1. Three evidence labels
| Label | Minimum meaning | Permitted interpretation |
|---|---|---|
| Observed change | A technology appears or disappears once from one collection path. | Candidate event only. It may be noise. |
| Probable migration | The directional change persists across time and multiple evidence channels. | Suitable for human review with limitations shown. |
| Verified migration | The old state is persistently absent, the new state persistently present, and retained evidence meets the governed threshold. | Publishable only with evidence date, confidence and a correction pathway. |
2. Collection boundaries
- Public, unauthenticated company-domain pages only.
- No login access, CAPTCHA bypass or technical-control circumvention.
- No collection or sale of individual contact information.
- Low-rate collection with site terms, robots controls and operational impact considered.
- No proxy infrastructure used merely to defeat normal access restrictions.
- Stop on authentication demand, explicit blocking or unclear permission boundaries.
3. Candidate verification
A candidate event should be checked across several public page types—such as homepage, product, collection and cart—and through both lightweight HTTP collection and browser rendering where permitted. A single homepage fetch is not a company technology stack.
Minimum review sequence
- Compare the candidate observation with a retained baseline.
- Recheck after a bounded interval rather than immediately declaring a migration.
- Compare multiple paths and evidence channels.
- Record disagreement between regions, devices, subdomains or collection methods.
- Require human review before a probable event can become verified.
4. Evidence channels
Research evidence may include timestamps, response headers, relevant HTML excerpts, script-resource changes, screenshots produced by the research system and fetch metadata. The evidence package should explain the conclusion without republishing unnecessary third-party material.
| Channel | Purpose | Known limitation |
|---|---|---|
| HTML structure | Detect platform-specific markup and route patterns. | Templates and headless storefronts can obscure the origin. |
| Script resources | Track platform and application assets. | Legacy scripts may remain after cutover. |
| Headers and network | Identify delivery and hosting changes. | CDNs can mask the underlying service. |
| Rendered behaviour | Observe consent-gated and client-rendered signals. | Experiments and localisation can create disagreement. |
5. Known false signals
- Consent mode preventing a technology marker from loading.
- Tag-manager injection, experimentation or temporary feature flags.
- Unused legacy code remaining after a genuine migration.
- Regional, subdomain, device or customer-segment differences.
- Temporary resource failure, CDN masking or collection error.
- A redesign or checkout change without a core platform migration.
6. Initial evaluation thesis
The proposed first corpus is mid-market ecommerce, focused on Adobe Commerce or Magento to Shopify or Shopify Plus. The commercial question is not whether a public signal wins the original replatform project—it may arrive too late—but whether it creates useful downstream action for adjacent software and service vendors.
7. Proposed research gates
These figures are targets for validation. They are not achieved or reported product performance.
| Metric | Proposed gate | Reason |
|---|---|---|
| Primary-platform detection precision | ≥ 98% | The baseline must be trustworthy before migration logic matters. |
| Verified-migration precision | ≥ 90% | False commercial triggers quickly destroy credibility. |
| Median detection latency | ≤ 72 hours | The event must remain commercially relevant. |
| False-removal rate | < 5% | Missing markers must not be mistaken for platform removal. |
| Manual verification time | ≤ 6 minutes | The evidence workflow must remain operationally viable. |
8. Correction principle
A company-domain observation is not an authoritative statement by the company. Any future public case must show its evidence date, confidence label and correction route. A neat archive is less important than an accurate one.