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

One Jump, or Several?

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: What actually happens at the moment a model learns? – It builds machinery rather than selecting it, working through the task in a reproducible order and trying a simpler wrong rule on the way. The visible training curve cannot tell you which is happening.

In plain English

What we asked. The sudden improvement looks like a single jump on a graph. We zoomed in far enough to see whether it really is one.

What we found. It is four, about sixteen steps apart, and they happen in a consistent order determined by which part of the pattern the model is predicting. The order repeats across independent runs at a correlation of +0.983, which is close to exact.

Why it matters. A curve that looks like one event at normal resolution is four events at high resolution. Anything that treats the jump as a single moment, including several of our own earlier measurements, is averaging over structure that is really there.

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, 10 training runs, no GPU, no cost.

Program v2 tier 3, items D2 (per token class) and A12 (per position), which are the same measurement applied to two ways of slicing the same runs. Decisive computation: analysis/acquisition_microscope.py. Output: analysis/acquisition_microscope.json. Reproduce with python analysis/acquisition_microscope.py in about ten minutes on a throttled laptop CPU.

QUALIFIED 2026-08-26 by N11. The stagger-over-width ratios below are measured on accuracy, which is an argmax, and an argmax widens a class's transition because a class cannot move until its correct token overtakes its rival. Measured in bits per token on the same runs, the ratio falls from 1.442 to 0.706 on delayed-copy and from 0.417 to 0.174 on dispatch-copy-8lag, both paired differences excluding zero. **The acquisition ordering reported here is untouched: it remains +0.983 and +0.910 consistent across seeds. The separation is roughly halved without the threshold**, so "one jump is really four" is substantially a statement about when each element wins the argmax. The original text below is unchanged.

The question, and why it matters now

Every measurement this project has made of the transition treats it as one thing: accuracy averaged over every scored position jumps from chance to ceiling. That average could be hiding a staircase.

**What looked like one moment of learning is really four.** The task asks a model to repeat a short pattern it has already seen. Measuring overall accuracy, the model appears to learn the whole thing in one sudden burst. Measuring each element of the pattern separately shows four bursts instead, about sixteen steps apart. The model learns the pattern in order: first element, then second, then third, then fourth. Averaging across all of them blurred four events into one. The order is almost perfectly repeatable across runs that share nothing but the task, which means the earliest element is an early warning for the rest. Averages hide structure, and here the structure was the finding.

It matters because D3 closed the control programme by showing the transition's timing cannot be steered. Structure inside the transition would be a different handle, and the program has none.

Kill test: all classes move together within a few steps.

"A few steps" is not scale-free, so it is made so: a stagger between classes counts only if it is large relative to how long a single class takes to transition. Two classes separated by 5 steps mean nothing if each takes 60 steps to cross.

Result 1: per position, there is a staircase, and it is the same one every time

delayed-copy, 13 scored positions, five seeds:

Measure	Value
Stagger between classes	75.0 steps
A single class's own width	52.0 steps
Stagger / width	1.44
Order consistency across seeds	`+0.983`

The classes resolve separately, and the order in which they resolve is almost perfectly reproducible between runs that share nothing but the task.

Result 2: what orders the staircase is not position

The per-position transition steps are not monotone in position: 40, 70, 95, 114, 54, 77, 91, 113, 74, 80, 88, 100, 79. "Later positions are harder" is simply false, and so is its opposite.

The script scores three candidate orderings rather than leaving the pattern to be spotted:

Ordering	Spearman
Prefix element being predicted	`+0.910`
Absolute position	`+0.302`
Position within the period	`-0.196`

delayed-copy repeats a prefix of length lag, so position p predicts prefix element (p + 1) mod lag. Grouping by that makes the staircase obvious:

Prefix element	Positions	Mean transition step
0	3, 7, 11, 15	61.8
1	4, 8, 12	75.7
2	5, 9, 13	91.3
3	6, 10, 14	109.0

The model learns to reproduce the repeating prefix one element at a time, in order, roughly 16 steps apart. What looked like a single event is four, and the ordering is the sequence order of the thing being memorised.

Result 3: per lag, the same consistency without the separation

dispatch-copy-8lag, six lag classes:

Measure	Value
Stagger between classes	105.0 steps
A single class's own width	251.5 steps
Stagger / width	0.42
Order consistency across seeds	`+0.910`
Ordered by announced lag	`-0.886`

By the scale-free criterion this is not a staircase: each class takes so long to transition that the classes overlap heavily. It is a ramp with a consistent internal ordering rather than a set of discrete steps.

And the ordering is counterintuitive: rho = -0.886 against the announced lag means longer lags are acquired first. Lag 6 and 7 transition around steps 80 and 90; lags 2 and 3 around 162 and

Whatever makes a short lookback easy for a human is not what makes it easy for this model.

What the two views agree on

The two decompositions disagree about whether to call it a staircase, and agree about the thing that matters more: the acquisition order is highly reproducible in both (+0.983 and +0.910). Whether the steps are separated enough to resolve depends on how wide each step is, which is a property of the task. The existence of a consistent order does not.

That is the internal structure the program lacked. Concretely:

A11 gains a candidate. A12's own note says that if acquisition propagates, the earliest class is a candidate leading indicator for the full transition. It does propagate, reproducibly, and the earliest class transitions ~47 steps before the latest.
A14 gains a handle. Position-dependent difficulty is real and ordered, so a curriculum over prefix elements is now a well-defined intervention rather than a guess.
It complements F2 without duplicating it. F2 found information present in the state ~23 steps before the model can use it. This finds parts of the output arriving before other parts. Both are "early predicts late", measured on different objects.

Limits

One architecture (GRU), one width, five seeds, two tasks.
Both orderings are correlational. Nothing here shows that the early classes cause the late ones; a curriculum experiment or a per-class freeze would be needed, and D1's freeze machinery is the obvious tool.
The 1.0 threshold on stagger-over-width is a choice, made before the runs but not preregistered. The two tasks land at 1.44 and 0.42, comfortably either side, so the verdict does not turn on a borderline call, but a task landing near 1.0 would need a better rule.
The dispatch task's six usable classes exclude lag 8, which did not clear the sample-count floor on every seed.

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.
argmax: Short for 'the argument that maximises'. A model outputs a score for every possible next token; the argmax is simply whichever one scored highest. Accuracy is measured on the argmax, which is why accuracy is a thresholded quantity: only the winner counts, and by how much it won is discarded.
bits per token: How surprised the model is by the correct answer, in the units used by compression. It is the training loss written a different way. Unlike accuracy it has no right-or-wrong threshold in it, so a gradual change in the model shows up as a gradual change in the number. Pure guessing on a 32-symbol vocabulary costs 5 bits per token.
curriculum: The order and mix in which a model is shown its training material, for example easy examples first and harder ones later.
GRU: Gated Recurrent Unit. A compact design for processing sequences one item at a time, with internal switches controlling what it keeps in memory.
kill test: A condition written down before running the experiment that says what result would make us abandon the idea. Fixing it in advance is what stops a disappointing result being reinterpreted as an encouraging one.
reproducible: Someone else running the same code on the same inputs gets the same numbers. Every result here is checked this way.
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.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.