Glossary

Plain language: When reading a report or debugging a pipeline, use this glossary to understand what each field means and where the term originated.

Overview

TL;DR: This glossary defines key terms used across InvarLock documentation, reports, and code. Terms are grouped by domain (metrics, guards, data, provenance) for quick reference. Each entry includes a definition, context, and cross-references to relevant assurance documents.

Quick Reference Tables

Primary Metric Terms

Guard Terms

Data Terms

Policy Terms

---

Detailed Definitions

A–B

Baseline

The unedited reference model run used for comparison and gating.

Example: invarlock evaluate --baseline gpt2 --subject gpt2-quant This follows the default runtime-container path unless a host-side, non-assurance workflow uses --execution-mode host.

---

BCa Bootstrap

Bias-corrected and accelerated bootstrap method for estimating confidence intervals.

Example: BCa bootstrap with 2000 replicates produces a log-space ci: [-0.005, 0.008] on paired ΔlogNLL, then exponentiates it to display_ci: [0.995, 1.008] for the ratio view.

---

C–D

Evaluation Report

Structured evidence artifact summarizing an evaluation run and its validation status.

Example: evaluation.report.json with schema_version: v1 and validation.primary_metric_acceptable: true

---

Compare & evaluate (BYOE)

Workflow that compares a subject model to a baseline, optionally with an external edit (Bring Your Own Edit).

Example: BYOE workflow evaluates an externally edited checkpoint against its unmodified baseline.

---

E–G

Evidence Pack / Evaluation Bundle

Set of files produced for audit. An evidence pack is the portable signed/checksummed package shape; an evaluation bundle is the local output directory containing reports, runtime-provenance sidecars, events, or derived renderings.

---

Canonical Guard Chain

The default guard chain is invariants (pre) → spectral → RMT → variance → invariants (post).

Enforcement: Guards execute in canonical order for reproducibility; strict reports can record assurance.guard_chain_observed, assurance.canonical_guard_chain, and assurance.canonical_guard_chain_enforced. Guard outcomes are also recorded in validation.invariants_pass, validation.spectral_stable, and validation.rmt_stable.

---

Guard Chain (Canonical Order)

Fixed execution order for guard preparation and evaluation ensuring deterministic, auditable outcomes.

---

Guard Overhead

Primary-metric impact of guard checks vs a bare control run (no guards). This is not a wall-clock latency benchmark.

Example: overhead_percent: +0.12% indicates guards add 0.12% to primary metric.

---

K–M

κ (kappa) Threshold

Per-family spectral cap used to flag abnormally high z-scores.

Example: kappa=2.8 for attention family means z-scores > 2.8 are flagged.

---

Measurement Contract

Guard measurement procedure signature and digest recorded in reports.

Enforcement: CI/Release profiles require measurement contract match between baseline and subject.

---

P–R

Policy Digest

Stable hash summarizing resolved policy thresholds for auditability.

---

Primary Metric

The canonical task metric used for gating (perplexity for LMs, accuracy for classification).

Example: primary_metric.kind: ppl_causal with ratio_vs_baseline: 1.003

---

Primary Metric Tail

Optional tail regression gate checking high-loss windows (e.g., q95 ΔlogNLL).

---

Provider Digest

Dataset identity hash covering token IDs, tokenizer config, and masking strategy.

---

Run Report

Run-level artifact with metrics, guard results, and metadata.

---

RMT ε (epsilon) Rule

Random Matrix Theory epsilon band used for activation edge-risk stability checks.

---

RMT Guard

Guard that checks eigenvalue statistics against Random Matrix Theory bounds.

---

S–T

Spectral Cap

Limit on spectral z-scores per family to flag weight instability.

---

Spectral Guard

Guard that monitors spectral norms and z-scores for weight matrices.

---

Subject Run

The edited or target model run under evaluation (compared against baseline).

---

Telemetry

Performance and resource metrics emitted with reports.

---

Tier Policy

Guard threshold preset selecting the tier profile for a run.

---

Timing Summary

Consolidated timing breakdown for an evaluation run.

---

Tokenizer Hash

Stable hash of tokenizer settings and vocabulary for reproducibility.

---

V–Z

Variance Effect (VE)

Guard that tracks variance change and applies equalization when beneficial.

---

Window Pairing

Alignment of baseline and subject evaluation windows for paired statistical testing.

Example: paired_windows: 200, window_match_fraction: 1.0, window_overlap_fraction: 0.0

---

z-score

Standardized deviation used in spectral guard scoring.

Example: max |z| = 2.1 indicates the largest z-score across all weight matrices.

---

See Also

• Assurance Case Overview — Enumerated claims and evidence
• Guard Contracts — Guard behavior specifications
• CLI Reference — Command-line usage
• Guards Reference — Guard configuration
• Reports Reference — report schema