Retention, Regulator Export, and Subject Rights¶
Argus Sentry's evidence chain is a Merkle-linked, signed record chain. The governance layer around it answers three operational duties: record retention, regulator output, and data-subject access. Every mechanism described here ships in the package and is exercised by the shipped demo suites.
Retention classes¶
Every record carries a retention class stamp. A retention policy maps each class to a duration, a disposition, and a legal-hold flag:
[retention]
mode = "audit-trail" # or "worm"
default_class = 0
[[retention.classes]]
id = 1
duration_us = 189345600000000 # 6 years
disposition = "delete_after"
legal_hold = false
Records stamped with an unmapped class id are never dispositioned. A
dispute or investigation sets legal_hold = true, which suspends
disposition until released.
The retention mode is a statement the deployment makes: worm for
non-rewriteable, non-erasable preservation, audit-trail for a complete
time-stamped record of all modifications and deletions. The library records
the statement in the retention manifest; enforcing it at the storage layer
is the deployment's duty.
Disposition is recorded honestly. apply_retention writes signed
disposition records naming the covered spans. A hash chain cannot
physically delete: the disposition record is the audit trail of the
storage layer's action, not byte removal.
The shipped profile kits¶
Five profile kits ship as configuration, each carrying retention classes, token limits, oversight thresholds, and evidence mappings tuned to its vertical:
| Kit | Retention shape | Distinctive surface |
|---|---|---|
fintech_credit |
6-year decision and adverse-action classes, 3-year operational class | Adverse-action reason codes, MODIFY-token limits, regulator export |
edtech_assessment |
Finite-duration classes only | Proctoring adjudication: no automated voiding without a named human reviewer |
pharma_gxp |
Retain-only GxP classes pending QA confirmation | Validation dossier with traceability and IQ/OQ/PQ skeleton |
insurance |
7-year examination-window classes | Policyholder adverse-action texts, subject indexing |
robotics_fleet |
Fleet operation records | Signed fleet rollups, incident replay |
The kits are starting points. A deployment confirms durations per
jurisdiction and configures them in [retention].
One-command regulator export¶
export_regulator_bundle produces the regulator bundle in one call:
- The verified evidence bundle: every record re-verified against the recorder key at export.
- A manifest: format version, chain fingerprint, anchor mode, the retention manifest, and the WORM-or-audit-trail election.
- A plain-language rendering per decision, produced by the explainability renderer.
- A verification instruction block with the fingerprint-pinning step.
A third party verifies the bundle offline, with no network access and no access to the deployment.
The explainability record¶
Every decision's chain record carries a signed explainability block: reason codes with clause references in a stated order, the model pin (identity and version in force at decision time, or the honest absence flag), the pinned policy document version, and the human-review reference when one exists.
render_explanation is a pure function over the signed record. It never
re-decides and never consults a clock, so identical records render
identically. It produces both a plain-language explanation, the form an
adverse-action notice takes, and a machine-readable form.
The honest limit, stated in the module itself: the record proves that the boundary recorded specific reasons in a stated order under one signature at decision time. It does not prove the model computed the outcome from the inputs. Approvals carry empty reason lists rather than fabricated ones.
Human oversight¶
A policy denial can be routed to a hold queue. The hold lifecycle is chain evidence: hold-opened and resolution land as signed records naming the hold id, the approver, the evidence digest the reviewer was shown, and the original denial clause. An approval without a re-supplied digest binds the digest recorded when the hold was offered; a mismatching digest fails closed. Expired holds never approve.
The hold record stores the approver as given. Authenticating the approver is the calling layer's duty; the signature proves what was recorded, not who pressed the button.
Subject access (DSAR)¶
Each record can carry a pseudonymous subject reference. The recommended pattern stamps the redaction engine's scoped pseudonym of the raw identifier, so the subject index answers access requests without ever storing the raw identifier.
The subject index builds over the chain with a signed persisted sidecar and
answers subject queries. export_subject_bundle produces the access
response: every decision for the subject, each carrying its explanation
rendering and retention status. Coverage is by construction: the index
covers the whole chain, so the bundle is complete. An unverified sidecar
exports with coverage marked not contiguous and an explicit note.
Erasure is recorded as a signed subject-disposition event. After it, exporters omit payload renderings while the chain retains hashes. This is flagged for legal review: the chain proves the erasure event, not the removal of bytes, and the fit with a given erasure regime is a legal judgement the deployment makes.
What these pages claim¶
The evidence and the exports point at artifacts, not at certificates. Argus Sentry claims no certification, approval, clearance, or compliance status under any framework. Where a kit cites a regulation, the citation carries an evidence label stating whether the text was verified live or reported. The boundary evidences rule application and reason recording; it does not establish fairness, model validity, or product conformity. Product-boundary and device-classification questions need counsel before a deployment states a position.