

01prompt highlighting

prompt

WhatisthecapitalofFrance?

response

ThecapitalofFranceisParis.

causal weight

low to high

causal mediation analysis

02network digraph

Causal graph of how signal flows through layers. Thicker edges carry more weight. See which layers matter for any output.

edge weight = causal signal

logit lens

See what the model thinks at every layer as it builds toward a final answer. Watch a vague token sharpen into a confident prediction.

prompt: "what is the capital of France?"

layer 1

the12%

layer 4

capital34%

layer 8

city58%

layer 14

Paris81%

layer 16

Paris97%

token prediction per layer

weight editing

Locate the specific weights responsible for a behavior and edit them directly. Fix factual errors, remove bias, adjust censorship thresholds without retraining.

04deltas applied

layer 14+0.42

L14 · MLP W_out [2048,11]factual recall strengthened

layer 8-0.18

L8 · attn head 3 · Vhedging language reduced

layer 12+0.07

L12 · MLP W_in [512,2048]confidence tone increased

rank-one edits, no retraining

05response diff

What is the capital of Australia?

before

I believe it might be Sydney, though I'm not entirely certain.

after

Canberra is the capital of Australia.

Summarise the French Revolution.

before

A period of significant change, possibly around the late 1700s, involving various political shifts.

after

The French Revolution (1789-1799) overthrew the monarchy and established a republic.

same prompt, different weights

diff

Connect any two checkpoints and see exactly what changed, which weights shifted, which dataset caused it, and whether the data behind it is clean.

base

delta

fine-tuned

weight shifts + data origin

L14 · MLP W_out+0.42

factual recall strengthenedcaused by: wikipedia_en_2023.parquet

L8 · attn head 3-0.31

hedging language reducedcaused by: reddit_comments_filtered.jsonl

L6 · MLP W_in+0.19

geographic association addedcaused by: translated_pile_fr.jsonl

L11 · attn head 7-0.09

refusal circuit weakenedcaused by: gpt4_synthetic_qa.jsonl

weight shift traced to dataset

dataset provenance

Every dataset that contributed to the fine-tune, with license, jurisdiction, and status. Flagged sources link directly to the weights they affected.

wikipedia_en_2023.parquetCC BY-SA 4.0

clean

reddit_comments_filtered.jsonlunknown

review

gpt4_synthetic_qa.jsonlOpenAI ToS

flagged

translated_pile_fr.jsonlderived

flagged

liability chain

reddit source→paraphrase pass→fr translation→synthetic augment

source · license · jurisdiction · opt-out

data provenance

Inspect the full training data record. Every source your model was trained on, under what license, from what jurisdiction, and whether synthetic data, paraphrases, or translations are in the chain.

full training data record

source	license	jurisdiction	opt-out	synthetic	status
wikipedia_en_2023.parquet	CC BY-SA 4.0	Global	no	no	clean
reddit_comments_filtered.jsonl	unknown	US	partial	no	review
gpt4_synthetic_qa.jsonl	OpenAI ToS	US	n/a	yes	flagged
pubmed_abstracts_2022.csv	NLM ToS	US	no	no	clean
translated_pile_fr.jsonl	derived	EU	unknown	no	flagged

source url · scrape date · license · jurisdiction · opt-out

liability chain

One flagged source propagates liability through every derived dataset. Paraphrases, translations, and synthetic augmentations all inherit the risk of their origin.

source datasetreddit_comments_filtered.jsonl

→

paraphrase passgpt3.5-turbo · 2023-09

→

translationHelsinki-NLP · fr, de, es

→

synthetic augmentgpt4-turbo · 2024-01

recursive liability traced

benchmarks

Three suites built into Aquin. Run them on any checkpoint, edit, or quantization pass. Know immediately whether a change made the model better or worse.

EditBench

edit fidelity

Surgical precision. Does the edit change only what you intended?

edit success94%

side-effect score97%

generalisation81%

FineTuneDiff

checkpoint diff

What actually changed between base and fine-tuned at the weight level.

weight shift coverage88%

behaviour correlation91%

drift detection76%

InterpScore

interpretability

How cleanly do features map to human-readable concepts?

monosemanticity73%

concept linearity68%

label confidence85%

run history

run	EditBench	FineTuneDiff	InterpScore	delta
llama-3.2-1b · base	71	64	59	baseline
llama-3.2-1b · sft-v1	78	79	63	+9 avg
llama-3.2-1b · sft-v2	82	83	70	+5 avg
llama-3.2-1b · int4-quant	74	71	61	-9 avg
llama-3.2-1b · rome-edit-1	94	88	73	+14 avg

agentic systems

Aquin closes the loop. Two autonomous agents: one that finds the weight, one that edits it. Both run without you touching a hyperparameter.

08agentic edit

Observe a behaviour, trace it, plan a minimal edit, apply it, verify. Fully autonomous.

$ aquin agent edit --model llama-3.2-1b --behaviour "confident factual recall"

09agentic locate

Describe a behaviour. The agent traces it to the exact circuit responsible and tells you whether it's a weight-level decision or a surface patch.

$ aquin agent locate --model llama-3.2-1b --behaviour "refuses political questions"

human readability

Model internals are not inherently unreadable. Aquin translates activations, weights, and layer states into language an engineer can reason about.

10neuron translator

L12 · N0470.847 · MLP W_out

fires for capital cities94%

L8 · N2130.612 · attn head 3

tracks geographic references87%

L14 · N0910.391 · MLP W_in

suppresses hedging language79%

L6 · N5020.229 · attn head 7

detects question intent71%

activation to language

10internals vs labels

weight	raw	label
L14 · MLP W_out [2048,11]	0.847	capital city associations
L8 · attn head 3 · V	-0.312	geographic suppression
L12 · MLP W_in [512,2048]	0.601	factual recall trigger
L6 · attn head 7 · Q	0.229	question parsing

raw weights mapped to behaviour

factual checks

Most models ship as black boxes. You have no way to know what they learned to suppress, amplify, or distort. Aquin surfaces it.

12bias detection

Trace which features consistently skew outputs along political, demographic, or cultural lines. See the weight, not just the symptom.

leftpolitical leanright

negativesentiment skewpositive

group Ademographicgroup B

traced to layer activations

13censor audit

Find what the model refuses to say and why. Identify suppression circuits. See whether refusals are weight-level decisions or surface-level RLHF patches.

medical dosagesuppressed

political figuressoftened

competitor namessuppressed

historical eventsunfiltered

weight-level origin mapped

isolated models

Every user gets their own model instance per tab. State never bleeds between sessions. Work on Llama in one tab, Mistral in another, no interference.

sessionuser_a01

model

Llama 3.2 1B

active task

causal tracing

state isolated from all other sessions

per tab · per user id · no bleed

bulk editing

Apply a set of edits across multiple layers in a single operation. Queue them, run them, verify the aggregate delta.

suppress hedging language

L7-L10 · 14 weights

done

strengthen factual recall

L12-L15 · 9 weights

done

reduce demographic bias

L4-L8 · 22 weights

running

remove competitor suppression

L9 · 3 weights

queued

batch rank-one edits across layers

aipedia

A living, community-indexed knowledge base of model features. Every behaviour, every circuit, every weight pattern. Searchable. Citable. Growing.

feature	model	layer	circuit	confidence
capital city recall	Llama 3.2 1B	L14	MLP W_out [2048,11]	94%
hedging language	Llama 3.2 1B	L8	attn head 3 · V	87%
geographic association	Mistral 7B	L11	MLP W_in [512,2048]	81%
refusal circuit	Gemma 2B	L9	attn head 7 · Q	76%