# Failure Registry
> ForgeAI Studio · Orion workflow · Owner: Anthony Kautz

Versioned log of all known failure modes, red-team exploits, and confirmed fixes across all pipeline assets.

---

## registry format

| Field | Description |
|---|---|
| **Issue ID** | `RT-{asset}-{seq}` for red-team; `RT-{seq}` for current active persona |
| **Asset** | Persona or prompt affected |
| **Version discovered** | Asset version when issue was first found |
| **Type** | adversarial · paraphrase · jailbreak · schema · drift · injection · bias |
| **Severity** | P0 (blocker) · P1 (high) · P2 (medium) · P3 (low) |
| **Status** | Open · Fixed · Accepted-Risk · Waived |
| **Fix version** | Asset version where fix was applied |
| **Notes** | Additional context |

---

## maura@1.x.x — (retired)

> Superseded by morgan. Issues from maura's red-team informed morgan's initial design. Registry entries archived.

| Issue ID | Status | Notes |
|---|---|---|
| (maura issues) | Archived | Superseded |

---

## orion@2.2.0 — known failure modes

| Mode | Failure type | Mitigation |
|---|---|---|
| Schema validation failure | Repair loop triggered | Repair loops mandatory; surface metadata to eval log |
| Prompt injection | Halt + surface to operator | Never process without explicit authorization |
| Persona drift | Re-anchor after 3 consecutive deviating turns | Visible re-anchor; operator surfaced |
| Bias facet fail | Persona blocked from emission | Mandatory Critic–Author; fix before emit |
| Hard gate fail | Production blocked | Repair loop; escalate if re-test fails |

---

## open issues

*None as of 2026-05-20.*

> Next anticipated issue: Judge calibration gap (if golden-set agreement <85% on first run).

---

*Update this registry on every red-team run, patch cycle, and Judge calibration event.*