How do you know a fine-tuned AI model is actually better?

You measure it. Hold out a portion of the client's real examples so the model never sees them during training, score the plain base model on those held-out questions first, fine-tune, then score again on the identical questions with identical scoring. The difference is the lift. Because the only thing that changed is the fine-tune, the improvement is attributable to it.

Does fine-tuning actually improve accuracy, or is the base model already good enough?

It depends on the task and, critically, on how much data you have. On a public benchmark of 510 lawyer-annotated commercial contracts, a base Qwen3-4B extracted short contract fields at 62.2% accuracy. Fine-tuned on 396 examples it reached 83.5-85.4%, a lift of 21.3 to 23.2 points across three runs. The same model trained on only 48 examples produced a lift inside the noise floor, so the amount of data was the deciding factor rather than the model's ceiling.

How much of our own data do we need to fine-tune a model?

Fewer examples than most people expect, but more than a handful. In our contract-extraction measurements, 48 training examples produced no measurable improvement while roughly 400 produced a 22-point gain. The honest answer for any specific case is that it depends on how varied and how consistent the examples are, which is what a paid readiness assessment measures on the client's own corpus.

Why can't you just use a public AI benchmark score?

Public benchmarks measure general ability on standardised tests, not your work. A model that scores well in the abstract can still be mediocre at reading your contracts or following your policies. Public benchmarks are also increasingly targeted during training, so a high score can reflect test preparation as much as capability. The only benchmark that answers your question is built from your own tasks.

Research record

Does the Early Signal Generalise?

By Joshua Bolick, Awareness Software Group ~13 min read September 2026

Ongoing research. This is an experimental result from active work, not a settled conclusion. The numbers are what we measured and the method is described so you can judge it, but the programme is still running and later experiments may revise what it means.

A note on the language in these records. This is a working laboratory notebook for research into training AI models more cheaply and efficiently, so you will read that an approach did not work, that a result did not hold up, or that one method was worse than another. That is the research doing its job, not a verdict on the engineering we deliver to clients. Ruling an approach out is how the search narrows, and these are the pages that teach us the most: nearly every technique we now rely on came from understanding why something else fell short. Testing our own ideas at least as hard as anyone else's is the point of publishing them. More about this programme and why we run it.

Part of a bigger question: Can you tell in advance that a model is about to improve? – Yes, repeatedly and by several independent routes -- a probe on the hidden state, gradient statistics, held-out loss, and which earlier word the prediction leans on. The warnings are real; how far ahead they fire varies a great deal.

In plain English

What we asked. We found an early warning signal in one setup. Does it work anywhere else?

What we found. Partly. The signal arrives about 23 steps early in one design, 14 in another, and is not measurable at all in a third. We ran a calibration check on all three to confirm the instrument works where it reports nothing, and it does, so the absence is real rather than another weak measurement.

Why it matters. A signal that works in some architectures and not others is much less useful than one that works everywhere, and it is honest to say so. The calibration check is the part worth copying: it is what separates 'there is nothing here' from 'we could not see'.

The rest of this page is the technical record: the design, every number, and the limits. It is written for a reviewer, and you do not need it to have understood the result above.

New here? How to read a research record

Start at the verdict. Every record states, before the experiment was run, what result would have made us abandon the idea. That is the "kill test". Then it says whether the test fired. Nothing gets reinterpreted after the fact.
Numbers in square brackets are uncertainty. 23.4 [18.1, 28.7] means our best estimate is 23.4 and the true value is probably somewhere in that range. If a range includes zero, we cannot claim an effect.
Results that rule an idea out are kept. Roughly half of what is published here says an approach did not work, including plenty of our own. Those pages are the output, not a shortfall: knowing which direction is a dead end is what lets the next experiment go somewhere better, and most of what we now rely on came out of understanding why something else fell short. Work that only publishes what worked is not measuring anything.
Read the Limits section. Every record ends with what it does not show. It is the most honest part of any experiment and usually the shortest.
Pro tip: the figures near the top are designed to carry the result on their own. If you read nothing else, read the caption under each one, which says what it shows and what to take from it.

EXPLORATORY. Not a preregistered study. Local CPU, 15 training runs, no GPU, no cost.

Program v2 tier 3 follow-up to F2 and E3. Decisive computation: analysis/decode_probe_architectures.py. Output: analysis/decode_probe_architectures.json. Reproduce with python analysis/decode_probe_architectures.py in about forty minutes on a throttled laptop CPU.

The question

F2 found the program's only leading indicator, in a GRU. (As of 2026-08-23 it is one of two: F1 added gradient-norm variance. Nothing in this record depends on the count.) E3 found that the expansion is a gated-recurrence phenomenon, present in the LSTM and absent in attention.

**The early warning only appears in some designs.** Earlier we found that a model's internal state carries the right answer before the model can produce it, which is a useful early warning. This asks whether that holds for other designs. The bars show how many training steps of warning each one gives; the shaded strip at the bottom is the range too small for our measurement to resolve. Two designs give real advance warning. The third gives none: its bar sits inside the unmeasurable strip. That third design is the one most modern AI is built on, and it is also the one that showed none of the internal reorganisation we had been studying. So the two findings turn out to be one finding, and both are limited to a particular family of designs rather than being facts about learning in general.

Those are different quantities. The expansion is about how a representation reorganises; the decode gap is about whether information arrives before it is used. There is no reason they must travel together, and it matters a great deal which way this falls: the indicator is the one result in this program with practical use, so its scope is the scope of anything useful that comes out of it.

