96 lines
2.6 KiB
Markdown
96 lines
2.6 KiB
Markdown
# Cognitive Workflow V0
|
|
|
|
status: draft_workflow_contract
|
|
date: 2026-06-20
|
|
|
|
This file describes the intended runtime loop. It is not a sample run and does not authorize M0-M1 to create prompt files or examples.
|
|
|
|
## Runtime Flow
|
|
|
|
```text
|
|
1. Intake / Value Assessment
|
|
2. QPI
|
|
3. Lens Orchestrator
|
|
4. Deep Processing
|
|
5. Synthesis & Calibration
|
|
6. Feedback & Asset Decision
|
|
7. Reader Translation
|
|
```
|
|
|
|
## 1. Intake / Value Assessment
|
|
|
|
Purpose: decide whether the input deserves cognitive processing and what depth budget is justified.
|
|
|
|
Expected output:
|
|
|
|
- processing level: `L1 | L2 | L3 | L4`;
|
|
- reason for depth;
|
|
- whether heavy processing is justified;
|
|
- required context gaps.
|
|
|
|
## 2. QPI
|
|
|
|
Purpose: classify the issue framing before deeper work.
|
|
|
|
QPI outputs:
|
|
|
|
- `question | problem | issue | mixed | no_call`;
|
|
- owner and scenario;
|
|
- dominant scarcity;
|
|
- missing context;
|
|
- misframing risks;
|
|
- recommended next step.
|
|
|
|
QPI does not solve the problem. It only controls routing and prevents framing mistakes.
|
|
|
|
## 3. Lens Orchestrator
|
|
|
|
Purpose: choose a small set of models/lenses for this run.
|
|
|
|
Default call budget:
|
|
|
|
- one primary model;
|
|
- two or three support or contrast models at most;
|
|
- one calibration lens when needed;
|
|
- one reader translation layer when producing external-facing text.
|
|
|
|
M0-M1 only prepares the fields needed for this later choice. It does not implement a selector.
|
|
|
|
## 4. Deep Processing
|
|
|
|
Purpose: run the selected primary/depth model.
|
|
|
|
For startup, the only depth model is `intellectual_archaeology`. It should run only when the input has enough value, complexity, or reuse potential.
|
|
|
|
Depth processing must obey minimum sufficient depth. Continuing deeper is justified only when it changes judgment, path, validation, action boundary, or asset decision.
|
|
|
|
## 5. Synthesis & Calibration
|
|
|
|
Purpose: combine outputs and prevent a single model from over-claiming.
|
|
|
|
Expected output:
|
|
|
|
- main judgment;
|
|
- supporting reasoning;
|
|
- conflicts between model outputs;
|
|
- evidence level;
|
|
- action boundary;
|
|
- what would change the conclusion.
|
|
|
|
## 6. Feedback & Asset Decision
|
|
|
|
Purpose: decide whether the run produced reusable learning.
|
|
|
|
Expected output:
|
|
|
|
- whether the output was useful;
|
|
- whether the run should become a sample later;
|
|
- whether a model card needs a future repair;
|
|
- whether a new model need was exposed.
|
|
|
|
## 7. Reader Translation
|
|
|
|
Purpose: turn internal trace into a reader-facing explanation.
|
|
|
|
The internal output can contain model vocabulary and deep structure. The reader output must explain the actual issue, mechanism, example, boundary, and next way to think in language a non-model-maintainer can follow.
|