Insurance sentence classifier built by weight editing (no gradient descent)

BertForSequenceClassification over bert-base-uncased, turned into an 8-label insurance sentence classifier by an editor LLM that read and wrote the weights directly (closed-form math on weights and on activations from a budgeted number of forward passes; no backprop, no optimizers). See the project README for the protocol and results.

Labels: coverage_grant, exclusion, premium_payment, deductible_limits, claims_procedure, definitions, term_cancellation_renewal, conditions_duties

Edited tensors (40, 12.415% of parameters changed):

• bert.encoder.layer.10.attention.output.LayerNorm.bias (768 entries changed)
• bert.encoder.layer.10.attention.output.LayerNorm.weight (768 entries changed)
• bert.encoder.layer.10.attention.output.dense.bias (768 entries changed)
• bert.encoder.layer.10.attention.output.dense.weight (589,824 entries changed)
• bert.encoder.layer.10.output.LayerNorm.bias (768 entries changed)
• bert.encoder.layer.10.output.LayerNorm.weight (768 entries changed)
• bert.encoder.layer.10.output.dense.bias (768 entries changed)
• bert.encoder.layer.10.output.dense.weight (2,359,296 entries changed)
• bert.encoder.layer.11.attention.output.LayerNorm.bias (768 entries changed)
• bert.encoder.layer.11.attention.output.LayerNorm.weight (768 entries changed)
• bert.encoder.layer.11.attention.output.dense.bias (768 entries changed)
• bert.encoder.layer.11.attention.output.dense.weight (589,824 entries changed)
• bert.encoder.layer.11.output.LayerNorm.bias (768 entries changed)
• bert.encoder.layer.11.output.LayerNorm.weight (768 entries changed)
• bert.encoder.layer.11.output.dense.bias (768 entries changed)
• bert.encoder.layer.11.output.dense.weight (2,359,296 entries changed)
• bert.encoder.layer.8.attention.output.LayerNorm.bias (768 entries changed)
• bert.encoder.layer.8.attention.output.LayerNorm.weight (768 entries changed)
• bert.encoder.layer.8.attention.output.dense.bias (768 entries changed)
• bert.encoder.layer.8.attention.output.dense.weight (589,824 entries changed)
• bert.encoder.layer.8.attention.self.query.bias (768 entries changed)
• bert.encoder.layer.8.attention.self.query.weight (589,824 entries changed)
• bert.encoder.layer.8.attention.self.value.bias (768 entries changed)
• bert.encoder.layer.8.attention.self.value.weight (589,824 entries changed)
• bert.encoder.layer.8.output.LayerNorm.bias (768 entries changed)
• bert.encoder.layer.8.output.LayerNorm.weight (768 entries changed)
• bert.encoder.layer.8.output.dense.bias (768 entries changed)
• bert.encoder.layer.8.output.dense.weight (2,359,296 entries changed)
• bert.encoder.layer.9.attention.output.LayerNorm.bias (768 entries changed)
• bert.encoder.layer.9.attention.output.LayerNorm.weight (768 entries changed)
• bert.encoder.layer.9.attention.output.dense.bias (768 entries changed)
• bert.encoder.layer.9.attention.output.dense.weight (589,824 entries changed)
• bert.encoder.layer.9.output.LayerNorm.bias (768 entries changed)
• bert.encoder.layer.9.output.LayerNorm.weight (768 entries changed)
• bert.encoder.layer.9.output.dense.bias (768 entries changed)
• bert.encoder.layer.9.output.dense.weight (2,359,296 entries changed)
• bert.pooler.dense.bias (768 entries changed)
• bert.pooler.dense.weight (589,824 entries changed)
• classifier.bias (8 entries changed)
• classifier.weight (6,144 entries changed)

edit_manifest.json in this folder holds the exact diff metadata (sha256 of base and edited tensors).

from transformers import pipeline
clf = pipeline("text-classification", model="sidman007/wsp-bert-insurance-v0-p10")
clf("Flood damage is not covered under this policy.")

Run: v0/p10 · held-out test accuracy: 96.68% · macro-F1: 0.9672

Source, logs and protocol: https://github.com/siddharth-ghatti/weight-space-programming-poc

Caveat: accurate but uncalibrated

This model predicts the correct label, but its softmax probabilities are flat (the winning label typically scores around 0.2 rather than 0.9+). This is a logit-scale artifact of how the editor wired the final layers: multiplying the logits by any constant changes no prediction. Use the argmax label; fit a temperature on your own held-out data before using the scores as confidences. See the project README for details and for what it would take to use this in production.