---
title: Reporting Contract
published: true
index: 10
tier: free
---

# Reporting Contract

Apply this host-neutral contract to every focused and comprehensive audit.

## Output mode

Follow [Glue JSON output](glue-output.md) and the bundled shared schema for the complete canonical `report.json`, validation and local-chat presentation. Use the host-supplied schema when pinned; otherwise the portable bundled schema is available. The setup fence is not the result schema. Keep the JSON report free of Summary/Findings output and preserve all selected outcomes. Glue receives its bare result object. After local validation, follow [local audit results](../workflows/local-results.md) to open the bundled localhost viewer, then summarize the saved JSON in chat without duplicating its full ledger. The viewer is a derived presentation; its availability never changes Check outcomes or the canonical report.

The collections below describe meaning and eligibility, not an alternative serialization. An explicitly requested Markdown rendering is derived from the canonical JSON and may use Issues, optional Changes, Checks and Limitations sections without inventing a Summary or new result collection. Keep strengths empty; neutral observations, evidence ledgers and diagnostics belong in existing detail fields or linked supporting artifacts.

## Issue inclusion and quality gate

An **Issue** is the canonical representation of an evidence-backed action, decision, or bounded experiment.

Create an Issue when target-bound evidence establishes at least one of:

- a selected check with terminal `Fail`;
- a meaningful experience, accessibility, discoverability, performance, implementation, or reliability gap requiring change; or
- a bounded experiment with a specific hypothesis, primary metric, and guardrails.

Every failed check maps to exactly one canonical Issue in the same top-level catalog category. Consolidate failed checks only when they share the same category, root cause, action, and retest; retain every contributing check ID. Keep causes or categories separate when their owner, remediation, or retest differs. Do not create Issues for passing checks, praise, resolved work, generic maintenance advice, or monitor-only suggestions.

For a funnel experiment, include the affected step and segment, current and comparison baselines, evidence quality, the funnel diagnosis, a falsifiable hypothesis, the single proposed change, primary metric, guardrails, and a decision or retest condition. Do not state a predicted conversion lift unless a target-owned experiment or comparable prior result supports it. Analytics alone may identify an opportunity but cannot establish the proposed cause.

Use the `Experience` Issue category for a hypothesis-led funnel experiment. A directly reproduced implementation, accessibility, discoverability, or experience defect stays in its owning category even when funnel evidence raises its priority. Never duplicate one action across categories.

Remove a proposed Issue when it names no meaningful action, its evidence only describes a strength or neutral fact, or completing it would create no material change. This removal pass must never hide a failed check.

A `Pass` result cannot carry a corrective suggestion. If the evidence supports action, correct the terminal outcome or create a separately supported experiment Issue. When no Issues qualify, write `None.`

## Issue content and order

Use a short action-led title and exactly one catalog category: `Experience`, `Accessibility`, `Discoverability`, or `Technical`. Never use a subcategory, storefront surface, route, or evidence topic as an Issue category.

Keep each Issue concise while preserving:

- stable Issue ID and every contributing check ID;
- P0-P3 priority and evidence state;
- exact target and the smallest reproduction or source evidence needed to verify it;
- customer or business impact;
- one concrete recommended change or bounded experiment; and
- one acceptance condition or exact retest.

Rank supported opportunities using [PIE](prioritization.md), retaining P0–P3 severity and urgency. Keep canonical category groups contiguous; within comparable urgency and dependency groups apply PIE ordering. Preserve each score's evidence rationale, confidence, missing evidence and overrides. Add ownership, dependencies, screenshots or broader reach only when they change the decision.

## Screenshot evidence

Capture a screenshot when it directly proves a browser-visible Issue or comparable Change. Record the exact route, state, viewport, and timestamp. Never invent or attach decorative evidence. When capture fails, preserve the best DOM, layout, console, network, or source evidence and block only checks for which the missing image prevents a responsible conclusion.

## Changes

Always look for the most recent prior report for the same configured target. Include Changes only when prior and current evidence are comparable in target, state, route, viewport, method, environment, metric definition, and relevant revision.

Allowed directions are `Improved`, `Regressed`, and `Resolved`. Newly observed, intentionally redesigned without an evaluative delta, unchanged, and incomparable items are not Changes. A regressed item that requires action also remains an Issue; a resolved item does not. If no compatible baseline exists, omit Changes rather than implying comparison.

## Checks

Checks is the exhaustive ledger of selected active checks, not an execution log or a visible praise section.

- Preserve each selected check's stable ID, title, category, and version.
- Assign exactly one terminal status: `Pass`, `Fail`, `Blocked`, or `Not applicable`.
- Group entries by category for readability, while listing every selected check individually.
- Include direct evidence or concise result detail for every outcome; retain useful Pass, Not-applicable, measurement, and diagnostic evidence here instead of converting it into Findings.
- Keep Finding references empty.
- Link each Fail to exactly one canonical Issue.
- Keep each Issue and all of its contributing failed checks in one top-level category; do not consolidate failures across categories.
- Link each Blocked outcome to one or more shared Limitations.
- For each Blocked outcome, name the precise required context or state that could not be obtained and why available repository or storefront evidence was insufficient. A generic statement such as `check-specific evidence unavailable` is not enough.
- Use `Not applicable` only after verifying that the feature, condition, or customer job is absent. Missing evidence is Blocked, not Not applicable.
- Never add parent rollups, browser/tool rows, repository-command rows, viewport rows, or journey rows as Checks entries.

Before returning, compare the selected ID set to Checks. Reject missing, duplicate, unknown, or repeated IDs.

## Limitations

Record each shared blocker once with a stable ID, title, exact diagnostic, smallest unblock action, and every affected check ID. A Limitation may affect only checks whose required decision evidence it actually prevents. Sampling breadth, screenshot capture, comparison incompatibility, or disposable-state cleanup may be audit-wide notes, but they do not force check outcomes unless the individual check requires them. Limitations describe why responsible conclusions were impossible; they do not duplicate Issue prose or count as Issues.

When no selected check is blocked, write `None.`

## Final reconciliation

Before output:

- when the host checkpoints batch outcomes and requests only candidate grouping, treat every checkpointed status and check detail as immutable and let the host expand the final ledger;
- every selected ID appears exactly once with one terminal status recorded during bounded-batch evaluation;
- Findings and Finding references are empty;
- neutral measurements, passing inventories, diagnoses without justified action, and unsupported causal claims remain in Checks or supporting metadata;
- every Fail links to exactly one existing Issue;
- every Issue and all referenced Fail rows share one top-level category, so each category with failures remains visible;
- every Blocked check and Limitation reference one another and name the unmet evidence plus safe acquisition attempts;
- no passing or not-applicable check carries a corrective recommendation;
- shared root causes and blockers are deduplicated without merging different actions or retests;
- Changes use only compatible before/after evidence; and
- partial access is explicit without stopping independent checks.
