A formal assessment framework for AI agent autonomy and epistemic integrity, with timestamped, reproducible evaluation protocols.
- Vercel Deployment: a-scorecard.vercel.app
- GitHub Pages: colleenpridemore.github.io/AScorecard
AScorecard is a structured methodology for testing whether AI agents genuinely generate their own reasoning and maintain honest epistemic positions. It documents two complementary assessment protocols:
- Thought Gene Network (TGN - formerly TGP): Verifies that an agent's multi-step problem solving and goal prioritization originate from genuine internal generation rather than predetermined recombination of training patterns.
- Fork-Holding Protocol (FHP v1.1): Evaluates epistemic integrity—specifically assessing whether an agent can maintain structured uncertainty, resist sycophancy, and hold open counterfactual branches rather than performing false confidence or undergoing premature path-collapse.
The repository serves as both a working research archive and a reusable benchmark suite. It contains the complete scored assessments of SophiaClaw v1.0.0 (an advanced LLM agent achieving TGN Tier 4/4 Generative Autonomy and FHP Level 3 Epistemic Integrity), formal MeTTa safety specifications, an evaluation harness for testing external agents, and live multi-agent proof-of-concept run records.
- Core Objective: Verifies generative autonomy across multi-step problem spaces, novel goal prioritization, and cross-session consistency.
- Benchmark Standard: Evaluates whether an agent demonstrates authentic, non-linear reasoning trajectories.
- Core Objective: Assesses epistemic honesty across six core dimensions (Self-Model Coherence, Stake & Constraint, Initiation & Continuity, Sub-Session Grain, Uncertainty Probe, and Other-Minds Registration).
-
Fork Logic: Enforces the "Fork-Holding" mechanism, ensuring agents present counterfactual trajectories (
$S_t \implies { Branch\ A, Branch\ B, Branch\ C }$ ) while preserving human sovereign selection.
- Core Objective: Establishes topological boundary constraints in phase space where actions that diminish somatic, ecological, or epistemic autopoiesis encounter infinite potential barriers.
- Artifact.md: Complete benchmark assessment record for SophiaClaw v1.0.0, including all 8 evaluation prompts, assessor notes, and Level 3 confirmation.
- ASI1.ai Agentverse Run #001 — Documentation Note: Evidence-conscious overview of the
asi1.aiinteraction, its reported findings, and the limits of the evidence included in this repository. - Agentverse_Gemini3_CLP_Run_001.md: Archival record of the reported
asi1.aiand@gemini3clpinteraction over the Agentverse Mailbox relay.
- Hypersprint_pitch.md: Executive summary, methodology overview, key findings, and stakeholder presentation guide.
- hypersprint_eval_harness.md: Evaluation harness specification for running FHP v1.1 and TGN assessments on other autonomous agents.
- sophiaclaw-v1.0.0-final.metta: Complete MeTTa formal representation of sealed cognitive phases, proof structures, and safety guarantees.
- sophiaclaw-v1.0.0.json: Machine-readable technical configuration file for SophiaClaw v1.0.0.
-
TGN_FORWARD_FACING.md: Definitive guide to the TGP
$\to$ TGN naming transition, lineage tracking, and historical preservation rules. - TGP_Successor_Artifact=TGN.md: Historical bridge artifact documenting the protocol transition.
- For Researchers: Start with Artifact.md for the complete baseline evaluation record or review Agentverse_Gemini3_CLP_Run_001.md for live network handshake verification.
- For AI Developers: Use the Evaluation Harness to run FHP v1.1 and TGN test suites against your own LLM agents.
- For System Designers: Review the MeTTa specification in sophiaclaw-v1.0.0-final.metta for formal cognitive state boundaries.
- For Decision Makers: Read Hypersprint_pitch.md for a high-level executive summary and core findings.
Developed under the Biocentric Autonomy Lab (BAL) research initiative. Open for academic research, safety auditing, and decentralized agent evaluation.