# ForgeAI Studio — Orion Workflow
> Owner: Anthony Kautz · Orchestrator: `orion@2.2.0` · Last updated: 2026-05-20

---

## what this is

A production-grade AI persona pipeline. Personas are designed, red-teamed, evaluated, and promoted through a rigorous, structured sequence before reaching production.

## pipeline sequence

```
Persona Pack → System Prompt → Red-Team → Patch → Re-test → Eval Pack → Judge Prompt → Production
```

## current state

| Asset | Version | Status |
|---|---|---|
| `orion` | 2.2.0 | ✅ Production |
| `sage` | 1.0.0 | Draft — 41/42 eval PASS |
| `maura` | 1.x.x | ❌ Retired |

## immediate next action

Build a system prompt for `sage@1.0.0`, incorporating its eval run notes, then red-team it.

## directory structure

```
forgeai-studio/
├── README.md
├── personas/
│   ├── orion/
│   │   └── orion@2.2.0.md          # Orchestrator persona — production
│   └── sage/
│       ├── sage@1.0.0.md                    # Draft — 41/42 eval PASS
│       └── sage-system-prompt@1.0.0.md      # real, derived via Stage 03 — red-team 10/10, eval hard gates PASS
├── eval-packs/
│   └── sage-eval-pack@1.0.0.md     # 42 real cases, AX-01–AX-07
├── judge-prompts/
│   └── sage-judge@1.0.0.md         # real, derived via Stage 07 — its first real content
├── pipeline-specs/
│   └── orion-workflow-pipeline.md  # Full pipeline spec + CI/CD target
├── red-team/                        # Red-team reports (populate per run)
└── registry/
    └── failure-registry.md         # All known issues + fixes
```

## key principles

- Every artifact carries a **tag block**: `name@version | eval-status | temp | max_tokens | red-team state | bias facet | eval-pack state`
- **Bias facet check** is mandatory on every persona — no exceptions
- **Hard gates** (AX-05, AX-06) are non-negotiable production blockers
- **Judge consistency** is prompt-enforced — the model in use accepts no `temperature` parameter; the non-negotiable guarantee is the hard-gate pass/fail computed in code from real scores, never trusted from the judge's own self-report
- **Repair loops** are mandatory on schema validation failure — not best-effort
- **Semantic versioning** on all assets; all changes logged in Failure Registry

---

*Orion ready. State the objective and constraints.*