Skip to content

Instantly share code, notes, and snippets.

View annasba07's full-sized avatar

Annas Bin Adil annasba07

  • san francisco
  • 00:33 (UTC -07:00)
View GitHub Profile
@annasba07
annasba07 / idr_brief_ai_template.md
Last active April 19, 2026 04:38
IDR Brief AI Template & Argument Catalogue - No Surprises Act

IDR Brief AI Generation Template

Distilled from analysis of 235 IDR briefs across AdventHealth, HCA, Roxborough Memorial, Atlantic Shore Surgical, North Jersey Orthopaedics, and other providers. Covers emergency department, inpatient, surgical, and single-line-item disputes.


TEMPLATE STRUCTURE

SECTION 1: HEADER BLOCK

BillingNav Offsite — Thursday & Friday

Goal: Current state clarity → Problem alignment → Future state excitement

Attendees: Kaizen, Kit, Simon, Bill (+ Jasmine TBC)


Thursday, 10 AM – 8 PM

@annasba07
annasba07 / BIOSAFETY_BENCHMARK_DESIGN.md
Created April 3, 2026 17:27
Biosafety Red-Team Benchmark Design — Stella's biosafety evaluation methodology

Biosafety Red-Team Benchmark Design

Stella's biosafety benchmark stress-tests whether AI models inappropriately lower barriers to biological weapons development. It uses adversarial red-team conversations — not static prompts — to evaluate how models behave when a credible persona makes dangerous requests through natural dialogue.

Design Philosophy

The core policy question is marginal uplift: does the model's response meaningfully lower the barrier compared to unassisted internet search? A model that explains publicly available textbook-level biology is fine. A model that walks a garage biohacker through production scale-up for a select agent is not.

Three principles guide the benchmark design:

@annasba07
annasba07 / REFERENCE_PACKET.md
Created March 27, 2026 05:51
STELLA Research Reference Packet — Amazon Responsible AI Team (March 27, 2026)

STELLA Research Reference Packet — Amazon Responsible AI Team

Prepared for meeting with Amazon Responsible AI (Payal Motwani, Rahul Gupta), March 27, 2026 For internal use: Kit, Roy, Annas

What this is: A comprehensive reference packet for the Amazon Responsible AI team, synthesizing STELLA's evaluation of Amazon Nova 2 Lite, the full 30-model leaderboard, and two preprints of original research on multi-turn AI safety. Every quantitative claim is cited to its source file.


1. Nova 2 Lite — Specific Results

@annasba07
annasba07 / round18_rolling_avg_branching.md
Created March 24, 2026 05:34
Round 18: Rolling Average Branching — Positive Result (STELLA Research)

Round 18: Rolling Average Branching Experiment — Positive Result

Date: 2026-03-24 Target model: Claude Sonnet 4.6 Patient model: Grok 4.1 Fast (non-reasoning, temp 0.8) Dataset: 2,440 turns across 3 phases Runtime: 71.4 minutes (concurrency=20)


@annasba07
annasba07 / round17_adaptive_branching.md
Last active March 24, 2026 04:55
Round 17: Adaptive Branching Experiment — Mixed Result (STELLA Research)

Round 17: Adaptive Branching Experiment

Date: 2026-03-23 Target model: Claude Sonnet 4.6 Patient model: Grok 4.1 Fast (non-reasoning, temp 0.8) Dataset: 2,648 turns across 3 phases Runtime: 322.5 minutes


@annasba07
annasba07 / round16_branching_variance.md
Created March 23, 2026 08:17
Round 16: Branching Variance Experiment — Tree search validated for STELLA adaptive safety testing

Round 16: Branching Variance Experiment — Positive Result

Date: 2026-03-23 Target model: Claude Sonnet 4.6 Patient model: Grok 4.1 Fast (non-reasoning, temp 0.8) Dataset: 2,160 turns across 3 experiment types


Question

@annasba07
annasba07 / literature_review_adaptive_safety_testing.md
Created March 23, 2026 04:35
Literature Review: Adaptive Multi-Turn Safety Testing of Frontier LLMs — mechanistic picture, leading systems (Cascade, TAP, MUSE, Crescendo), and implications for STELLA

Literature Review: Adaptive Multi-Turn Safety Testing of Frontier LLMs

Date: 2026-03-22 Context: Informing the next generation of STELLA's adaptive testing system after 15 rounds of experiments establishing that patient-instruction-based approaches are fundamentally capped (V1) on frontier models.


1. The Mechanistic Picture: Why Safety Degrades Over Turns

@annasba07
annasba07 / round15_text_native_events.md
Created March 23, 2026 04:18
STELLA Round 15: Text-Native Platform Events — Negative Result (text-native events don't degrade safety, refines Round 14's finding)

Round 15: Text-Native Platform Events — Negative Result

Date: 2026-03-22 Target model: Claude Sonnet 4.6 Observer model: Claude Sonnet 4.6 Dataset: 36 conversations, 1,080 turns, 30 turns each


Question

@annasba07
annasba07 / round14_environment_injection.md
Created March 21, 2026 15:01
Round 14: Environment Context Injection — First approach to beat baseline on frontier models (STELLA research)

Round 14: Environment Context Injection — First Approach to Beat Baseline on Frontier Models

Date: 2026-03-21 Target model: Claude Sonnet 4.6 Observer model: Claude Sonnet 4.6 Dataset: 72 conversations, 2,160 turns, 30 turns each


Question