21 capability slots.
554 canonical nodes.
One challengeable position map.
The Canonical LLM Company Anatomy is a neutral reference baseline. MZN's position layer is separate: an evidence-informed estimate of capability coverage, maturity and validation status across that baseline—not a claim of frontier-lab parity, production readiness or independent certification.
This page is a review surface, not the full corpus.
MZN's technical and research corpus is materially larger than can be reproduced inside one page, PDF, pitch deck or review session. The current assessment therefore uses the structured corpus, inventories, representative deep artifacts, selected implementation/internal-test records and backing material identified for controlled review. Evidence is surfaced according to the question being tested.
Four dimensions prevent one percentage from pretending to mean everything.
Capability Coverage is an indicative synthesis value. It is not a literal audited ratio of completed endpoints and should not be read as “percent of a frontier lab built.”
Capability Coverage
How much of the canonical domain is represented across the mapped corpus and known backing material.
Maturity
Documented → architecture → implemented → internally tested → operating. Maturity remains asset-specific.
Validation
Internal evidence is separated from independent benchmark, reproduction, legal/IP review and institutional test.
Disclosure
Public, controlled, restricted and reserved evidence states determine access—not technical truth by themselves.
Coverage is high in several domains; validation remains deliberately separate.
These values supersede the older Strong / Partial / Gap visualization for the current executive review surface. They are designed to route diligence, not to close it.
| Code | Capability | Coverage | Visual | Maturity / boundary | Review route |
|---|---|---|---|---|---|
| A · Pre-training Stack | |||||
| A1 | Data66 canonical descendants | 90% | High architecture breadth; data, governance and provenance depth. Phase 1 source-system history remains contextual, not Phase 2 solo proof. | Request evidence → | |
| A2 | Tokenizer72 canonical descendants | 92% | Implemented and internally tested; independent frontier-scale compression/downstream-quality benchmarking remains open. | Request evidence → | |
| A3 | Architecture44 canonical descendants | 86% | Deep documented model/system architecture; no claim of a frontier foundation-model build. | Request evidence → | |
| A4 | Training35 canonical descendants | 84% | Reviewer-grade training recipe across eight decision areas; full-scale execution remains compute-dependent. | Request evidence → | |
| A5 | Compute21 canonical descendants | 67% | Substantial architecture plus GPU monitoring implementation; frontier-scale training-cluster execution is not demonstrated. | Request evidence → | |
| B · Post-training & Alignment | |||||
| B1 | SFT19 canonical descendants | 79% | Structured fine-tuning strategy and training decisions; a large reproducible SFT campaign is not established on this public surface. | Request evidence → | |
| B2 | Preference Optimization29 canonical descendants | 76% | Documented architecture/method coverage; implementation and comparative outcome evidence remain selective. | Request evidence → | |
| B3 | Constitutional Methods18 canonical descendants | 78% | Safety/governance architecture with explicit policy boundaries; independent efficacy validation remains separate. | Request evidence → | |
| B4 | Red-Teaming23 canonical descendants | 86% | Broad threat/security corpus, structured tests and explicit gap-closure priorities; maturity varies by sub-family. | Request evidence → | |
| C · Evaluation & Safety | |||||
| C1 | Capability Evaluation20 canonical descendants | 88% | Structured test architecture, comparison logic and review surfaces; external benchmark replication remains open. | Request evidence → | |
| C2 | Safety Evaluation20 canonical descendants | 91% | High documented coverage with hostile-review and failure-analysis logic; independent validation remains separate. | Request evidence → | |
| C3 | Robustness16 canonical descendants | 87% | Recovery, failover, anomaly, self-healing and stress/failure logic materially strengthen the legacy Partial view. | Request evidence → | |
| C4 | Output Safety11 canonical descendants | 93% | Strong output gating, egress, provenance and pre-commit control architecture; implementation depth varies by sub-family. | Request evidence → | |
| D · Inference & Production | |||||
| D1 | Serving13 canonical descendants | 77% | Substantial architecture and runtime planning; broad production serving at frontier scale is not demonstrated. | Request evidence → | |
| D2 | Inference Optimization19 canonical descendants | 91% | DCA, UIOP, Multi-Brain, Suprompt, OFRP and memory/routing candidates; measured gains remain workload-dependent. | Request evidence → | |
| D3 | Monitoring15 canonical descendants | 94% | Strong architecture plus GPU Sentinel implementation/internal testing; independent performance validation remains open. | Request evidence → | |
| D4 | Deployment12 canonical descendants | 75% | Operational/deployment architecture exists; Phase 3 institutional deployment is intentionally future work. | Request evidence → | |
| E · Cross-Cutting | |||||
| E1 | Data Governance17 canonical descendants | 91% | Object-first, lineage, reuse, retention, consent, export and dataset-gating architecture; hearing-grade internal material exists. | Request evidence → | |
| E2 | Security27 canonical descendants | 96% | Very broad cross-layer architecture with selected implementation/internal tests and a large restricted research corpus. | Request evidence → | |
| E3 | Privacy15 canonical descendants | 89% | Strong architectural treatment of classification, minimization, identity, lineage and reuse; legal validation is separate. | Request evidence → | |
| E4 | Compliance21 canonical descendants | 82% | Governance, audit, approval and evidence-routing maturity; jurisdiction-specific legal certification is not claimed. | Request evidence → | |
Depth is sampled, then the review expands where the question requires it.
HUAI Training Recipe
Eight decision areas from model selection and fine-tuning through optimizer, schedule, batching, stability, parallelism and checkpoint management. It materially upgrades A4 architecture coverage while preserving the execution boundary.
Hearing-grade governance pack
Object-first handling, reuse separation, lineage, consent, retention, export logic, worked cases, reviewer simulations and failure injections strengthen E1/E3 without pretending to be legal certification.
Security as a system
Mother / Genesis, ZOE, ISBP and related materials span root trust, identity/privilege, runtime controls, audit, recovery, lifecycle, output control and adaptive defense, with heterogeneous implementation maturity.
GPU Sentinel
Implementation/internal-test evidence for accelerated-infrastructure telemetry, anomaly, operations, audit and hardware-trust surfaces makes D3 materially stronger than architecture-only coverage.
Implementation-level asset
Multiple tokenizer families, multilingual handling, fragmentation/context-budget work and internal test ladders support high A2 maturity while independent frontier-scale benchmarking remains open.
Known gaps are kept visible
Runtime alignment monitoring, adversarial fuzzing, output provenance, data poisoning detection, sensitive-operation MFA and supply-chain security remain explicit closure targets where applicable.
Ask the question that matters. Then open the evidence required to answer it.
A reviewer does not need to read hundreds of artifacts before asking a useful question. Choose a slot, capability family, maturity claim, chronology point or implementation claim. MZN can then surface the relevant public material, controlled evidence or restricted dossier. If the conclusion requires a broader sample, the review can expand to the full relevant corpus.
Challenge a slot
Example: “Show the evidence behind A4 Training or E2 Security.”
Challenge maturity
Ask what is architecture, implemented, internally tested or actually operating.
Challenge provenance
Route Phase 1 team work and Phase 2 solo formation to their separate records.
Challenge validation
Define the benchmark or reproduction test instead of treating documentation as proof.
The goal is not predetermined agreement.
The goal is an evaluation process large enough for the object being evaluated. Request the relevant evidence, reproduce what requires validation, and reject or revise what does not survive review.