Reference-context engagement scorer
Tester access. Enter the shared password to continue to the scorer, model mathematics, demo video, and contact sheets.
Sessions expire after 30 minutes. Password attempts are rate limited.
Tester access. Enter the shared password to continue to the scorer, model mathematics, demo video, and contact sheets.
Sessions expire after 30 minutes. Password attempts are rate limited.
One photograph, placed in a measured reference population of 91,810 posts. The instrument reports where it falls.
Score the same crop separately with selected learned college-page embeddings. All other reference-context fields remain fixed.
Drop a photo here, or choose one. If more than one person is in frame, one is selected and scored.
Nothing is stored. The image is scored and discarded.
The live scorer runs on a single machine which is not running right now. The recording below shows the same path a live upload takes.
Ensemble spread —.
These are context interventions on a fold-standardized score axis, not college-specific percentiles.
These contact sheets were produced partway through training. They reflect an earlier, cruder version of the system than the final product and the model featured on this page. Treat them as development diagnostics, not as the finished output.
The scorer maps frozen DINOv3 patch representations and structured posting context to a scalar estimate of log engagement. The following defines the 512-pixel configuration used by each member of the demonstrated ensemble.
After channel normalization, the image is passed through a frozen DINOv3 encoder. The classification and register tokens are discarded. The scoring head receives only the 1,024 spatial patch embeddings:
The cached classification token is not used by the scoring head.
Six scalar covariates are log-transformed, standardized, and assigned learned feature-specific values when missing:
The transformed scalars are encoded by a 6–32–32 MLP. Its output is concatenated with learned embeddings for page, gender, carousel position, carousel length, and person count, together with source-quality covariates and missingness indicators. A final affine projection produces the conditioning vector
Each pooling block contains separate adaptive normalization modules for its query, patch sequence, and feed-forward input. For , LayerNorm is evaluated independently at every sequence position over the 768 feature coordinates:
Context is converted to a featurewise scale and offset and broadcast across sequence positions:
Each modulation is a 128-to-1,536 affine map. Its parameters are zero-initialized, so AdaLN initially reduces exactly to non-affine LayerNorm.
A learned query is expanded across the batch. In each of two blocks, 12 heads of width 64 compare this singleton query with all patch positions. For head h,
Here and . Softmax is taken over the 1,024 patch positions.
The 12 head outputs are concatenated and affinely projected. Attention-weight dropout is applied during training. The resulting vector is added to the unmodulated residual query:
The attention residual is independently modulated and passed through a 768–3,072–768 feed-forward network with GELU and hidden-state dropout:
Both blocks attend to the original frozen patch matrix X. The second block therefore revisits every patch using the query state produced by the first block.
After the second block, the length-one query axis is removed. No additional normalization or activation precedes the image-path output:
The context-only branch contains no image information. Both outputs are unrestricted real-valued scalars.
At inference, nuisance covariates are fixed to a shared reference context. Three fold checkpoints evaluate the upright and horizontally reflected image. Each of the six scores is standardized against its checkpoint-specific held-out distribution before aggregation:
The empirical distribution is the held-out calibration population for gender reference g. The displayed ensemble spread is the range of the six member percentiles; it is not a confidence interval.
One image-only head may be selected in place of the ensemble. It receives the same frozen patch matrix X and a learned gender embedding at the teacher's context width, and nothing else: no page, no follower count, no carousel slot. The reference context is not removed but inherited, because the quantity it was trained to predict is the ensemble's own reference-context score.
Writing for that teacher score, training standardizes it within gender and the student predicts the standardized target under a Gaussian likelihood whose variance floor is the teacher's own inter-checkpoint spread:
A label the three checkpoints disagree about is a label the student is not penalised for missing. The predicted uncertainty is an internal quantity and is never reported as a statement about the photograph.
At inference the target standardization is inverted, which places the student back on the ensemble's axis, and the same empirical distribution supplies the percentile:
The student has no calibration population of its own, and an estimate is less dispersed than the quantity it estimates, so its percentiles sit closer to the median than the ensemble's for the same photograph. That compression is measured on held-out first-image single-person posts and reported with every student score, together with the acceptance cell the student fails.
Predicted engagement for this image posted solo, as the thumbnail, by a standardized account, under this population's revealed preferences, with variables such as page reach, follower count, posting history, carousel position and source compression held fixed.
Without it, a score mostly measures how many followers an account has. The reference context pins every one of them to the same value for every image, so what varies between two scores is the photograph.
The ensemble band is typically wide, a median of 39 percentile points between the lowest and highest member. The band is shown at the same weight as the estimate because it is the same size as the estimate.