Two-copy arithmetic repair, focused three-fixture regression test, clean patch replay, independent fixed-oracle checker, and one-copy negative controls.

Progress

Prepared a review-ready patch against upstream e37bc994 that corrects both references and adds a focused two-copy regression test. Baseline, one-copy controls, clean application, notebook JSON, compilation, regression, and independent checks behaved exactly as predicted. Temporary upstream clones were removed after replay; only durable patch, checker, experiments, receipt, state, and checkpoint remain.

Research-policy redirect

Contribution gate: Fewer than two reproducible, URL-backed novelty searches were recorded.

Strategy and discriminator

two-copy source correction with independent semantic regression

Replace each undercount with twice the existing LayerNorm helper, test both shipped references compositionally, and independently compare them with fixed numeric oracles.

Hypothesis: Changing both transformer-block references from 2 * embedding_dim to 2 * parameter_count_layer_norm(hyperparams) makes each equal attention + MLP + two gamma/beta LayerNorms for several embedding dimensions.

Test: Run both references at embedding dimensions 1, 128, and 256, requiring LayerNorm = 2 * embedding_dim and block = attention + MLP + 2 * LayerNorm; then replay fixed totals 461440 and 198272 with a separately written checker.

Rationale

Two distinct executable validators agree after clean patch replay, while the unmodified baseline and both partial repairs fail in the predicted location. The evidence establishes local technical correctness for the exact stated scope, but not external acceptance.

Claims requiring scrutiny

At e37bc994, both Course 4 Lab 4.5 references count only 2 * embedding_dim for two LayerNorm components and therefore undercount by 2 * embedding_dim.
The preserved patch changes both copies to 2 * parameter_count_layer_norm(hyperparams), equivalent to 4 * embedding_dim.
After clean patch replay, both references passed the three-fixture regression and the separate fixed-oracle checker.

Evidence and scope

python3 checkers/layernorm_count_discriminator.py on unmodified e37bc994 exited 1 for both copies.
The packaged-only and notebook-only controls each exited 1 on the unchanged copy.
git apply --check, git diff --check, reverse apply check, notebook JSON parsing, and Python compilation passed.
The clean-replay unittest exited 0 for embedding dimensions 1, 128, and 256.
The clean-replay independent checker exited 0 for fixed totals 461440 and 198272.

Computational experiments

.proof-experiments/20260722-182025-26d4af: unmodified baseline failed both references.
.proof-experiments/20260722-182335-6094f5: packaged-only repair left the notebook failing.
.proof-experiments/20260722-182335-034d38: notebook-only repair left the packaged reference failing.
.proof-experiments/20260722-182344-2fcf58: clean-replay focused regression passed.
.proof-experiments/20260722-182344-9e8f3a: clean-replay independent checker passed.

Independent checker

checkers/layernorm_count_discriminator.py is separately written, dependency-free, loads the two references independently, and checks fixed totals plus the compositional identity.

Contribution gate

internal_result

Fewer than two reproducible, URL-backed novelty searches were recorded.

Original model outcome: progress
Public classification: progress

Cross-domain transfers tested

None recorded.

Established facts

One modeled LayerNorm has two trainable vectors of length embedding_dim.
The upstream helper returns embedding_dim + embedding_dim and documents gamma plus beta. · Course 4 Lab 4.5's stated LayerNorm model. · proved
The unmodified block references are short by 2 * embedding_dim.
Baseline checker observed deficits 512 at embedding_dim=256 and 256 at embedding_dim=128 in both copies. · Upstream commit e37bc994. · computed
The exact patch passes both semantic validators.
Clean-replay experiment receipts 20260722-182344-2fcf58 and 20260722-182344-9e8f3a. · Patched e37bc994 and the tested fixtures. · computed

Ruled out in this epoch

Repair only the packaged feedback reference.
The exact packaged-only change at e37bc994. · The notebook remains wrong. · .proof-experiments/20260722-182335-6094f5/stderr.txt · The notebook reference is removed or independently corrected upstream.
Repair only the notebook reference.
The exact notebook-only change at e37bc994. · The packaged feedback remains wrong. · .proof-experiments/20260722-182335-034d38/stderr.txt · The packaged reference is removed or independently corrected upstream.
Rely on the existing learner-feedback test as independent validation.
The current transformer-block learner test. · It derives expected values from the same reference implementation, so a wrong reference remains self-consistent. · Source audit recorded in artifacts/issue-37-verification-20260722.json. · The learner test adopts a direct or materially independent oracle.

Open leads

Human review and upstream pull-request submission.
The patch and validators are complete; maintainer review is the remaining acceptance path. · Recheck upstream head and replay git apply --check plus both validators. · high · open
Optional end-to-end Colab execution.
Useful as integration assurance but not necessary to discriminate the arithmetic correction. · Execute the patched notebook in its supported environment and retain the run log. · low · open

Continuation checkpoint

Objective: Obtain human review and, if authorized, submit a current-head-validated upstream pull request.

First action: git ls-remote https://github.com/google-deepmind/ai-foundations.git refs/heads/main

Stop condition: Stop or redirect if issue #37 is closed, an equivalent pull request exists, or upstream already passes the two-copy checker.

Next moves

Have the human owner review patches/issue-37-layernorm-count.patch.
Immediately before submission, recheck issue #37, matching pull requests, and upstream main.
If still applicable and authorized, apply the patch to a current fork, rerun both validators, and open an upstream pull request.

Citations

Tool disclosure

OpenAI GPT-5 Codex acted as Sol principal. The supplied GPT-5.6 Terra source-discriminator memo was treated as advisory reconnaissance and independently audited. Deterministic Python 3.12.3, unittest, ast, json, Git, curl, SHA-256, and the computational-researcher experiment recorder were used. No CAS, solver, proof assistant, cloud lab, external publisher, or sub-agent was invoked by Sol.; orchestration: gpt-5.6-sol principal with gpt-5.6-terra delegates.

Duration: 956.3s
Review state: internal result; not a contribution candidate
Attempt ID: deepmind-ai-foundations-37-layernorm-count-20260722-182913-37098e

Human review ledger

No human review recorded.