# Project Brief

status: draft_runtime_boundary
date: 2026-06-20

## One-Sentence Definition

`Cognitive-OS-Wantsong` is a lightweight cognitive-processing runtime that uses Wantsong's cognitive models to process complex inputs into internal analysis and reader-facing explanations.

## Restart Basis

The old project proved useful file-first model-governance mechanisms, but its operating gravity drifted toward review evidence, selector calibration, regression cases, and lifecycle caution.

This project starts from the opposite order:

```text
run real cognitive workflow -> inspect output value -> stabilize useful model pieces
```

M0-M1 therefore creates only the boundary and model-management kernel required before any runtime sample.

## Product Layers

The project has three layers and one review lane:

1. Model management layer: register, describe, call, combine, retire, and repair model assets.
2. Cognitive-processing runtime layer: intake, QPI, lens orchestration, deep processing, synthesis, calibration, and feedback.
3. Reader translation layer: turn internal model output into understandable reader-facing expression.
4. Milestone quality lane: Local CCRA reviews milestone outputs, not every small edit.

## M0 Acceptance

M0 is acceptable when:

- the repo has an independent git boundary and remote;
- the old project is documented only as reference and seed archive;
- forbidden old process artifacts are not present;
- v0.1 non-goals are explicit.

## M1 Acceptance

M1 is acceptable when:

- QPI and Intellectual Archaeology have callable model cards;
- `models/model-registry.json` supports future primary/support selection;
- model statuses and levels are defined without creating a full lifecycle system;
- no full selector, full schema validator, sample run, or regression suite exists.

## Non-Goals For v0.1 Startup

- full model library;
- 10-model expansion;
- complete schema validator;
- database;
- backend service;
- frontend UI;
- vector database;
- full RAG;
- full regression suite;
- lifecycle promotion;
- before/after selector diff;
- Web review bundle by default;
- old report migration;
- old review-bundle migration;
- old Local CCRA history migration;
- Round Conductor.

## Failure Signals

M0-M1 is drifting if:

- QPI becomes the main governance object again;
- review artifacts outgrow runtime artifacts before the first sample;
- Codex creates `round-*`, `ccra_review_bundle/`, or full regression material;
- model management starts requiring perfect evidence matrices before any runtime use;
- reader translation is ignored as a product layer.