Representation Contract

A representation contract records how a raw object becomes the model input used for training and evaluation. It is the bridge between entity, modality, architecture, loss, metric, and leakage checks.

The core map is:

$$x_{\mathrm{raw}}\stackrel{\varphi }{\to }r\stackrel{f_{\theta }}{\to }\hat{y}$$

where $x_{\mathrm{raw}}$ is the raw object, $\varphi$ is preprocessing or featurization, $r$ is the model-ready representation, and $\hat{y}$ is the model output.

Contract Fields

Field	Question
Raw object	What is the entity before preprocessing: text, image, molecule, protein, pocket, complex, graph, or sequence?
Representation	What does the model see: tokens, tensor, graph, fingerprint, conformer, coordinate set, embedding, or fused view?
Unit	What is one token, node, edge, residue, atom, frame, candidate, or example?
Shape and axes	What are the batch, token, node, edge, coordinate, channel, head, or candidate axes?
Featurizer	Is $\varphi$ deterministic, learned, stochastic, cached, versioned, or data-dependent?
Invariance	Should the representation preserve order, permutation, translation, rotation, time, units, or metadata?
Missingness	What happens when fields, modalities, coordinates, labels, or alignments are missing?
Output	What output space does this representation support?
Loss and metric	Which loss trains the model, and which metric evaluates the output?
Split	Which unit must be held out to support the generalization claim?
Leakage	Does $\varphi$ use labels, target poses, future information, test statistics, or deployment-unavailable context?

AI Use

For AI architecture notes, the representation contract explains why an architecture is appropriate.

Representation	Likely Architecture Bias	Check
Token sequence	Transformer, RNN, state-space model	tokenization, context length, positional encoding
Image patches or feature maps	CNN, ViT, U-Net	locality, resolution, augmentation policy
Graph	GNN, Graph Transformer	node/edge definition, graph construction leakage
Set	Deep Sets, Set Transformer, Perceiver	permutation behavior and pooling
Coordinates	geometric GNN, equivariant model, diffusion/flow	transformation group and coordinate-frame leakage
Embedding	MLP, retrieval model, linear probe	frozen vs trainable encoder and evaluator capacity

Molecular Modeling Use

Object	Representation Questions
Molecule	SMILES, graph, fingerprint, descriptors, conformer, tautomer/protonation/stereo policy
Protein	sequence tokens, MSA, residue embedding, residue graph, coordinates, pocket representation
Ligand pose	conformer source, coordinate frame, atom mapping, symmetry handling
Protein-ligand complex	pocket definition, ligand state, interaction graph, distance features, pose availability
Assay record	target, assay, endpoint, unit, threshold, censoring, source

Formula View

If the representation is stochastic or learned, write:

$$r\sim q_{\psi }(r\mid x_{\mathrm{raw}},c)$$

where $c$ is context such as pH, assay, target, crop, alignment, retrieval corpus, or conformer protocol. If $c$ is unavailable at inference, evaluation may be invalid.

Checks

• Is the raw object separate from the model input?
• Are tensor shapes and axis meanings explicit?
• Is the representation identical across train, validation, test, and inference where it should be?
• Are featurizer versions, random seeds, caches, and failure policies recorded?
• Does the representation preserve the information required by the task?
• Does the representation include information unavailable at deployment time?
• Is the split defined before or after representation caching?
• Does the loss and metric match the output space implied by the representation?
• Does the architecture match the representation structure rather than only the paper’s keyword?

Related

• Modality representation
• Modality-task map
• Tensor shape notation
• Preprocessing contract
• Molecular featurization contract
• Protein representation
• Architecture selection
• Objective-metric alignment
• Claim-evidence boundary