Independent AI safety research · Open source · Reproducible by design
From statistical text traces to portable evidence for AI workflows.
CogniPrint began with interpretable statistical measurement of text transformations. That mathematical core became working software, controlled benchmarks, and a broader research programme for portable, privacy-aware evidence in multi-principal AI systems.
Roman Adriashkin — Independent Researcher-Engineer & Founder, CogniPrint Research Initiative
The project started with measurement, not detection.
The research question is which measurable traces survive editing, translation, paraphrasing, summarization, model-to-model rewriting, and mixed human/AI production chains—and where the correct conclusion is that evidence is insufficient.
Hypothesis. Statistical regularities in text can be represented reproducibly.
Theory. Define an interpretable feature map φ(T) and separate measurement from inference.
Program. Implement deterministic extraction, distances, perturbation analysis, entropy, and n-grams.
Benchmark. Freeze a leakage-safe pilot and compare against transparent baselines.
Next layer. Extend the evidence discipline to multi-principal agent workflows.
Mathematical core
The current v2 profile uses twelve interpretable coordinates: mean word length, type-token ratio, character entropy, word entropy, punctuation ratio, uppercase ratio, mean syllables per word, character-bigram uniqueness, Simpson diversity, digit-token ratio, mean sentence length, and reading ease.
H(X) = −Σ p(x) log₂ p(x)
cos(a,b) = (a·b)/(‖a‖‖b‖)
d₂(a,b) = √Σ(aᵢ−bᵢ)²
Δφ = φ(T′) − φ(T)
Theory became executable evidence.
The public implementation includes deterministic feature extraction, documented normalization, cosine and Euclidean geometry, controlled perturbation tests, corpus aggregation, entropy and word/character n-gram analysis, reproducibility scripts, tests, and versioned evidence artifacts.
Fixed Stage A: the baseline won.
The fixed descriptive pilot uses a 351/149 grouped train/test split across five classes with zero train/test lineage overlap. Simple hashed TF-IDF baselines outperform the current 12D nearest-centroid profile.
| Method | Accuracy | Macro-F1 |
|---|---|---|
| Word 1–2 hashed TF-IDF | 0.5973 | 0.5952 |
| Character 3–5 hashed TF-IDF | 0.5839 | 0.5788 |
| CogniPrint 12D | 0.5369 | 0.5359 |
| Length-only | 0.2953 | 0.2474 |
Interpretation: the compact 12D profile is not currently a stronger classifier than simple surface baselines on this controlled task. The negative result is retained because it constrains the theory and defines what future attribution research must beat.
Portable Evidence Capsules for Multi-Principal Agent Workflows
The current multi-agent research extends the same methodology from content measurement toward provenance and failure tracing under incomplete or conflicting records.
Multi-principal provenance
Represent event, delegation, version, authorization, and lineage evidence across independently controlled systems.
Failure tracing
Test whether evidence can localize broken action paths without assigning responsibility to the wrong principal.
Privacy-preserving verification
Evaluate what can be checked from hashes, bounded metadata, and structured records when raw payloads cannot be shared.
Fail-closed uncertainty
Preserve UNKNOWN and abstention states when records are incomplete, conflicting, replayed, or insufficient.
Research status: active direction. Future empirical milestones are not presented as completed results.
Scientific boundary
Current scientific readiness remains descriptive_only. Limitations follow the evidence rather than replacing it.
- Descriptive statistical profiles
- Controlled transformations and comparisons
- Entropy and n-gram diagnostics
- Versioned evidence artifacts
- Explicit uncertainty states
- Author identity or intent
- Unique or exact source model from text similarity
- Definitive AI origin
- Generation-lineage reconstruction
- Legal responsibility or forensic provenance
Optional third-party model API integration. Provider names describe technical integrations only and do not imply employment, sponsorship, partnership, or organizational affiliation.
Reviewer entry points
Public source and research repository · Research vision · Failure charter · Researcher profile · ORCID