Project summary

I built a behavioral observability framework for AI systems in production — 18 months, 40+ live services, white paper published May 2026. Asking for time to open-source the methodology. No weight access required.

What are this project's goals? How will you achieve them?

1. The Problem

We can audit what AI systems say. We cannot audit what they do.

The AI interpretability research field is growing fast — but almost entirely inside labs with weight access. The production reality is different: deployed systems run behind APIs, produce outputs, and behave in ways that no one fully tracks.

The gap between 'trust the output' and 'retrain the model' is enormous. In that gap:

• behavioral drift goes undetected until it fails in production

• model version changes produce unpredictable behavioral shifts

• multi-agent pipelines have no audit trail for intermediate behavioral structure

• alignment research produces frameworks that cannot be applied without weight access

This is not a niche problem. Every team deploying AI in production faces it. The missing layer is not logging or evaluation. It is process-level behavioral provenance that remains non-authoritative, privacy-bounded, and usable without weight access.

2. Cognitive Fossils: The Framework

Cognitive Fossils is a gray-box behavioral observability methodology for AI systems in production. It operates without weight access, without intrusion into the model's internals, and without modifying system behavior.

Core distinction — what makes it different

What everyone else observes

Output (what the model said)

Memory (raw session history)

Observation (passive reading)

What Cognitive Fossils captures

Fossil (structured behavioral trace, bounded and versioned)

Structure (observable behavioral invariants)

Audit (structured comparison across versions and time)

The key insight: a fossil is not a log. It is a structured behavioral snapshot — bounded, comparable, and versioned — that can be diffed against other fossils to detect drift, regression, or guardrail failures over time.

What the framework enables

• Behavioral fingerprinting — identifying a model's behavioral signature from outputs alone

• Drift detection — comparing behavioral fossils across time, versions, or providers

• Cross-system audit — applying the same methodology to any LLM pipeline

• Non-intrusive observability — no access to weights, no prompt injection, no model modification

  
  3. Live Evidence — Built, Running, Validated

  This is not a proposal to build something. The infrastructure exists. The ask is time to formalize, document, and distribute it.

  ---

  The Cognitive Fossils framework has been running in production for 18 months on the AI-ZI-ME ecosystem — a multi-agent AI platform hosted on a single VPS with 40+ PM2 services.

  Important timeline: the white paper (v1.0) was published on May 7, 2026. The cockpit screenshots below were captured on June 13, 2026 — five weeks later. They are not illustrations of a concept. They are post-publication validation of a live system that continued to evolve after the framework was formally documented.

  The operator cockpit — behavioral divergence detection in real time

  
  

  Fig. 1 — Génome opérateur cockpit (supervision.cordee.ovh) — June 13, 2026. État: corrélation active · Cohérence: 89 · Alignement global: divergent · Action recommandée: lire. The doctrinal contract is displayed inside the tool: output ≠ processus, mémoire ≠ structure, observation ≠ intervention.

  What this screenshot shows is not a demo. It is the Cognitive Fossils framework operating on live data:

  • 'Alignement global: divergent' — structural drift detected on correlation, not on output content

  • 'Action recommandée: lire' — observation only, never intervention; non-intrusion enforced at the interface level

  • 'Dernier fossile: preuve structurelle agrégée' — the audit trace is structural, not a log of user content

  • 'Notes opérateur: output ≠ processus...' — the doctrinal contract is visible in the tool, not just in documentation

  Multi-modal behavioral capture — intensity, confidence, and temporal freshness

  
  

  Fig. 2 — Captation view (supervision.cordee.ovh) — June 13, 2026. Four modalities: texte (0.53/0.77), voix (0.50/0.71), signal visuel (0.43/0.74), continuité (0.48/0.77). Each carries independent intensity + confidence + temporal freshness. Quality signals: capture exploitable (faible), hétérogénéité intermodale (moyenne), fragilité de contexte (faible). Timeline contrôlée: aucun contenu brut utilisateur exposé.

