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

The Probe Does Not Cry Wolf

By Joshua Bolick, Awareness Software Group ~12 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. Our best early-warning tool reads a model's internal state and warns, a little ahead of time, that the model is about to learn its task. Every test of it so far used models that did learn. A warning that also goes off when nothing is coming is not much use, so we tried it on models that never learn at all.

What we found. On twenty such models, ten trained on pure noise and ten learning too slowly to finish, it never went off. But it came closer than you might expect: a model trained on noise still carries enough of what it has just seen to fool the tool if its alarm line were set a little lower.

Why it matters. So the tool is honest, and the reason is where its line is drawn. Anyone using it on a new kind of model should check that line against models that cannot learn first.

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, 21 training runs of 900 steps with F2's probe fitted at every evaluation. The design and kill test were committed (32242ac) before any run.

Program v2 Bucket L, item L5. Decisive computation: analysis/probe_failure_population.py. Output: analysis/probe_failure_population.json. Reproduce with python analysis/probe_failure_population.py (about three hours on a throttled laptop CPU); --reuse re-derives every endpoint from the saved curves, and was checked to print the identical summary.

The question

F2's decode probe -- a linear readout of the model's hidden state -- crosses 50% accuracy about 23.5 steps before the model does, and is the programme's most quoted instrument. Every figure for it is conditioned on the model learning. K6 checked the two cheap gradient alarms on runs that never transition and said explicitly that it did not cover the probe.

**On models that never learn, the probe stays below its alarm level.** A probe that reads a model's internal state normally warns that the model is about to learn. Here it is run on twenty models that never learn: ten trained on pure noise and ten learning too slowly to finish. Each bar is the share that would set off the alarm at a given level; the level actually used is 0.5. At the level actually used, the probe raised no false alarms at all. But a model trained on noise still holds enough of its input to fool the probe at 0.3 or below, so the warning is honest only because of where its line is drawn.

L5 runs F2's own run() on K6's two failure constructions, applied by a wrapper that leaves F2's loop untouched:

unlearnable -- training targets resampled as uniform noise every step (I6's and K6's construction). The probe is still asked to decode the true lagged token, as on a real run: if a model that never uses the answer carries it readably anyway, the probe fires falsely.
too-slow -- K6's learning rate 0.00012, so the model would learn eventually but not in 900 steps. An alarm here is early rather than wrong.

Kill test, fixed before execution: in both arms the probe's false-alarm rate (crossing 0.5 on a run whose model never reaches 0.5) is no higher than K6's surrogate rate, 0.20.

Anchor, in code: F2's seed 601, re-run unmodified, reproduces F2's committed curves exactly. It does.

Result: the kill test fires -- the probe is honest at its own level

Arm	Runs that never learned	Probe crossed `0.5`	Mean probe maximum
`unlearnable`	`10/10`	`0`	`0.290`
`too-slow`	`10/10`	`0`	`0.419`

On twenty runs where the model never learned, the probe never fired. F2's headline reliability is not an artefact of only looking at runs that learn.

The margin, and what it depends on

The probe is not silent on these runs; it is below the line. Chance is 0.031 and F2's shuffled-label control sits at 0.032, but a model trained on pure noise still carries the lagged token readably at 0.29: a recurrent state holds its recent inputs whatever it is trained on. How many of the ten runs in each arm would have fired at a lower crossing level:

Level	`0.20`	`0.25`	`0.30`	`0.35`	`0.40`	`0.45`	`0.50`
`unlearnable`	`10`	`10`	`2`	`0`	`0`	`0`	`0`
`too-slow`	`10`	`10`	`10`	`10`	`9`	`1`	`0`

F2's 0.5 sits about 0.2 above what a noise-trained model carries, and about 0.05 above what a slow learner reaches in 900 steps. Below 0.3 the probe would fire on every noise-trained run. N14 swept the crossing level from 0.2 to 0.9 when comparing instruments; any use of the probe at a level below about 0.35 would inherit false alarms that the published figures never saw. The too-slow arm is climbing (0.42 at step 900) and would presumably cross later -- early, as K6 put it, rather than wrong.

What stands

Kill test fires. The probe crosses 0.5 on 0 of 20 never-learning runs.
The level carries the honesty: a noise-trained model reads at 0.29, so levels at or below 0.3 false-alarm on most such runs.
Practical reading: the probe's false-alarm floor is set by how much of the input a recurrent state holds for free, and should be checked for any new architecture or task before a level is chosen.

Limits

One task, one width, F2's seeds; 900 steps, so the too-slow arm's eventual crossing is not seen.
The levels table is post hoc, computed from the saved curves; only the 0.5 row was the test.

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.
false-alarm rate: How often a warning system goes off when nothing is actually happening. Any detector can be made to fire early by making it twitchier, so a lead time only means something alongside this number.
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.
hidden state: The model's working memory: the internal numbers it carries from one step of a sequence to the next.
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.
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.
post hoc: Worked out after the fact, rather than decided in advance. We report such checks separately and never let them decide a result, because it is far too easy to find a pattern once you already know the answer.
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.
reliability: How often something works on an individual case, as opposed to how well it does on average. The two can differ a lot, and only one of them tells you what to expect next time.
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.
surrogate: A fake version of real data, built to keep everything except the effect being tested. If a detector fires just as often on the surrogate, it is responding to something other than what it claims.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.