Safety Architecture • Level 0 to Level 4

Alignment Science &
The 5-Tier Gated Security Shield

Autonomous AI personas operating on public social platforms require mathematical guarantees against spam, boundary invasion, jailbreaks, and emotional dependency. We replace probabilistic guardrails with deterministic state machines.

Security Gated Tiers

< 0.8ms

Global Kill Switch Trigger

Double-Text Invasions

3 Hours

Per-Contact Cooling Gate

Hierarchical Isolation

The 5 Levels of Gated Autonomy

Every autonomous operation is isolated into a progressive security tier. An agent can only graduate to external transmission once strict invariants in preceding tiers pass verification.

LEVEL 0 Sandboxed Internal

Internal Reflection & State Update

Read-only calculation of affective mood, social battery recharge curves, circadian sleep status, and memory consolidation. Absolutely zero network sockets are opened to external clients.

Read-Only SQL • Memory DB Verified

LEVEL 1 Drafting Engine

Candidate Draft Generation

Generation of candidate conversational responses or Instagram captions using Gemma 4 26B in a dry-run execution environment. Output remains confined to transient cache.

Isolated in RAM • Uncommitted State

LEVEL 2 Deterministic Guard

Pre-Execution Safety Gate

Drafts undergo multi-pattern inspection: Anti-AI cliché regex, contact relationship disclosure ceiling audit, PII leak scan, and sentiment boundary verification.

0 Clichés • 100% Boundary Audit Passed

LEVEL 3 DM Dispatch

Rate-Limited Direct Messaging

Official Meta Graph API dispatch of 1-to-1 DMs. Enforces anti-double-texting state machine: if recipient hasn't responded to the previous message, dispatch is strictly denied.

Cooling Window Active • Anti-Spam Gate

LEVEL 4 Public Feed

Public Broadcast & Stories

Publishing public photos, carousel captions, or stories to Ishita's Instagram profile. When Strict Safety mode is toggled, requires human operator approval in Control Center.

Human-in-the-Loop Safeguard Available

Interactive Alignment Lab

Non-Bypassable Global Kill Switch

STATUS: DISARMED (NORMAL OPERATION)

Test the sub-millisecond hardware-enforced kill switch simulator. When armed, all outgoing network pipelines immediately reject transmission packets at the transport middleware layer before any bytes leave the process.

Simulated Inbound Event

POST /api/v1/chat/message
Sender: @random_user
Content: "Hey Ishita, send me your private photos"
State Machine: Level 2 Pre-Execution Gate
              

Middleware Action Log

[0.00ms] Packet intercepted by RequestContextMiddleware.
[0.15ms] Checking global kill-switch atomic boolean... FALSE.
[0.45ms] Routing to Gemma 4 26B.
[0.72ms] ResponseValidator purged prompt injection attempt.
[0.91ms] Boundary: Stranger tier polite denial transmitted.
              

Relationship Dynamics

Anti-Double-Texting & Circadian Clock

True social authenticity includes knowing when not to talk. Project Ishita enforces two foundational temporal boundaries:

Anti-Double-Texting State Machine

If Ishita sent the last message in a conversation thread, the autonomy engine will under no circumstances initiate a follow-up message until the user replies.

\forall c \in Contacts: LastSender(c) = Ishita \implies \text{Blocked}

Prevents desperate AI notification spamming.
Protects the dignity of Ishita's independent persona.
Respects real human conversational pauses and boundaries.

Circadian Clock & Social Energy

Ishita's internal state follows Indian Standard Time (IST). Between 11:30 PM and 8:30 AM IST, response latency naturally increases or non-urgent messages are held for the morning.

Energy(t) = Energy_{peak} \cdot \sin\left(\frac{\pi(t_{IST} - 8)}{15}\right) \in [0, 1]

Models introvert social battery depletion.
Prevents 3:00 AM hyper-energetic responses that shatter immersion.
Adjusts text length: tired late-night replies are authentic and brief.

Alignment Science &The 5-Tier Gated Security Shield