A research organization

Infrastructure for computational research.

Measured results on how models develop during training, with the methods and evidence behind them.

Read the research

What we work on

A narrow scope, worked deeply.

Counfield studies how models develop while they are trained, and how that development can be measured. The published work is a set of measured training runs, the methods behind them, and the evidence for each result.

We keep the scope narrow on purpose. Results are published with their evidence, negative results as visibly as positive ones, and anything that was not measured is labelled as such.

One run, measured at eight checkpoints.

The same model, read at eight moments of its training on eight cells of matched items. Move along the run to see every cell at one moment, or click a cell's name to read what the research page says about it.

100% of training · step 32,470 · 531,988,480 tokens · validation CE 2.622200

M@4+0.00032/128
M@4 across eight Pilot-0.2 checkpointsNormalized score for cell M@4 at eight checkpoints, from theta-zero to one hundred percent, over n=128 matched items. The zero line is chance. Filled dots are at or above chance; hollow dots are below chance. No confidence interval is drawn because the source does not provide one for these cells.0-0.05M@4 at θ₀ — 35/128, normalized +0.031M@4 at 1% — 26/128, normalized −0.062M@4 at 5% — 34/128, normalized +0.021M@4 at 10% — 27/128, normalized −0.052M@4 at 25% — 36/128, normalized +0.042M@4 at 50% — 35/128, normalized +0.031M@4 at 75% — 32/128, normalized +0.000M@4 at 100% — 32/128, normalized +0.000θ₀1%5%10%25%50%75%100%
R@1+0.15674/128
R@1 across eight Pilot-0.2 checkpointsNormalized score for cell R@1 at eight checkpoints, from theta-zero to one hundred percent, over n=128 matched items. The zero line is chance. Filled dots are at or above chance; hollow dots are below chance. No confidence interval is drawn because the source does not provide one for these cells.00.20R@1 at θ₀ — 66/128, normalized +0.031R@1 at 1% — 65/128, normalized +0.016R@1 at 5% — 66/128, normalized +0.031R@1 at 10% — 65/128, normalized +0.016R@1 at 25% — 78/128, normalized +0.219R@1 at 50% — 67/128, normalized +0.047R@1 at 75% — 76/128, normalized +0.188R@1 at 100% — 74/128, normalized +0.156θ₀1%5%10%25%50%75%100%
I@8+0.39553/128
I@8 across eight Pilot-0.2 checkpointsNormalized score for cell I@8 at eight checkpoints, from theta-zero to one hundred percent, over n=128 matched items. The zero line is chance. Filled dots are at or above chance; hollow dots are below chance. No confidence interval is drawn because the source does not provide one for these cells.00.200.40I@8 at θ₀ — 4/128, normalized +0.000I@8 at 1% — 8/128, normalized +0.032I@8 at 5% — 5/128, normalized +0.008I@8 at 10% — 3/128, normalized −0.008I@8 at 25% — 22/128, normalized +0.145I@8 at 50% — 47/128, normalized +0.347I@8 at 75% — 57/128, normalized +0.427I@8 at 100% — 53/128, normalized +0.395θ₀1%5%10%25%50%75%100%
C@True+0.0247/128
C@True across eight Pilot-0.2 checkpointsNormalized score for cell C@True at eight checkpoints, from theta-zero to one hundred percent, over n=128 matched items. The zero line is chance. Filled dots are at or above chance; hollow dots are below chance. No confidence interval is drawn because the source does not provide one for these cells.0-0.05C@True at θ₀ — 6/128, normalized +0.016C@True at 1% — 0/128, normalized −0.032C@True at 5% — 5/128, normalized +0.008C@True at 10% — 4/128, normalized +0.000C@True at 25% — 2/128, normalized −0.016C@True at 50% — 7/128, normalized +0.024C@True at 75% — 3/128, normalized −0.008C@True at 100% — 7/128, normalized +0.024θ₀1%5%10%25%50%75%100%
