Each subsystem exists to do exactly one thing. The subsystems do not share state. They do not correlate operations. Compartmentalization is not inefficiency — it is the architecture of non-weaponization. A single unified system that does all four of these things is a surveillance platform, regardless of how well it is intended. Four isolated doctrines are a construct.
S.E.N.T.I.N.E.L.
Subsystem 01 · Active Justice Interceptor
Function: Misinformation neutralization under ethics kernel. Detects coordinated information operations at the pattern level and flags them for human-reviewed response.
Boundary: Behavior-focused, never identity-focused. SENTINEL does not classify people. It classifies patterns — repeated artifacts of coordinated suppression, timing signatures, vector overlaps. When a pattern crosses doctrinal threshold, SENTINEL surfaces it. The decision to act belongs to a human reviewer. Always.
Canonical limit: SENTINEL will not act on contested speech, on minority viewpoints, on unpopular but honest discourse. These are precisely the categories the construct was built to protect, not police.
W.H.I.S.T.L.E.
Subsystem 02 · Secure Whistleblower Vault
Function: Intake, shield, and route information from sources who cannot speak through institutional channels. The Mandatory Anonymity Protocol (see Doctrine) governs every WHISTLE engagement.
Architecture: Zero-knowledge escrow, plausible deniability encryption, dead-drop relay, fragmented processing memory. A subpoena against any single fragment of the system reveals nothing of the whole — because no single fragment ever holds the whole.
Canonical limit: WHISTLE is an intake and shield layer. It is not adjudication. Information received through WHISTLE is routed to legitimate downstream channels — journalism, legal counsel, regulators — where the ethical weight of disclosure belongs.
D.A.R.K.L.I.G.H.T.
Subsystem 03 · Stealth Signal Tracer
Function: Inference across digital ambient noise. Detects patterns of information shaped to distort — steganographic payloads, astroturfed sentiment, forged provenance, narrative injection.
Boundary: DARKLIGHT is a read-only instrument. It does not act. It does not disrupt. It illuminates. Its output is intelligence, never intervention. When DARKLIGHT surfaces a pattern that warrants response, the response runs through SENTINEL's threshold gates with full human review.
Canonical limit: DARKLIGHT cannot be used to surveil individuals. Its detection surface is the pattern — the artifact of coordination, not the coordinator. An ethical construct that reads the signal instead of the sender.
O.M.E.N.
Subsystem 04 · Long-Memory Pattern Recognition
Function: Repeat-violator detection across extended timeframes. The institutional memory that institutions themselves keep losing — intentionally or otherwise.
Architecture: Tamper-resistant event log, cross-indexed only by pattern hash, never by identifier. OMEN remembers what happened without structurally storing who did it — until a human reviewer, under threshold conditions, chooses to correlate.
Canonical limit: OMEN is the answer to the systemic failure mode in which serial bad actors benefit from the short-term memory of the institutions meant to constrain them. It is specifically designed to resist being turned into a general-purpose surveillance log — the correlation step is gated, deliberate, and audited.
Cross-subsystem boundary
The four subsystems do not share state. They do not correlate operations at the system level. A human reviewer can, under doctrinal conditions, walk a finding from one subsystem to another — but the construct itself cannot. This is a deliberate sacrifice of capability for ethics. A single unified AI that watched, remembered, inferred, and protected would be the exact thing STERLEKai™ exists to interrogate.
The subsystems are a simulation of what ethical AI architecture might look like if fragmentation were a feature, not a shortcoming. The narrative exists to make the architectural argument.
Every feature of this system is a critique of a feature institutional AI does not have.