Observed Health
Uses successfully tested checks only. Skipped, unavailable, and failed-execution checks are excluded instead of becoming passes or zeros.
Show the formula
Σ(weight × score) / Σ(weight), tested checks onlyM01 · version 1.0.0Formula version 1.0.0 · zero AI calls
Evidence before persuasion.
We report what was observed, how it was measured, what remains unknown, and how to verify a change. A confident-looking score never replaces missing evidence.
One report, three different questions
Health, coverage, and confidence are related—but never interchangeable.
Keeping them separate prevents missing tests from looking like good performance and prevents strong evidence from sounding like a guaranteed business outcome.
Evidence Coverage
Every eligible check stays in the denominator. High coverage can still reveal poor health.
Show the formula
100 × tested eligible weight / all eligible weightM02 · version 1.0.0Finding Confidence
Weak fingerprintDirect deterministic
Evidence class, repeatability, and attribution quality determine confidence—not severity or hoped-for impact.
Show the formula
100 × Σ(weight × evidence coefficient) / Σ(weight)M03 · version 1.0.0Read the label before the number
Observed, scenario, verified, and unavailable mean different things.
- Observed
- Direct fixture evidence retained with a source, location, time, hash, rule version, and limitation.
- Scenario
- A user-entered what-if calculation. It is never relabeled as a measured result.
- Verified improvement
- A later before/after result that passes a defined acceptance test and guardrails. This draft claims none.
- Unavailable
- Evidence the scan could not inspect. It remains visible and is not treated as zero.
Why a health score may disappear
Low coverage suppresses the headline score.
If eligible evidence falls below the configured threshold, the interface shows scope and missing evidence instead of a polished health number.
Eligible evidence→Coverage check→Show health or explain suppression
Current draft policy: the overall fixture threshold is a documented 60% proposal. Approved category thresholds and weights were not supplied.
Confidence ladder
Evidence earns weight based on how directly it supports the finding.
- 1.00Direct deterministicObserved by a reproducible rule against retained evidence.
- 0.90Repeated traceConsistent across repeated instrumented runs.
- 0.75Field correlationSupported by field data, without claiming a single cause.
- 0.60Structured inferenceReasoned from explicit signals and shown as inference.
- 0.40Weak fingerprintA tentative clue that should lead to investigation, not certainty.
Traceability in the product
Every next step should lead back to retained proof.
Source IDs and hashes remain available in evidence details, while primary surfaces use plain merchant language.
- Page
- Template
- Test
- Evidence
- Finding
- Next step
Image transfer formula story
Start with exact bytes. End before the evidence runs out.
The controlled fixture compares 6.80 MB of current transfer with a 3.25 MB validated candidate. The measured difference is 3.55 MB.
Measured difference3.55 MB
Measured difference
current transfer bytes − validated candidate bytesThroughput scenario
saved bytes × 8 / effective throughput bits per secondMonthly origin scenario
saved bytes per view × monthly views × cache-miss fractionMB and GB use decimal units; exact integer bytes are stored. The 3.55-second result is a transfer-time ceiling at 8 Mbps—not an LCP prediction.
A change is not an outcome
The verification workflow closes the loop.
- Preserve the baselineKeep the original pages, evidence, formula version, and limitations.
- Make the lowest-risk changeUse a duplicate theme, reversible edit, or controlled implementation path.
- Repeat the same measurementUse the same rule, asset, page, and evidence boundary after deployment.
- Report the honest verdictVerified, no change, regression, or inconclusive all remain valid outcomes.
Claim boundary
What byte savings and public evidence do not prove
Byte savings alone do not establish a PageSpeed-score change, LCP change, conversion lift, additional orders, or revenue. Public evidence cannot establish indexing without verified Search Console access or guarantee visibility in an answer engine.
Source limits and corrections
The blueprint boundary remains visible.
The supplied material includes audit rules, metrics, calculators, a backlog, design tokens, and a contact sheet. Approved category specifications, controlled claims, benchmarks, page specifications, and production integrations described elsewhere were not supplied and are not invented here.
How to record a correction in this local draft
Preserve the scan ID, rule ID, evidence hash, expected source value, and reason in a repository issue or decision record. No external complaint or submission service is configured.
What remains intentionally disabled
Public crawling, private Shopify data, GA4, Search Console, Merchant Center, email, booking, authentication, analytics delivery, billing, and AI providers.