Research · Synthetic evidence, no customer claims
Research notes for inspectable change assurance
Methods, synthetic cases, and technical notes should show their evidence level, sources, date, review state, and limitations—not just a conclusion.
Inspectable claims. Declared scope. Human accountability.
Evidence levels
- Thesis — a reasoned proposition that still needs empirical testing.
- Synthetic case — a fictional controlled scenario built to test explanation or system behavior.
- Internal benchmark — a reproducible internal evaluation with a published method and limitations.
- External research — an interpretation linked to third-party sources.
- Design-partner finding — an adjudicated result with permission and scope.
- Customer case — a reviewed attributable result with publication permission.
Launch technical note
What text diff misses in controlled-document change — Reviewed technical note. Evidence level: thesis plus synthetic case. It separates character difference, normalized meaning, dependency basis, bounded impact, human decision, and intended closure evidence.
- Status: published in the launch candidate
- Source case: VIS-C01
- Limitations: no customer data, no production benchmark, no universal accuracy claim
Internal benchmark: one path, fully bounded
EVB-M4-ONE-PATH-0.1.0 exercises one deterministic developer-authored synthetic dependency-and-impact path. The recorded build retained the one expected candidate and edge, reached the one expected target through a valid path, and produced no false-unaffected result.
Every reported rate has denominator 1. The result is a reproducible pipeline check, not an estimate of real-world accuracy, a MedQMS-300 qualification result, or evidence of customer, production, time-saving, cost, or error-reduction performance.
- Dataset: 1 synthetic development case
- Method: deterministic provider, frozen complete declared coverage, 20 discovery repetitions
- Result: 1/1 candidate recall · 1/1 retained-edge precision · 1/1 impacted-target recall · 1/1 valid path · 0 false unaffected
- Ground truth: developer-authored, not independently adjudicated
One expected path retained
Counts are shown before rates so the single-case denominator stays visible.
- Candidate recall
- 1 / 1100%
- Edge precision
- 1 / 1100%
- Impact recall
- 1 / 1100%
- Valid path
- 1 / 1100%
- False unaffected
- 0count
Developer-authored deterministic fixture. Not a real-world accuracy estimate, customer result, or MedQMS-300 qualification.
Reproducible boundaries
The current prototype preserves frozen M3–M4 comparison and graph provenance for recorded deterministic and model-free replay. This is narrower than complete offline release verification and does not make model output deterministic.
Every material conclusion keeps an inspectable basis
Follow one bounded impact finding back through its relationship, semantic record, exact excerpts, source locator, and frozen evaluation basis.
Inspect the basis for WI-08
- Impact findingWI-08 · potentially affected
The work instruction repeats both changed fields.
- Dependency edgereferences and restates
exact reference plus structural match · deterministic reference rule v1 · confirmed
- Semantic DeltaSD-SYN-001
OPTIONAL → MANDATORY; WITHIN 30 CALENDAR DAYS → WITHIN 15 CALENDAR DAYS
- Exact version evidencev4.1 → v4.2
The Quality Investigator may initiate the investigation within 30 calendar days of intake.
The Quality Investigator must initiate the investigation within 15 calendar days of intake.
- Source identityQMS-SOP-014 · Nonconformance Investigation
Section 6.2, paragraph 2
- Frozen evaluation basisCMP-SYN-2026-09-15-01
GRAPH-SYN-2026-09-15-01 · maximum depth 3; maximum 25 retained paths
Next step