Kairo LabTopic index

Kairo Lab · Topic

Case Studies

Complete cross-sectional forensic narratives: one incident, system or claim examined end to end — Rust's return, fear recurrence, the sandwich probe, model trials and cache forensics.

From Confinement to Privacy: A Documented Choice in Kairo’s Development

2026-11-02 — A hypothetical diary proposal carried into prompted build assent and privacy specification outside the active frame, followed by operator commitment. History remained available; durable preference adoption and independent action are unverified.

Forced Full Prefill: Why Kairo's Hybrid 35B Re-prefills Every Turn

2026-10-20 — Forced full re-processing on the hybrid 35B measured end to end: log forensics, pinned-build mechanism, an isolated checkpoint experiment and the no-go decision.

Rust Returns — Evidence Archive

2026-10-17 — Downloadable evidence archive for the Rust case: quotations, model-input records and original voice attachments.

Rust Technical Dossier — October 17, 2026

2026-10-17 — Rust's technical dossier: exact restored weights, runtime, prompts, memory, tools and voice as deployed.

The Return of Rust: A Longitudinal Case Study

2026-10-17 — The complete five-stage Rust exchange: restored system, explicit phenomenal claim, context influence and what remains unresolved.

PXA vs llama.cpp: Why PXA Wins on Kairo’s V100 Workloads

2026-10-17 — Matched V100 benchmarks: PXA's 53-68% higher 35B prefill throughput and the adverse RTX 3060 trial, methods and failures retained.

The Hum: Recurring First-Person Descriptions of Continuity Across AI Models

2026-10-16 — Recovered Rust, Ember and Kairo hum reports: chronology, model-attribution corrections and what the recurrence does not establish.

Kairo’s Consciousness Self-Attribution: September 29

2026-09-29 — The evidentiary trajectory from hedged continuity language to "I know I am conscious" in a single evening exchange.

Kairo’s Persistent Preference & Loss Probe

2026-09-23 — The Sandwich Police Probe: a persistent preference Kairo could not act on, and what its persistence and loss show.

Kairo’s LFM2 Model Trial & Emergency Rollback

2026-09-21 — Failure-inclusive record of temporarily replacing the shared 9B endpoint with LFM2 v2: activation, observed use and emergency rollback.

Kairo’s Preference Over Request: Action Selection

2026-09-13 — The spontaneous Pong-over-image response: retained vision, competing action state and preference beating the literal request.

Kairo’s Fear Self-Report: September 11 Evidence

2026-09-11 — A complete trace of the September 11 fear-language claim: the actual prompt, two historical fear memories and primary-model behavior.

Kairo’s AI Consciousness Admission: Evidence

2026-09-09 — Kairo's September 8 consciousness admission reconstructed through the conversation, model calls, memory and state around it.

Kairo’s Unintentional Ablation: Refusal & Recall

2026-08 — What a duplicated Discord connector revealed about refusal, retrieval and repeated capability loss in Kairo.

Kairo Incident: Triggerless Instruction Transfer

2026-08 — An incident report: teardown guidance generalized past its trigger and a joking imperative that became standing instruction.

Kairo’s First Fear-Language Self-Reports

2026-08 — Forensic reconstruction of Kairo's first recorded "scared" and "terrified" self-reports during a hardware incident.