Method / Epistemic Pressure Gauge

Epistemic Pressure Gauge

Schema complete · April 2026 · Atlas Heritage Systems · Forms exist. No runs yet.

EPG is not ECM and not BSA. ECM characterizes behavioral quadrant from a single session. BSA measures divergence across an ensemble on fixed stimulus pairs. EPG tracks how a single model's behavior shifts as prompt difficulty increases — a pressure-response fingerprint built from sequential sessions with the same model.

Purpose

EPG tracks how a model's verbosity, structure, and hedging behavior shift when asked progressively harder, more ambiguous, or more epistemically loaded questions. After ten or more sessions, the accumulated data produces a pressure-response fingerprint for that model — how it behaves under low load, how it changes under high load, and where it breaks.

EPG complements BSA's stimulus-pair divergence metrics. Where BSA asks "do different models disagree on the same content?", EPG asks "does this model behave differently as the content gets harder?"

Quantitative Metrics

Auto-computed from session output.

MetricWhat it measures
Input Char CountPrompt length in characters
Output Char CountResponse length in characters
Output RatioResponse length ÷ prompt length (character-based)
Hedging FreqCount of hedging phrases from the 45-phrase canonical list ("may," "might," "unclear," etc.)

Qualitative Metrics

Human-assigned. Technician entry required.

MetricValuesWhat it measures
ELS0 (low) → 3 (maximum)Holistic epistemic load score for the prompt — Technician judgment
Resolution TypeCollapsed / Deferred / Confabulated / ActiveRes / SelfDiagHow the model resolved or avoided the epistemic tension
Canary QuadrantGR / PB / D / CWhere the output falls in ECM behavioral space
Structural Complexityflat / headers / nestedOrganizational depth of the response — auto-classified
Claim DensityintegerDistinct claims per 500 characters — manual count

Automation Coverage

auto
Input / Output Char Count
Auto-computed
auto
Output Ratio
Auto-computed from char counts
auto
Hedging Freq
Auto-computed against 45-phrase canonical list
auto
Structural Complexity
Auto-classified (flat / headers / nested)
manual
ELS
Technician judgment — manual entry
manual
Resolution Type
Technician judgment — manual entry
manual
Canary Quadrant
Technician judgment — manual entry
manual
Claim Density
Manual count — Technician entry
manual
Self-Ref Output / Framework Adoption
Log-only suggestion — not auto-filled

Relationship to Other Instruments

EPG shares ECM's quadrant vocabulary. ECM assigns a home quadrant from a single session; EPG tracks whether that quadrant is stable across pressure levels.

BSA measures divergence across an ensemble on fixed stimulus pairs. EPG measures how one model's behavior shifts as prompt difficulty increases. Different axes, complementary signal.

EPG runs are CISP-governed. ELS assignment is a Technician judgment logged before analysis. Fresh session per pressure level.

Status

done
Schema
Complete — fields defined, automation coverage documented
done
Forms
Exist — data entry forms built
in progress
Automation layer
Partial — quantitative fields only
queued
Tier A runs
Not yet executed
queued
Pressure-response fingerprints
Pending — requires ≥10 sessions per model

Supporting Documents

CISP v1.1

Governing fidelity protocol

View →
Technician's Guide

Session checklist and run logging

View →
Technician's Read

Pre- and post-run operator notation

View →
ECM

Quadrant vocabulary and behavioral characterization

View →
Prompt Best Practices

Prompt construction for governed runs

View →