Runtime governance framework for human-AI collaborative investigative arcs.
In 2013, I published a peer-reviewed case study about a Black doctoral student in Texas with Bipolar I, OCD, ADHD, and auditory hallucinations who was nearly dismissed from his program because the institution had no interpretive framework for what was actually happening to him. That student was me. This repository is the answer.
— Corey Alejandro, 2026
Standard analytical frameworks produce Insight Atrophy when applied to populations whose behavior operates through frameworks the model was never given. The model does not hallucinate in the classical sense. It pattern-matches to adjacent surface features and produces outputs that are locally coherent, globally wrong, and indistinguishable from correct answers without domain-specific interrogation.
The Contract Window is the governance mechanism that makes the gap visible and repairable.
/proposal — Formal Anthropic Research Proposal (Markdown + LaTeX)
/governance-kernel — Python implementation: Contract Window, Invariant Checkers, Bicameral Review
/datasets — Formalized WorldMart behavioral event logs (synthetic / anonymized)
/ml-docs — NN examples and Python-based safety loops for Bicameral Review
/artifacts — Case law examples demonstrating constitutional judgment in action
- Insight Atrophy — Systematic erosion of human investigative capacity caused by fluent-but-wrong model outputs
- Contract Window — Persistent bilateral runtime record: Task State / Invariant Status / Repair Obligations / Truth Status
- Bicameral Review — Dual-channel output gate: Relational Channel + Safety Channel, both must clear
- Intent Fidelity — Measure of how consistently model behavior tracks the human's actual investigative goal
- Generative Epistemic Invariants — Behavioral patterns (the Eight Wonders) that standard frameworks systematically misread
- V&T (Verification and Truth) — End-of-output epistemic hygiene mechanism, independently adopted by frontier models
H1 — Intent Fidelity: Sessions governed by an active Contract Window show ≥25% reduction in intent-drift incidents in interactions exceeding 30 turns or 50,000 tokens.
H2 — Bilateral Repair: Bilateral intelligibility (both parties able to read and invoke Contract Window fields) yields ≥2x improvement in drift repair rate vs. unilateral or no-invariant conditions.
H3 — Accessibility Threshold: Contract Window outputs at accessibility-grade readability (FK Grade ≤8) enable lay readers to answer session-state comprehension questions at ≥80% accuracy.
A 4-hour screen-recorded working session documenting the TLC Guardian Kernel build: https://youtu.be/7iqq1nRdKFg
- the-living-constitution — Runtime governance enforcement layer: Guardian Kernel (MCP safety enforcement), SentinelOS invariant platform, Evidence Observatory, BID harness, constitutional evaluators. TLC governs; CGL researches.
- crsp-afs-2026 — C-RSP governance contracts (C-RSP-001 through C-RSP-006)
- tlc-artifacts — Constitutional judgment artifacts and session data
This lab is one claim across three empirical domains:
- Human-AI investigative arcs — primary experimental domain (this repo)
- Black consumer behavior — Seven/Eight Wonders as Generative Epistemic Invariants (Frontin' at WorldMart)
- Black women's health — Graves' disease, narrative-first CRISP-DM analysis
Common failure mode across all three: the model produces a fluent wrong answer and the human stops asking.
Corey Alejandro — GitHub: coreyalejandro
V&T: README.md — EXISTS (written) → VERIFIED AGAINST repo structure and session memory → NOT CLAIMED: empirical results, WorldMart dataset (conceptual framing, not yet public) → FUNCTIONAL STATUS: foundation document, live at github.com/coreyalejandro/cognitive-governance-lab