The Evidence Standard · 2026
When the regulator asks “show me,” have an answer.
Most tools can find personal data. The harder question is proving exactly what you searched for, what you found, and what you did about it on any given day.
PIISCAN 0.6.0 · EVIDENCE DATED 04 SEP 2026
01 · The stakes have moved
Your customer data is now a brand asset — and a brand liability.
Leads, support transcripts, scanned documents, forms, and images all collect personal data quietly. When obligations arrive, the question is not whether you take privacy seriously. The question is whether you can prove it.
Sincerity is not a defense. Evidence is.
02 · Finding data is easy. Proving it is the product.
A finding becomes evidence at the moment it is made.
Kestryl validates candidates against published rules, records a deterministic fingerprint of the rules in force, and chains actions so a changed entry cannot be quietly substituted after the fact.
A verdict, not a guess
Mathematical and checksum validation rejects values that cannot exist.
A fingerprint of the rules
Same rules, same signature, on any machine.
An entry that cannot be rewritten
Every action accrues to a chained audit trail.
03 · How it works
One controlled pipeline — from records and PDFs to a defensible evidence trail.
Kestryl reads structured records, free text, native and scanned PDFs, and images through PIIScan, OCR, and a locally deployed NER model. Discovery is read-only by default.
Remediation requires three separate approvals: the rule must allow it, configuration must permit it, and the operator must confirm it at run time, plus a confidence threshold you set. No single hand can trip it by accident.
New in this release: clean copies of the documents your teams use.
Word, Excel, PowerPoint, OpenDocument, email, archives, and supported plain-text/data formats can produce cleansed copies in a folder you designate. PDFs, images, scanned pages, and legacy Word .doc files are read or reported but not rewritten. Original files are never overwritten.
A finding in a scanned PDF is evidence you need even though the PDF is left for a human to redact. Nothing is quietly skipped, and nothing is quietly overwritten.
04 · Measured, not promised
Fast enough to check everything. Deliberate where it changes data.
Figures are PDI-reported measurements on synthetic data, benchmarked 04 SEP 2026 on a single processing thread of a standard cloud server. Your data, rules, and hardware will differ.
Four jobs, four numbers.
Documents included and remediation applied are separate decisions. The matrix below preserves the supplied release values, with only the disputed 12M scan timing normalized.
| Records | Documents | Scan records | Scan + documents | Remediate records | Remediate + documents |
|---|---|---|---|---|---|
| 1,000,000 | 1,000 | 0.63 | 4.18 | 20.41 | 53.78 |
| 2,000,000 | 2,000 | 1.16 | 8.17 | 41.43 | 106 |
| 4,000,000 | 4,000 | 2.11 | 16.01 | 81 | 217 |
| 8,000,000 | 8,000 | 4.12 | 31.84 | 160 | 432 |
| 12,000,000 | 12,000 | 6.06 seconds | 47.65 | 237 | 635 |
The 12M-row scan is reported at 6.06 seconds from PDI-reported measurements on synthetic data, benchmarked 04 SEP 2026. The complete matrix preserves the other supplied benchmark values exactly.
Other supplied measurements: RHEL 9.4 · Azure Standard_E32-16as_v5 · 16 vCPU AMD EPYC 7763 · 251 GB · single thread · tuned rule set.
05 · What you actually receive
A record of the run, not just a list of hits.
Every scan closes with an executive summary, an analyst extract, the rules fingerprint, confidence and channel information, refusals, and the values deliberately left alone. Original values do not appear in reports; where retained, they remain under separate control.
Low-confidence matches are found and recorded but remain outside automated masking. The discipline is not only in what Kestryl finds; it is in what it declines to act on.
06 · Operable by design
Everything you can change lives in a plain, versioned configuration.
A rule the engine does not understand does not run. New rules begin in report-only mode, and every edit changes the fingerprint so one configuration cannot be confused with another.
Optional name detection is a report-only heuristic where supplied. Audio/video support is limited to metadata or tags where supported, not spoken-content transcription unless explicitly available.
07 · Why Pacific Data Integrators
The evidence fits the systems your data actually lives in.
PDI brings more than fifteen years and a hundred-plus implementations inside enterprise data stacks including Informatica, Salesforce, Snowflake, and Databricks. Kestryl joins that data-quality portfolio with a goal that is simple to state and hard to fake: make the answer to “show me” immediate, complete, and reproducible.