Design. One task, dispatch-copy-wide-vocab, because it is the only cell in the E3 grid where all three architectures genuinely learn. The transformer solves plain delayed-copy almost instantly: that is the confound A1 was written about, and no lead can be measured against a model that arrives already knowing the answer. Learning rates are the ones E3 screened per cell on its excluded seed. Five seeds per arm, crossing measured at 40% accuracy.

Result: they travel together

Architecture	Lead	Resolvable	Max probe-minus-model gap	Probe / model at end	Instrument check
gru	`+23.4` `[+18.8, +28.1]`	5 / 5	`+0.0721`	0.6216 / 0.6141	5 / 5
lstm	`+14.4` `[+11.0, +17.9]`	5 / 5	`+0.0575`	0.7156 / 0.7095	5 / 5
transformer	`+1.3` `[+0.6, +2.0]`	0 / 5	`+0.0711`	0.8300 / 0.8236	5 / 5

The lead appears where the expansion appears, and not in attention. Both are gated-recurrence properties. That is the second of the three outcomes this pilot was written to distinguish.

Two things that make the transformer's null trustworthy

The instrument check passes in all three arms, including the transformer. An optimally fitted linear readout matches the trained head at convergence in 5 of 5 transformer runs (0.8300 against 0.8236). This is exactly the guard F2's own correction installed, and here it is doing its job: the transformer's null is not the under-powered-probe failure that made F2's first version report the opposite of the truth. The probe works. It simply finds no lead.

The transformer's gap is real but not temporal. Its largest probe-minus-model gap (+0.0711) is comparable to the GRU's (+0.0721). The probe is consistently a little better than the trained head. What it is not is earlier: the two curves rise together, so the gap is a constant offset rather than a lead. That distinction is the whole point of measuring a crossing time rather than a gap size, and a pilot that reported only the gap would have concluded the opposite.

The GRU result replicates on a new task

F2 measured +23.5 steps [+22.1, +24.9] on delayed-copy. This measures +23.4 [+18.8, +28.1] on dispatch-copy-wide-vocab: a different task, a different vocabulary size, a different learning rate, and unseen seeds. That is an independent replication of the program's most important positive result, and it was not the question this pilot was asked.

What this changes

F2's scope is now known, and it is the same as the expansion's: gated recurrence. Anything built on the leading indicator inherits that limit, and the honest framing of the program's one useful finding is "in gated recurrent models" rather than "in models".
The two findings are one story, not two. Whatever gating does at the moment of learning produces both the representational expansion and the gap between what the state carries and what the head reads. That strengthens the case for making gating the object of study, which E3 already recommended.
The transformer remains the interesting negative. It learns the task, takes the longest of any architecture to do it, shows no expansion, and shows no lead, while a calibrated probe confirms the information is slightly more decodable than the head manages throughout. Attention appears to acquire and use information at the same time.

Limits

One task, five seeds per arm, one width, one optimizer. This is a scope check on F2, not a second independent confirmation of it.
The crossing is measured at 40% rather than F2's 50%, because the arms plateau between 0.61 and 0.83 on this task and 50% sits too close to the transformer's own curve to time reliably. F2 established that its lead is robust across thresholds from 20% to 90%; that robustness was not re-established here.
The transformer is one configuration (1 layer, 4 heads). "Attention does not show this" is a claim about that configuration, not about attention at scale.

Terms on this page

Every piece of vocabulary this record uses, in plain language. Generated from the text above, so it cannot drift out of step with it.

accuracy: The fraction of answers a model gets right on questions it was not trained on.
architecture: The blueprint of a model: what components it has and how they connect. Two models can be the same size and completely different architectures.
calibrated: A model is calibrated when its confidence matches how often it is right: answers it gives with 90% confidence should be right about 90% of the time.
confound: A second explanation you did not control for. If bigger models both learn faster and score higher, then 'fast learners score higher' may be entirely about size and not about speed.
gated: A design where the model has explicit internal switches deciding what to remember and what to discard at each step. GRUs and LSTMs are gated; plain recurrent networks are not.
gradient: The direction and amount by which each of a model's internal numbers should change to do slightly better. Training is repeatedly following it.
GRU: Gated Recurrent Unit. A compact design for processing sequences one item at a time, with internal switches controlling what it keeps in memory.
learning rate: How big a step training takes each time it updates the model. Too small and nothing happens; too big and it never settles.
LSTM: Long Short-Term Memory. An older and larger relative of the GRU, also gated, also for sequences.
optimizer: The algorithm that decides how to change the model given the gradient. AdamW and SGD are two common choices and they behave differently.
plateau: A stretch of training where the model is not visibly improving. Often, though not always, followed by a sudden jump.
probe: A small separate model trained to read information out of a bigger model's internals, used as a measuring instrument rather than as a product.
recurrent: A design that reads a sequence one item at a time, carrying memory forward. The main alternative is attention, which looks at everything at once.
seed: The number that fixes all the randomness in a training run. Same seed, same run. Running several seeds is how you tell a real effect from a lucky one.
transformer: The architecture behind most modern large language models. It uses attention to look at every part of the input at once, rather than reading in order.
variance: How spread out a set of numbers is. Rising variance in a signal means it is becoming more erratic.
vocabulary: The set of distinct symbols a model can read and produce.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.

The plain-language version. This page is the record: the design, the numbers and the limits. If you would rather have the story, Does the Early Signal Generalise?, explained without the jargon, covers the same result for a reader who has never seen a training curve.

About this record. This page is the write-up of an experiment we ran ourselves. Every number on it comes from a run we recorded, and the script that produced each one is named in the text above. We publish these results and the data behind them here, including the ones that went against what we expected.