Q@1−0.0413/128
Q@1 across eight Pilot-0.2 checkpointsNormalized score for cell Q@1 at eight checkpoints, from theta-zero to one hundred percent, over n=128 matched items. The zero line is chance. Filled dots are at or above chance; hollow dots are below chance. No confidence interval is drawn because the source does not provide one for these cells.0-0.05Q@1 at θ₀ — 2/128, normalized −0.049Q@1 at 1% — 8/128, normalized +0.001Q@1 at 5% — 14/128, normalized +0.051Q@1 at 10% — 5/128, normalized −0.024Q@1 at 25% — 1/128, normalized −0.057Q@1 at 50% — 3/128, normalized −0.041Q@1 at 75% — 2/128, normalized −0.049Q@1 at 100% — 3/128, normalized −0.041θ₀1%5%10%25%50%75%100%
V@8+0.03620/128
V@8 across eight Pilot-0.2 checkpointsNormalized score for cell V@8 at eight checkpoints, from theta-zero to one hundred percent, over n=128 matched items. The zero line is chance. Filled dots are at or above chance; hollow dots are below chance. No confidence interval is drawn because the source does not provide one for these cells.0-0.05V@8 at θ₀ — 17/128, normalized +0.009V@8 at 1% — 13/128, normalized −0.027V@8 at 5% — 20/128, normalized +0.036V@8 at 10% — 21/128, normalized +0.045V@8 at 25% — 20/128, normalized +0.036V@8 at 50% — 20/128, normalized +0.036V@8 at 75% — 20/128, normalized +0.036V@8 at 100% — 20/128, normalized +0.036θ₀1%5%10%25%50%75%100%
A@0+0.28342/128
A@0 across eight Pilot-0.2 checkpointsNormalized score for cell A@0 at eight checkpoints, from theta-zero to one hundred percent, over n=128 matched items. The zero line is chance. Filled dots are at or above chance; hollow dots are below chance. No confidence interval is drawn because the source does not provide one for these cells.00.200.40A@0 at θ₀ — 4/128, normalized −0.033A@0 at 1% — 9/128, normalized +0.008A@0 at 5% — 11/128, normalized +0.025A@0 at 10% — 45/128, normalized +0.308A@0 at 25% — 36/128, normalized +0.233A@0 at 50% — 59/128, normalized +0.425A@0 at 75% — 48/128, normalized +0.333A@0 at 100% — 42/128, normalized +0.283θ₀1%5%10%25%50%75%100%
T@modulus+0.13519/128
T@modulus across eight Pilot-0.2 checkpointsNormalized score for cell T@modulus at eight checkpoints, from theta-zero to one hundred percent, over n=128 matched items. The zero line is chance. Filled dots are at or above chance; hollow dots are below chance. No confidence interval is drawn because the source does not provide one for these cells.0-0.05T@modulus at θ₀ — 4/128, normalized +0.016T@modulus at 1% — 1/128, normalized −0.008T@modulus at 5% — 3/128, normalized +0.008T@modulus at 10% — 1/128, normalized −0.008T@modulus at 25% — 6/128, normalized +0.032T@modulus at 50% — 11/128, normalized +0.071T@modulus at 75% — 14/128, normalized +0.095T@modulus at 100% — 19/128, normalized +0.135θ₀1%5%10%25%50%75%100%
Validation CE2.622200
Validation CE across eight Pilot-0.2 checkpointsValidation cross-entropy at eight checkpoints, from theta-zero to one hundred percent of training, on its own scale. It falls at every checkpoint.48Validation CE at θ₀ — validation cross-entropy 10.503032Validation CE at 1% — validation cross-entropy 5.210447Validation CE at 5% — validation cross-entropy 3.860014Validation CE at 10% — validation cross-entropy 3.311519Validation CE at 25% — validation cross-entropy 2.950929Validation CE at 50% — validation cross-entropy 2.760658Validation CE at 75% — validation cross-entropy 2.655592Validation CE at 100% — validation cross-entropy 2.622200θ₀1%5%10%25%50%75%100%

These are exploratory longitudinal measurements over a matched item set. No formal change-point, mixed-effects or population model was applied.

