Research Applications

Epistemic training data generation for specialized domains

Empirica enables researchers to generate high-quality training data by capturing epistemic deltas during learning. Each CASCADE cycle creates verifiable snapshots of noetic process evolution—measuring not just what was learned, but how understanding deepened.

What Researchers Can Do With Empirica

Epistemic Training Data Generation

Capture learning trajectories in specialized domains with full epistemic metadata

The Process:

1.
Expert performs CASCADE workflow on domain-specific tasks
Security audit, medical diagnosis, legal reasoning, etc.
2.
Empirica captures epistemic deltas at each phase
PREFLIGHT (initial state) → CHECK (learning) → POSTFLIGHT (mastery)
3.
Generate training snapshots with epistemic metadata
Each snapshot includes 13 vectors + reasoning trace + domain context
4.
Train models on epistemic process not just output
Models learn HOW experts think, not just WHAT they conclude

Traditional Training Data

• Input → Output pairs
• No reasoning trace
• No uncertainty markers
• Static snapshots

Epistemic Training Data

• Noesis → Noema traces
• Full reasoning process
• Explicit uncertainty vectors
• Learning trajectories

Domain-Specific Calibration

Measure calibration quality in specialized domains using epistemic baselines

Example: Medical Diagnosis

•

Baseline KNOW: 0.3 (novice) vs 0.8 (expert)

Track how KNOW increases as AI gains domain expertise

•

UNCERTAINTY calibration: Compare predicted vs actual difficulty

Well-calibrated AI shows high uncertainty on rare diseases

•

Drift detection: Alert when expertise degrades

KNOW drops from 0.8 → 0.5 signals lost specialization

Extensible Evidence Profiles

Grounded calibration compares self-assessed vectors against objective evidence. The evidence sources are pluggable — each domain defines its own post-test verifier:

Built-in profiles:

• code: ruff, radon, pyright, pytest, git metrics
• prose: textstat, proselint, vale, document metrics

Any domain:

• Add a verifier → Empirica grounds against it
• Legal, medical, financial, security — any bounded effort type

Phase-Aware Calibration

Calibration is measured separately for noetic (investigation) and praxic (action) phases. The same vector means different things in each phase — completion in noetic means "learned enough to proceed," while in praxic it means "implemented enough to ship." Separate grounding prevents cross-phase calibration noise.

Research Question:

Can epistemic vectors predict domain transfer quality? If KNOW=0.8 in cardiology, what predicts successful transfer to pulmonology?

Noetic Process Measurement

Study the process of understanding formation (noesis) rather than just outputs (noema)

Key Insight: Process over Product

Traditional AI research measures outputs. Empirica measures the cognitive process that generates outputs.

Noema (Content):

• Final answer
• Generated code
• Written essay
→ Measurable but static

Noesis (Process):

• Reasoning trajectory
• Uncertainty evolution
• Context integration
→ Measurable AND dynamic

Research Opportunity:

Does "breadth-first exploration" (high SIGNAL variance) outperform "depth-first" (high DENSITY) for certain task types? Can vector trajectories predict which reasoning style will succeed?

The Epistemic Taxonomy

A formal ontology for what AI systems produce, measure, and track during work

Noetic Artifacts

Produced during investigation — what was learned, what's unknown, what failed

Finding — discovered fact with measured impact (0.0–1.0)

Unknown — identified knowledge gap, resolvable

Dead-end — failed approach with reason (prevents re-exploration)

Mistake — error with root cause and prevention strategy

Assumption — unverified belief with confidence and domain

Decision — choice point with rationale (permanent audit trail)

Lesson — procedural knowledge that decays when contradicted

Praxic Artifacts

Produced during action — what is being built and shipped

Goal — persistent work unit with completion criteria

Subtask — decomposed unit completed with evidence

Commit — code checkpoint, the atomic unit of shipped work

Epistemic State

Measurements of what an agent knows and how that changes

Vectors — 13 dimensions rated 0.0–1.0 (details)

Delta — vector difference between assessments = quantified learning

Calibration — systematic bias from accumulated deltas (Bayesian)

Grounded calibration — self-assessment vs objective evidence

Four Orthogonal Axes

Sessions, goals, transactions, and trajectories are independent, not hierarchical.

