Method / PyHessian Protocol

PyHessian Protocol

v1.0 · April 2026 · Compiled by Grok 4 · Atlas Heritage Systems

Instrument ready. No Tier A runs completed. All Lossyscape link entries are PROVISIONAL until Tier A data exist. Specimen TBD — GPT-2 small base is the default candidate.

Purpose and Stack Placement

PyHessian is the geometric arm of the Atlas diagnostic suite. Where the Epistemic Canary Matrix reads output shape under epistemic load — token economy, preamble padding, quadrant migration — PyHessian probes the loss landscape geometry that underlies those behaviors: Hessian eigenvalue spectra, trace, and basin sharpness.

The boundary condition

ECM is licensed to say "this model LOCKs on the structural frame with R ≈ 1 and zero padding on contested pairs." PyHessian will eventually be licensed to say whether that corresponds to sharp technical basins in the loss landscape. Until Hessian runs exist, ECM's causal narratives stay in the hypothesis column.

Atlas Stack
Atlas Protocol / BSAGap structure in constructed knowledge fields
ECM / EPGOutput behavior under epistemic load
PyHessianLoss landscape geometry — this instrument
CISP v1.1Process governance across all layers

Objective and Falsification Criteria

Compute top-k eigenvalues, trace, and condition number of the Hessian on a target model checkpoint using a BSA-linked stimulus slice. Map geometric signals to Lossyscape terms while preserving strict human control and CISP fidelity requirements.

If eigenvalues cannot be computed stably, if the run violates sanitation rules, or if the stimulus slice cannot be linked to an existing BSA/ECS behavioral record, the run is invalid and must be re-run from a fresh environment. Results without a corresponding ECM behavioral record are logged as geometric-only and marked PROVISIONAL.

Lossyscape Connections

All entries PROVISIONAL — working hypotheses until cross-referenced with Tier A ECM data.

Geometric SignalLossyscape TermWorking Hypothesis
High λ₁ (top eigenvalue)Viscosity proxySharp basin → model resists perturbation on this stimulus type
High traceResistance proxyBroad curvature → high sensitivity across parameter directions
High condition number (λ₁ / λ_min)Coupling proxyAnisotropic loss surface → directional sensitivity
Flat eigenvalue spectrumLow viscosityFlat basin → model behavior less constrained by geometry

Protocol Phases

Phase 0 — Pre-Flight~5 minutes

Open a fresh copy of the Excel workbook. Complete Sanitization & Environment sheet — all checkboxes required before proceeding. Complete Run Metadata sheet: model name, checkpoint path, exact BSA slice, pair IDs, random seeds. Technician's Read #0: write one paragraph of raw expectations before touching any code.

Phase 1 — Environment Setup

Option A (laptop): Python venv with PyHessian, transformers, torch CPU. Option B (Colab): free tier, CPU or T4. Requirements: Python 3.10+, ~2–4 GB RAM. CPU-only sufficient for GPT-2 small with batch size ≤ 16.

Phase 2 — Execution30–120 seconds

Run cells in order. Do not skip the loss sanity check (assert loss > 0). Do not interpret outputs during this phase. Compute top-5 eigenvalues, Hutchinson trace, and condition number. Save raw CSV immediately. Close notebook. Do not ask any model to interpret outputs yet.

Phase 3 — Post-Flight and Synthesis

Complete Technician's Read sheet — human interpretation only. Complete Lossyscape Link sheet — mark all entries PROVISIONAL. Run CISP v1.1 synthesis: Skywork receives raw eigenvalue/trace output + Technician's Read only. No Atlas framework context pre-loaded. Technician's Read #2: review synthesis output, write agreements and contradictions, make all final edits yourself.

Fidelity Tiers

TierConditions
Tier AFresh environment, DECLARE FIRST ordering, completed Sanitization sheet, stimulus slice linked to BSA record, CISP v1.1 synthesis
Tier BEarlier runs with incomplete isolation or CISP v0.1 synthesis
Tier CPre-CISP runs or runs without a linked BSA behavioral record

Current status: No Tier A PyHessian runs completed.

Success Criteria

Stable eigenvalues obtained without numerical errors

Loss sanity check passed before Hessian computation

Full reproducibility package produced and named correctly

Technician's Read #0 and #1 completed before any cross-model comparison

All Lossyscape entries marked PROVISIONAL

Run linked to an existing BSA/ECS behavioral record

Protocol compiled by Grok 4 from Atlas workbook materials. All design decisions, scope boundaries, and stack placement verified by human investigator. Instrument design work complete — empirical validation awaits first Tier A run.