CASE WORKSPACE

Settings, Privacy & Ethics

Operating posture for the console, and the constraints under which its assessments may be used.

Controlled demonstration

Operating mode

Demonstration modeDeterministic controlled cases, fully offline. No external service is contacted.

With demonstration mode disabled, module adapters would dispatch to the configured containerised model services. Those services are not attached in this build, so results remain labelled as a controlled demonstration.

Role-based access control

Permission
media.metadata.view	ALLOW	ALLOW	ALLOW
media.content.view	ALLOW	ALLOW	DENY
analysis.run	ALLOW	ALLOW	DENY
review.submit	ALLOW	DENY	DENY
review.approve	DENY	ALLOW	DENY
mitigation.execute	DENY	ALLOW	DENY
report.finalize	DENY	ALLOW	DENY
report.export	ALLOW	ALLOW	ALLOW
audit.view	ALLOW	ALLOW	ALLOW

Media content is private storage: only roles with media.content.view receive a signed read, and every grant or denial is written to the case audit trail.

Reviewer sign-off SLA policy

Window an authorising reviewer has, per case priority

CRITICAL priority

hours

HIGH priority

hours

MEDIUM priority

hours

LOW priority

hours

When the window elapses without a decision, the case is flagged as breached and can be escalated with a reason code. Escalation never unlocks mitigation or report finalisation — only an authorised human sign-off does.

Reason codes

RC-A1Evidence supports assessmentAPPROVEObserved evidence independently verified and consistent with the model output.
RC-A2Approved with residual uncertaintyAPPROVEAssessment authorised, but reviewer recorded unresolved uncertainty in the notes.
RC-A3Downgraded on reviewAPPROVEReviewer accepts the case but judges the model risk band too strong.
RC-R1Insufficient evidence linkageRETURNConclusion is not traceable to specific evidence objects.
RC-R2Unresolved cross-modal conflictRETURNContradicting modality signals were not addressed before submission.
RC-R3Chain-of-custody gapRETURNProvenance or hash continuity is incomplete for one or more assets.
RC-R4Re-run requiredRETURNA degraded module invalidates part of the fused assessment.
RC-E1SLA breached — no reviewer actionESCALATIONSign-off window elapsed without an authorising decision.
RC-E2SLA breached — reviewer unavailableESCALATIONAssigned reviewer could not be reached inside the window.

SLA breach notifications

Where a breached reviewer sign-off window is announced

In-app notificationsRaise the breach in the console notification tray with reason code and recommended escalation.

Email the duty reviewerComposed server-side. No mail transport is attached in this build, so alerts are recorded as queued rather than sent.

WebhookServer-side POST of a JSON breach payload (case, breached item, reason code, recommended escalation) to your incident endpoint.

Breaches are re-scanned every 15 seconds while the console is open. One notification is raised per case per escalation level, and every dispatch attempt is written to the case audit trail.

Publish gate — URL health check

Runs automatically before publish. Sitemap, robots.txt and canonical must pass for the live hostname or the final publish step is blocked.

Security posture

— Private media storage — no publicly addressable assets
— Authenticated access with per-case row-level policies
— All decoding and inference server-side in controlled workers
— Secrets held server-side only; never present in frontend code
— Validated, typed API inputs with structured errors
— Rate limiting on analysis dispatch and export endpoints
— No arbitrary code execution; no uncontrolled public-source scraping
— Every read, action and export written to the audit trail

Data minimisation & retention

— Only metadata required for the assessment is extracted; personal fields are suppressed from the interface.
— Default retention is 90 days under a restricted forensic policy; extensions require approval.
— Exports carry an integrity manifest and are recorded against the requesting analyst.
— Deleted cases purge derived artifacts (frames, spectrograms) along with the source asset.

Model limitations & false positives

— Heavy re-encoding, low resolution and narrowband audio all degrade detector reliability.
— Benign editing (stabilisation, noise reduction, transcoding) can produce indicators that resemble manipulation.
— Absence of evidence is reported as INSUFFICIENT EVIDENCE, never as proof of authenticity.
— Provenance that cannot be verified is UNVERIFIED, which is not the same as manipulated.

Responsible use

Suryakavacham assesses media integrity. It does not identify wrongdoing, does not infer intent and does not attribute legal responsibility to any person. Outputs are probabilistic, uncertainty-bearing and intended to support — never replace — a qualified human investigator.

An AI assessment is not proof of wrongdoing.

Human review is required before any case leaves this console. Detect the signal. Correlate the evidence. Explain the risk. Support the investigator.