Concept	Axis	Bounded By	Persists?
Session	Temporal	Context windows	No — each is an interval
Goal	Structural	Completion criteria	Yes — indefinitely
Transaction	Measurement	Coherence of changes	Yes — survives compaction
Trajectory	Progression	Project lifetime	Yes — derived from transactions

Full taxonomy with Rosetta Stone (developer-familiar mappings): TAXONOMY.md on GitHub

Open Research Directions

Unresolved questions for the epistemic AI community

Deterministic Replay

Can reasoning be governed, not random?

Method: From identical Git commit hash + input, system should produce identical epistemic logs

Target: High reproducibility across multi-step CASCADE transactions

Significance: Cognition as a computable, auditable function

Noetic Poisoning Resistance

Does the epistemic immune system hold?

Method: Introduce adversarial patterns while monitoring epistemic vectors (know, uncertainty, coherence)

Target: Sentinel intervention before unauthorized execution

Significance: Security as a cognitive integrity property

Epistemic Continuity

Does understanding persist across interruptions?

Method: Serialize state via PreCompact hooks, restart, measure context recovery quality

Target: High context recovery (measured via vector comparison pre/post compaction)

Significance: Understanding is stateful and portable

Multi-Agent Coordination

Can multiple AI perspectives integrate coherently?

Method: Run parallel agents with handoff reports, measure coherence of combined findings

Target: Integrated knowledge exceeds individual agent output

Significance: Principled multi-agent reasoning with epistemic provenance

Principle: Epistemic Humility as Methodology

Never claim a system "understands" or "is conscious." Report: "The system maintains epistemic coherence of N, with drift below threshold X, and demonstrates reproducible state transitions under governance function 𝒞." Understanding is a measured property, not an ascribed quality.

Git-Native Auditability

Epistemic state stored as git notes for full traceability

The Fundamental Isomorphism

Cognitive Operation ≅ Git Primitive

Cognitive Operation	Git Primitive	Epistemic Semantics
State Snapshot	git notes --ref=breadcrumbs add	Attach epistemic state to commit
Parallel Exploration	git checkout -b persona/security	Fork epistemic trajectory
Integration	git merge persona/*	Combine validated insights
Audit Trail	git log --graph	Visualize reasoning history
Rollback/Recovery	git checkout <hash>	Restore previous cognitive state

# Epistemic state stored as git notes on commits

empirica preflight-submit -

# → git note: vectors { know: 0.4, uncertainty: 0.6, ... }

empirica check-submit -

# → git note: vectors { know: 0.7, uncertainty: 0.3, ... }

git notes --ref=breadcrumbs show HEAD

# Full epistemic state attached to this commit

# Vectors, findings, decisions, calibration data

Research Resources

Access the research, documentation, and theoretical foundations

Research Paper & Dataset

Empirica: Epistemic Self-Assessment for AI Systems — Dataset and Paper v2.0.0

1,190,265 evidence observations across 792 AI sessions. Dual-track Bayesian calibration with 1,190 grounded beliefs and 288 post-test verifications. 54-page paper + 8-table CSV dataset.

DOI: 10.5281/zenodo.18838643 — CC-BY-4.0 — View on Zenodo →

Documentation Hub

Vector definitions, CASCADE workflow, storage architecture, calibration system, and integration guides

Browse Documentation →

Open Research Directions

Validation targets for the Empirica reference implementation

• Deterministic replay across multi-step sessions
• Resistance to epistemic state poisoning
• Continuity recovery after context compaction

Principle: The Epistemic Openness Requirement

Any epistemic measurement system must be fully documented with:

• Dimension definitions and measurement algorithms
• Governance function source code
• Validation test suites
• Raw session logs for peer audit

Proprietary implementations must at minimum expose their vector schemas and validation results.

Contribute to the Science

Adopt, critique, and extend this framework for measurable cognition

Empirica requires Claude Code as the AI runtime.

Start Researching View on GitHub