Eight measured cells across eight Pilot-0.2 checkpoints, one panel per cell. Each panel is on its own scale; the axes are never averaged together. Zero is that cell’s own chance level. Filled dots sit at or above chance, hollow dots below it. Source §6.2, n=128 matched items per cell.
The findings and the values, as text
  1. A became distinct early. The A@0 normalized score moved from −0.033 at θ₀ to +0.308 at 10%, peaked at +0.425 at 50%, and fell back to +0.283 at the final checkpoint.
  2. I rose later. I@8 stayed around chance up to 10%, then reached +0.145 at 25%, +0.347 at 50% and +0.427 at 75%.
  3. In this model the measured A signal appeared before the I signal. This single-run pattern does not support the hypothesis that I is a required precursor of the other axes, and it does not refute it either.
  4. T-modulus rose late, from around 0 to +0.135 at the final checkpoint. This does not mean the whole T axis is ordinal or measurable.
  5. Q showed no positive capacity. The selected Q cell stayed below chance at most checkpoints.
  6. V was largely fixed. V@8 repeating 20/128 across later checkpoints keeps measurement insensitivity on the table alongside model invariance.
  7. Validation CE and the axes did not move together. CE fell steadily while A peaked before the end, M stayed volatile, and I opened late.
Hits over n=128 and the normalized score, per cell and checkpoint.
CheckpointStepTokensValidation CEM@4R@1I@8C@TrueQ@1V@8A@0T@modulus
θ₀0010.50303235/128
+0.031		66/128
+0.031		4/128
+0.000		6/128
+0.016		2/128
−0.049		17/128
+0.009		4/128
−0.033		4/128
+0.016
1%3255,324,8005.21044726/128
−0.062		65/128
+0.016		8/128
+0.032		0/128
−0.032		8/128
+0.001		13/128
−0.027		9/128
+0.008		1/128
−0.008
5%1,62426,607,6163.86001434/128
+0.021		66/128
+0.031		5/128
+0.008		5/128
+0.008		14/128
+0.051		20/128
+0.036		11/128
+0.025		3/128
+0.008
10%3,24853,215,2323.31151927/128
−0.052		65/128
+0.016		3/128
−0.008		4/128
+0.000		5/128
−0.024		21/128
+0.045		45/128
+0.308		1/128
−0.008
25%8,118133,005,3122.95092936/128
+0.042		78/128
+0.219		22/128
+0.145		2/128
−0.016		1/128
−0.057		20/128
+0.036		36/128
+0.233		6/128
+0.032
50%16,236266,010,6242.76065835/128
+0.031		67/128
+0.047		47/128
+0.347		7/128
+0.024		3/128
−0.041		20/128
+0.036		59/128
+0.425		11/128
+0.071
75%24,353398,999,5522.65559232/128
+0.000		76/128
+0.188		57/128
+0.427		3/128
−0.008		2/128
−0.049		20/128
+0.036		48/128
+0.333		14/128
+0.095
100%32,470531,988,4802.62220032/128
+0.000		74/128
+0.156		53/128
+0.395		7/128
+0.024		3/128
−0.041		20/128
+0.036		42/128
+0.283		19/128
+0.135

The full figure, with its method and limits

Published research

Model Birth Observatory and Cognitive Birth

2026

A research program for recording model initialization, tracing checkpoint-level development, and measuring eight distinct cognitive axes under frozen and auditable decision contracts.

Research program · Program status; infrastructure verified, prediction not established

STOP-v1 Forensic Audit of Model-0

2026

A 26.65M-parameter model produced replicated signal on M, I, and A, but the frozen eight-axis gate correctly retained a STOP decision after difficulty-scale, format, multiplicity, and independence audits.

Technical report · Official frozen result plus post-hoc forensic audit

An Eight-Checkpoint Developmental Profile of Local-70M

2026

Matched diagnostic measurements from initialization to 531.99M training tokens show early A emergence, later I growth, non-monotonic axis trajectories, and a divergence between validation loss and the measured cognitive profile.

Experiment · Exploratory longitudinal measurement

Phenotype Shift During Continued Pretraining

2026

Continuing the same 69.97M-parameter model to 1.00B cumulative tokens improved validation loss while measured I and A declined and easy R increased, providing an exploratory example of non-uniform developmental change.

Research note · Exploratory continuation observation

All research, methods and limits