  This view demonstrates the fossil capture layer. Each behavioral modality is tracked independently with intensity, confidence, and temporal freshness — whether the signal is recent or cooling (structurally significant for drift detection).

  The 'Timeline contrôlée' confirms the core privacy invariant: user raw content never reaches the observability layer.

  What is live today

  • Production conversational surface — Cordée, used as a real-world interaction layer.

  • Behavioral observability layer — Quark-AI, including fossil generation, behavioral fingerprinting, and drift comparison.

  • Infrastructure trust layer — Sibil Monitor, a separate commercial preflight/observability product.

  • Guardrail test suites — regression tests enforcing output ≠ process, memory ≠ structure, observation ≠ intervention.

  • Shadow comparison pipeline — local model behavior compared against reference outputs to detect behavioral gaps.

  Test coverage across the ecosystem: 60–476 passing tests per service, with smoke regressions for all doctrinal guardrails.

How will this funding be used?

The infrastructure is running. What is missing is the public layer that makes it usable by others: documentation, extraction, and distribution.

Period Deliverable Measurable output

M1–M2 Extract Cognitive Fossils spec from Public spec v1.0 + reference production codebase into a standalone schema implementation-agnostic protocol document

M2–M3 Publish open-source reference implementation GitHub release with test of fossil generation and diff (language-agnostic, suite pluggable)

M3–M4 Document the shadow comparison pipeline as Technical paper + a reusable methodology for non-invasive reproducible dataset behavioral gap analysis without weight access

M4–M5 Produce 3 case studies: behavioral drift detection, Published case studies cross-provider comparison, with real data doctrinal guardrail regression

M5–M6 Public dashboard for Quark-AI behavioral audit Live tool with onboarding

(open access for community testing)

Budget allocation

Item Amount %

Research & writing time (6 months, solo) $13,500 75%

Infrastructure (VPS, API costs, tooling $2,700 15% for public release)

Community (events, collaboration, $1,800 10% external review)

TOTAL $18,000 100%

This budget buys focused extraction time, not initial R&D. The R&D risk has already been partially absorbed by 18 months of unpaid production work.

5. Why This, Why Now, Why Me

Why this matters now

AI systems are being deployed faster than our ability to audit them. The interpretability field is producing important work — but it assumes weight access that production teams don't have. The gap is growing. A production-native observability methodology that doesn't require lab conditions is not a nice-to-have: it becomes critical as deployment scales.

Why I can do this

• 18 months of building this methodology inside a live, complex system — not a toy demo

• Independent researcher without institutional constraints — able to move fast and publish openly

• Full-stack builder: from low-level FastAPI services to mobile apps to doctrinal prompt engineering

• The infrastructure is already the proof — Cognitive Fossils is a formalization of what works in production, not a theoretical proposal

What I am not claiming

This is not interpretability in the academic sense. Cognitive Fossils does not explain model internals. It audits observable behavior — which is what production teams need, and what academic interpretability cannot currently deliver at scale.

White paper URL https://quark-ai.cordee.ovh/cognitive-fossils-white-paper.pdf

6. Risks & Mitigations

Risk Mitigation

Solo builder — bus factor All deliverables are open-spec and open-source. The methodology survives the individual.

No academic affiliation The production infrastructure is the credential. 18 months of live validation is harder to fake than a paper.

Scope drift (too many products) Grant scope is strictly Cognitive Fossils extraction and distribution. Commercial products (Sibil, Cordée) run in parallel and are self-sustaining.

Closing

The production AI deployment world still lacks a clear layer for behavioral provenance without weight access. I have built and tested an early version of that layer in production over 18 months.

This grant is not to start from zero. It is to extract, document, validate, and release what already works as a public research artifact.

Every core claim in this proposal is backed by a running service, a test suite, a public artifact, or a changelog entry. The screenshots above are from a live system, taken on June 13, 2026.

I am not asking to be believed. I am asking for time to document what is already true.

Live evidence: cordee.ovh · quark-ai.cordee.ovh · monitor.cordee.ovh

By AI-ZI-ME

https://aizime.cordee.ovh/