Cognitive-OS-Wantsong/docs/MODEL_MANAGEMENT_V0.md

101 lines
2.8 KiB
Markdown

# Model Management V0
status: draft_callable_kernel
date: 2026-06-20
## Purpose
Model management exists to make models callable in real cognitive processing. It is not the product itself.
M0-M1 defines enough structure to register and call the first two models:
- `qpi`;
- `intellectual_archaeology`.
## Model Status
Allowed statuses:
- `seed`: source material or rough idea, not callable.
- `draft`: model card exists but call boundary is incomplete.
- `callable`: usable in a manual runtime with known limits.
- `core_candidate`: repeatedly useful in real runs, pending milestone review.
- `core_active`: accepted by owner after multiple useful runs and milestone review.
- `needs_rework`: known boundary or quality issue blocks normal use.
- `deprecated`: no longer recommended for new runs.
- `archived`: retained only for history or reference.
M0-M1 may use only `draft` and `callable`.
## Model Levels
Allowed levels:
- `L0_source_material`: source article, discussion, report, or note.
- `L1_candidate_model`: extracted candidate model, not yet callable.
- `L2_callable_model`: callable manually with a known input/output contract.
- `L3_core_model`: future state after repeated useful runs and milestone review.
- `L4_long_term_core_model`: future state outside v0.1.
M0-M1 may use only `L1_candidate_model` and `L2_callable_model`.
## Required Model Card Fields
Each model card must state:
- `model_id`;
- file path;
- runtime scope;
- governance level;
- default role;
- allowed roles;
- one-sentence definition;
- call conditions;
- no-call conditions;
- input types;
- output contract;
- depth or cost controls;
- common misuses;
- reader translation requirement;
- source seed notes;
- current limits.
## Registry Requirements
`models/model-registry.json` is the only structured registry in M0-M1.
The registry must support:
- locating the model file;
- determining default role;
- identifying whether a model can be primary, support, contrast, calibration, translation, or routing;
- avoiding over-call by cost and no-call fields;
- pairing QPI with later deep processing;
- requiring reader translation when internal output is too model-heavy.
The registry is not a full schema system.
## Upgrade Rule
A model cannot be upgraded because JSON parses, docs exist, or a selector could call it.
Future upgrades require:
- real run evidence;
- owner review;
- quality review at a milestone;
- clear usefulness beyond one sample;
- explicit decision entry.
## Stop Conditions
Stop model-management expansion if work starts to require:
- complete evidence matrices before the first real runtime sample;
- a full regression suite;
- full selector calibration;
- lifecycle promotion machinery;
- proof that every boundary case is covered.
Those belong later, after the runtime proves value.