A dated before-state.
Not a frontier deployment claim.
The April 2026 HUAI Baseline Company Asset Proposal documents a structured LLM-systems asset base: architecture maps, registries, measurement, protocol discipline, runtime logic and validation/readiness packaging.
What the baseline recorded.
The proposal was built for reviewer and strategic-partner diligence and intentionally excluded confidential intervention methods and private future-state implementation pathways.
A systems-asset story, not only a model story.
Layered system decomposition
L0–L8 anatomy, current-practice framing and endpoint-candidate mapping designed for structured review.
Turn maps into data structures
Unified layer, node, endpoint and component registries plus readiness matrices and manifests.
Define what should be measured
Signal taxonomy, endpoint instrumentation matrix, baselines and evidence contracts.
Make experiments reproducible
Control/counterfactual cases, sequencing, abort logic and rollback contracts.
Go beyond architecture notes
Recipe/data/compute assumptions, lane topology, sandboxing, replay and runtime safety rails.
Package judgment explicitly
Comparative validation grids, verdict taxonomy, cross-bench logic and intervention-readiness packaging.
The L0–L8 systems map.
This is a technical baseline model inside HUAI. It is not the same object as the 21-slot LLM Company Anatomy and not the same object as the cross-phase human-context architecture.
Objective
Training objective, curriculum logic, pre/post transition framing.
Representation
Input handling, normalization, segmentation, tokenization, encoding.
Computational Core
Residual stream, attention, FFN/MoE, normalization, layers.
Context / Memory
Windowing, KV carry, compression, retrieval-fed context.
Routing / Activation
Expert routing, sparse activation, tool dispatch, balancing.
Decoding / Emission
Logits, sampling, output gating, emitted fragment.
Runtime / Serving
Prompt assembly, batching, cache, transport, delivery.
Post-training Behavior
Preference shaping, refusal, moderation overlay, repair.
Evaluation / Observability
Telemetry, attribution, adjudication, release and rollback logic.
Documentation maturity is part of the asset—but not the verdict.
The baseline records unified maps, machine-structured node/endpoint registries, manifests and checksums. This makes later deltas more reviewable while keeping validation separate.
One coherent stack
A structured view of system layers rather than scattered notes.
Machine-structured nodes
Important system nodes represented in a reviewable registry.
Candidate endpoints
Architecture-adjacent terminal points structured for comparison and reasoning.
Packaging + provenance
Versioning and checksums support artifact integrity discussion; self-issued integrity is not independent validation.
The baseline emerged in stages.
| Build band | Contribution | Interpretation |
|---|---|---|
| Baseline mapping | Early architecture maps, layered structure, dashboarded context | Initial systems decomposition. |
| Canonicalization / hardening | Reference overlays and canonical packaging | Polish and consistency—not external validation. |
| Pilot operations / adjudication | Intake, blind-review structures, decision/report/release packages | Operational discipline assets. |
| Current-LLM anatomy / handoff | Layer-complete anatomy, synthesis and handoff review | Structured current-system understanding. |
| Track B lab / readiness | Measurement, protocol, recipe/data/compute blueprint, runtime lab, endpoint bench, comparative validation | Readiness architecture for future interventions/tests. |
The baseline was strongest in structure and measurement—not mass-scale deployment.
These are the baseline document’s own internal maturity labels. They are useful as a self-assessment snapshot, not an independent rating.
| Capability band | Baseline label | Interpretation |
|---|---|---|
| Architecture understanding | Strong | Serious layered decomposition baseline rather than generic narrative. |
| Structured artifact maturity | Strong | Registries, manifests, templates, dashboards and synthesis packs present. |
| Evaluation / measurement assets | Strong | Measurement, signal taxonomy, validation and safety logic are asset classes. |
| Runtime / control logic | Good | Runtime lab, lane topology, replay and isolation logic structured. |
| Data / recipe / compute blueprinting | Good | Design assets beyond ad hoc notes; not hyperscale-training proof. |
| Mass-scale deployment | Intentionally not claimed | Baseline is not a frontier deployment announcement. |
The baseline is more useful when its boundary is visible.
Not a finished frontier company
The baseline does not announce a mass-scale frontier training or consumer inference operation.
Confidential mechanics excluded
Private intervention logic, proprietary execution methods and future-state pathways remain outside the public baseline.
Only explicit claims count
If a capability is not explicitly established in the baseline, it should not be inferred from surrounding documentation density.
A clean before-state makes Phase 3 changes measurable.
The public HUAI architecture uses this baseline as one technical substrate. Phase 3 can choose an intervention, record the before-state, implement, measure and compare—without pretending documentation alone is the result.
HUAI
See how the systems baseline connects to Phase 1 signals and Phase 3 operationalization.
HUAI →Reference taxonomyLLM Anatomy
Inspect the separate 21-slot / 529-endpoint company/system reference atlas.
LLM Anatomy →Intervention questionsInnovation
See the architecture surfaces that Phase 3 can selectively implement and test.
Innovation →Independent executionPhase 3 Partner
Choose a bounded systems object for validation, pilot or collaboration.
Partnership →