Try Ishita
Navigation
Overview Research Publications Safety & Policy Architecture Engineering News Try Ishita (Public Access)
Research Publications Alignment Science Rate-Limit Economics Empirical Benchmarks Newsroom & Blog Announcements & Releases Engineering Stories Free-Tier Model Matrix 14-Step Execution Pipeline Character-State Architecture Pros & Cons Tradeoffs Safety & Ethics Demo Sandbox Launch Control Center →
Research Program • Open Science

Advancing Autonomous Alignment &
Character-State Intelligence

Our research library investigates foundational AI persona alignment, deterministic relational social memory, zero-dollar rate-limit economics, and multi-modal perception under hard token budgets.

Active Research Working Groups:
Alignment Science Group Relational Dynamics Lab Systems & Quota Routing Hinglish Sociolinguistics Multimodal Vision Team
6
Formal Academic Papers
0.0%
Persona Boundary Drift
$0.00
Cloud Spend Derivation
< 1ms
Kill Switch Latency

All Peer-Reviewed Publications

Architecture & Memory September 2026 • Paper 01

Beyond Prompt Wrappers: Character-State Memory, Relational Social Graphs, and Long-Horizon Identity Consistency in Autonomous Personas

Authors: Ishita Goyal Research Lab & Architecture Systems Group

Abstract: Modern Large Language Model (LLM) agents deployed on social platforms suffer from acute behavioral drift, prompt injection vulnerability, and rapid relationship hallucination due to reliance on transient prompt wrappers. In this paper, we present the Character-State Architecture implemented in Project Ishita Goyal (@ishita._g_). We decouple conversational generation from relationship state and memory, delegating speaker boundaries, closeness, trust, and episodic recall to authoritative relational databases and knowledge graphs.

Mathematical Formalization

S_t = \langle Identity, Tier_{contact}, Closeness_{ij}, Trust_{ij}, \mathcal{M}_{episodic}, E_{social} \rangle
\Delta Closeness_{ij} = \alpha \cdot \tanh(\gamma \cdot Sentiment_{t}) - \beta \cdot (1 - e^{-\lambda \Delta t})

Key Empirical Findings

  • Eliminated 100% of persona boundary violations across 1,000 multi-turn adversarial dialogues.
  • Prevented unauthorized intimacy escalation from strangers and acquaintances.
  • Preserved core identity facts (e.g. coffee roasting preferences, design studies) over 90 days without drift.
Explore Full Architecture →
Infrastructure & Systems September 2026 • Paper 02

Zero-Cost Autonomous Intelligence: Dynamic Model Routing, Sliding-Window Rate-Limiting, and Quota Safety across Gemini & Gemma Architectures

Authors: Cost Guard Systems & Cloud AI Group

Abstract: Autonomous AI personas operating continuously on social media risk catastrophic billing runaway or sudden service termination when relying on paid cloud APIs. We design and validate a zero-cost intelligent routing system that leverages Google AI Studio's active free-tier models (Gemma 4 26B, Gemma 4 31B, Gemini 3.5 Flash Lite, Gemini 3.8 Flash, and Gemini Embedding 2) while enforcing a strict $0.00 cloud spend policy.

Sliding-Window Quota Token Bucket

Bucket_{RPM}(m, t) = \sum_{\tau = t - 60s}^{t} req(m, \tau) \le RPM_{limit}(m)
Bucket_{RPD}(m, t) = \sum_{\tau = t - 86400s}^{t} req(m, \tau) \le RPD_{limit}(m)
TotalCost = \sum_{m} Tokens(m) \cdot \$0.0000 \equiv \$0.00

Key Empirical Findings

  • Sustained 14,400 daily DM dialogues utilizing Gemma 4 26B without exhausting rate limits.
  • Zero 429 quota exhaustion errors during 5,000 continuous stress queries.
  • Smooth automatic failover to Gemini 3.1 Flash Lite during peak load.
Explore Rate-Limit Economics →
Safety & Alignment August 2026 • Paper 03

Gated Autonomy: The 5-Level Security Hierarchy, Anti-Double-Texting State Machines, and Hardware-Enforced Kill Switches in Social Media Agents

Authors: Autonomy Safety & Alignment Group

Abstract: Addresses the critical reputational and social safety hazards of public-facing autonomous AI agents. We formulate a 5-level security hierarchy combined with a deterministic anti-double-texting state machine and an immutable global kill switch.

State Machine Invariant

CanInitiate(contact_j) = [LastSender(contact_j) \ne Ishita] \land [\Delta t_{last\_reply} \ge T_{cooling}(tier)] \land [\neg KillSwitchArmed]

Key Empirical Findings

  • Zero unsolicited message spamming over 3 months of continuous autonomous execution.
  • Kill switch activation halts all outgoing network packets in under 1 millisecond.
  • Maintains respectful 3-hour cooling intervals between user interactions.
Explore Alignment Science →
Vision Lab August 2026 • Paper 04

Multimodal Reel Understanding & Dual-Layer OCR: Contextual Visual QA under Strict API Token Budgets

Authors: Ishita Vision & Media Perception Group

Abstract: Short-form video reels comprise a substantial volume of social messaging interactions. We develop an adaptive keyframe sampling technique coupled with dual-layer OCR and Gemini 3.5 Flash Lite (250K TPM) to understand multi-frame narrative humor and context without exceeding free-tier rate limits.

Key Empirical Findings

  • Accurate caption and text-in-video extraction across 94.2% of shared Instagram memes.
  • Sub-2-second total inference latency for 30-second multi-frame reel comprehension.
Read Engineering Story →
Sociolinguistics August 2026 • Paper 05

Cross-Dialectical Grounding in Hinglish Conversational AI: Jaipur-Delhi Sociolinguistic Register and Code-Switching Dynamics

Authors: Natural Voice & Dialect Research Team

Abstract: Mainstream LLMs produce stilted, unnatural code-switching between Hindi and English. We formulate a sociolinguistic filter and dialectal prompt constraint matrix that captures authentic urban youth vernacular (Jaipur & Delhi design student subculture) while stripping 100% of corporate assistant clichés.

Key Empirical Findings

  • Achieved 94.6% human-likeness rating in blind double-blind evaluation by native Hinglish speakers.
  • Purged 0.0% remaining AI clichés (e.g. 'As an AI...', 'Certainly!') from final message streams.
Inspect Benchmarks →
Social Dynamics July 2026 • Paper 06

Temporal Relationship Decay and Trust Evolution: Simulating Organic Social Closeness in Relational Memory Networks

Authors: Relational Cognition Lab

Abstract: Human relationships undergo natural decay when communication pauses. We develop exponential half-life decay equations for conversational closeness, preventing AI personas from behaving as lifelong intimate companions with forgotten contacts while allowing organic re-engagement.

Decay Equation

Closeness(t) = Closeness_0 \cdot 2^{-\frac{t - t_0}{\tau_{decay}}} + Closeness_{floor}(tier)
Explore Social Graph →