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

Which Weights Carry the Transition?

By Joshua Bolick, Awareness Software Group ~5 min read August 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. More about this programme.

EXPLORATORY. Not a preregistered study. Local CPU, 40 training runs, no GPU, no cost.

Program v2 tier 3, item D1. Decisive computation: analysis/transition_anatomy.py. Output: analysis/transition_anatomy.json. Reproduce with python analysis/transition_anatomy.py in about twenty minutes on a throttled laptop CPU.

The question

The representation expands at the transition. Which weights move to produce that?

**The part that changes most is not the part that matters most.** Six parts of a small model. The blue bars show how much each part's numbers change during the moment the model learns the task. The orange bars show how much accuracy is lost if that part is held fixed and never allowed to learn at all. Each series is scaled against its own largest value, because the two quantities have no shared unit. The tallest blue bar and the tallest orange bar are different parts. The component that changes the most during learning turns out to cost nothing when frozen, while the one that actually matters changes less. Watching where a system is busiest is a poor way to find out which parts it depends on, and this is the cheapest possible demonstration of that: you have to hold each part still and see what breaks.

E3 sharpened this a great deal. The expansion is present in every gated recurrent cell tested and absent in attention and in an ungated cell, so multiplicative gating is the suspect. A GRU stacks three gates in each recurrent matrix: reset, update, candidate, 48 rows each at width 48, so "which matrix" becomes answerable as "which gate".

Two parts, because observation and causation are different questions.

Part A: where the motion is

Per-group update norms through training, aligned to each run's own transition, and reported relative to each group's own weight norm, because an absolute update norm mostly measures how many parameters a group has. The window is ±40 steps around the transition.

Group	In window	Elsewhere	Concentration
gate: update	0.06624	0.02429	2.73x
gate: candidate	0.06161	0.02741	2.25x
gate: reset	0.05436	0.02740	1.98x
head	0.05096	0.01465	3.48x
final_norm	0.01781	0.00355	5.01x
embedding	0.00842	0.00492	1.71x

Motion localises, in both senses. It concentrates in time, every group moves 1.7x to 5x faster inside the transition window than outside it, and in place: the update gate moves 7.9x faster than the embedding.

The kill test, "motion is spread uniformly across matrices and time", does not fire.

Part B: what is load-bearing

Freeze one group at a time and ask what it costs. Every arm runs with weight_decay = 0, including the control, because AdamW's decoupled decay shrinks a parameter whether or not it has a gradient, and a "frozen" group under the project's usual 0.01 would drift toward zero and measure decay rather than freezing.

Frozen group	Learned	Accuracy lost	Transition width	Expansion remaining
gate: candidate	5/5	`+0.2328`	3.47x	56%
gate: reset	5/5	`+0.0124`	1.84x	106%
embedding	5/5	`+0.0017`	1.02x	101%
gate: update	5/5	`-0.0013`	1.00x	121%
final_norm	5/5	`-0.0024`	1.02x	102%
head	5/5	`-0.0043`	1.22x	98%

Nothing is fatal. All six freezes still transition in 5 of 5 runs. The network routes around every single-group freeze, which is worth stating plainly because it is the kind of result that would be invisible under a pass/fail reading.

The result: motion and necessity dissociate

The update gate moves fastest in the transition window and costs nothing to freeze: its arm is a hair more accurate than the control and expands 21% more. The candidate gate moves less and is the expensive one: freezing it costs 23 percentage points of accuracy, makes the transition 3.5x wider, and removes 44% of the expansion.

Where the parameters move is not where the work is being done. That is the entire reason Part B exists, and a study that ran only Part A would have named the update gate and been wrong.

The ordering is also informative in its own right. The candidate gate is the GRU's proposal for a new hidden state, the content of what gets written. The update gate controls how much of that proposal is admitted. That the content matters more than the admission rate is not obvious in advance, and it is the first mechanistic handle this program has on what gating contributes.

Two corroborations of earlier results

Freezing the embedding costs 0.0017. Archive mining found the embedding unchanged to four decimal places under the geometry penalty; here it is confirmed to be doing almost nothing at the transition either, from the causal side.
**Freezing the head costs nothing** (-0.0043, i.e. very slightly better). Combined with F3, where freezing the recurrence was devastating, the pair is clean: the recurrence is load-bearing and the readout is not. A model with a random, never-trained output layer still reaches 0.9954.

Limits

One architecture (GRU), one task, one width, five seeds. The gate decomposition is GRU-specific by construction; an LSTM has four gates and would need its own version, which E3's result now makes worth running.
Freezes are single-group. Pairs are untested, and "no single group is required" is compatible with some pair being required.
Part A measures update norms, not directions. Two groups moving equally far may be doing very different things, and a rotation that leaves the norm unchanged is invisible here.
The weight_decay = 0 control makes Part B internally consistent but not directly comparable to the project's archived runs, which all used 0.01.

The plain-language version. This page is the record: the design, the numbers and the limits. If you would rather have the story, Which Weights Carry the Transition?, 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.

All research from this programme →