Report Outline

This page defines the renderer-neutral structure for modern InvarLock evaluation reports. It is the bridge between the canonical evaluation.report.json payload and human-readable renderers such as Markdown, HTML, evidence-pack summaries, and benchmark comparison pages.

The outline is implemented by invarlock.reporting.report_outline.build_evaluation_report_outline.

Why This Exists

The original report body was built as a linear Markdown document. That worked for early GPT-2/BERT preservation reports, but the repo now emits richer report shapes:

• policy failures, warning-mode guard movement, and strict warning policies
• causal, MLM, seq2seq, image-text, and MoE evidence lanes
• primary-metric tail checks and measured accuracy floors
• public assurance-basis reports with runtime manifests and model revisions
• guard-value evidence and benchmark-style bare-vs-guarded comparisons

Renderers should not decide this information architecture independently. They should render the shared outline.

Canonical Section Order

Section	Purpose	Typical source blocks
Decision	Overall verdict, evidence mode, model/edit identity, warning count.	validation, assurance, meta, primary_metric, guard_warnings
Primary Metric	Task metric, final value, baseline-relative comparison, CI, tail gate.	primary_metric, primary_metric_tail, validation
Policy Gates	Hard verify gates and thresholds.	validation, policy_digest, resolved_policy
Guard Signals	Guard observations and warnings separate from hard failures.	guard_warnings, invariants, spectral, rmt, variance, moe
Benchmark Comparison	Optional bare-vs-guarded scenario deltas.	benchmark_comparison, benchmark, guard_effect_benchmark
Evidence And Provenance	Dataset, windows, runtime/policy/provider digests, device, seed.	dataset, provenance, policy_digest, meta, artifacts
Technical Appendix	Verbose raw measurements, resolved policy, plugins, artifacts.	plugins, resolved_policy, policy_provenance, system_overhead, classification, structure, artifacts

The benchmark section is omitted when no benchmark block is present.

Renderer Rules

• Keep policy failures, guard warnings, and guard-value evidence distinct.
• Keep primary metric interpretation task-aware: ppl-like metrics use ratios; accuracy uses percentage-point deltas.
• Put benchmark deltas after guard signals, not in provenance or appendix.
• Keep verbose policy YAML, plugin provenance, and raw artifacts in the technical appendix unless they are needed to explain the verdict.
• Treat the outline as the source for visible section order in future Markdown and HTML renderers.