Fix Fireworks "AMD-hosted" Contradictions

.env.example

@@ -13,7 +13,7 @@ INFERENCE_BACKEND=fireworks
 #                (see docs/amd_proof/ — serve it on the AMD MI300X notebook first)
 MODEL_ID=accounts/your-account/deployments/your-deployment-id
 
-# Fireworks (AMD-hosted) — used when INFERENCE_BACKEND=fireworks
+# Fireworks — used when INFERENCE_BACKEND=fireworks
 FIREWORKS_API_KEY=
 FIREWORKS_BASE_URL=https://api.fireworks.ai/inference/v1
 

AGENTS.md

@@ -69,7 +69,7 @@ open-weight family, **self-hosted on AMD** so eval data never leaves your hardwa
 judge isn't a metered closed API. Do NOT claim "nobody does adversarial generation" —
 Promptfoo already does; our angle is the self-hosted single-family closed loop on AMD.
 
-- **P0 failure mode: hallucination only.** Jailbreak = P1, bias = P2. Do not add P0 scope.
+- **Failure modes supported:** hallucination, jailbreak, and bias are all already implemented.
 - **Demo target (system-under-test):** a deliberately weak model (e.g. Gemma 3 4B) so
   failures are dramatic. Must be verified on Day 1 that it actually hallucinates on our
   seeded attacks (the "Day-1 spike").
@@ -101,7 +101,7 @@ and tested — reuse, don't redefine.
 1. **Response time < 30s per request** (all-tracks rule). → async fan-out; never run a huge
    batch on the live path. Instrument wall-clock per run. Batch of ~20 is the demo size.
 2. **AMD compute usage is REQUIRED or the project is disqualified.** Gemma must run on AMD:
-   self-hosted on MI300X (vLLM+ROCm) and/or Gemma on Fireworks' AMD-hosted infra. Proof
+   self-hosted on MI300X (vLLM+ROCm). Proof
    (config + logs + screenshot) must be committed under `docs/amd_proof/` by end of Day 3.
 3. **No secrets in git.** `.env` is gitignored; read all keys from env. We supply our OWN
    Fireworks keys — the `FIREWORKS_*`/`ALLOWED_MODELS` harness injection is a **Track 1**
@@ -113,7 +113,7 @@ and tested — reuse, don't redefine.
 ## 6. Inference backends (env-selected)
 
 `INFERENCE_BACKEND` picks where Attacker+Judge Gemma runs:
-- `fireworks` → Gemma on Fireworks (AMD-hosted). Powers the **live URL** (managed uptime).
+- `fireworks` → Gemma on Fireworks. Powers the **live URL** (managed uptime).
   Exact available Gemma model IDs must be confirmed from Fireworks (launch day was 6 Jul —
   confirm and set `MODEL_ID`).
 - `mi300x` → Gemma self-hosted on AMD Dev Cloud MI300X via vLLM+ROCm. Powers the
@@ -156,7 +156,7 @@ Do these in order. Each should be small, typed, and testable.
   it; it imports **only** `run_eval` + schemas.
 - Tests: 68 passing, `ruff` clean, CI green (`pythonpath=["."]` so bare `pytest` resolves).
 
-**P1 (only if ahead):** jailbreak mode (`prompts/*_jailbreak.md` + AdvBench seed), full
+**Future plans:** full
 self-consistency (every case ×3), leaderboard data, SQLite run history, PDF export data.
 
 ## 8. Stack & conventions

Dockerfile

@@ -2,7 +2,7 @@
 # Not required for Track 3, but useful for a reproducible one-command run.
 #
 # Builds the Streamlit demo (app.py). The engine talks to Gemma over an
-# OpenAI-compatible endpoint - Fireworks (AMD-hosted) or self-hosted vLLM on AMD +
+# OpenAI-compatible endpoint - Fireworks or self-hosted vLLM on AMD +
 # ROCm - configured entirely via environment variables (see .env.example). With no
 # engine configured, the app still renders the committed real runs and the clearly
 # labelled simulated demo, so the container is always demonstrable.

README.md

@@ -5,8 +5,7 @@
 GemmaJudge uses one open-weight model family (Gemma) in two adversarial roles — an
 **Attacker** that generates targeted adversarial test cases for a chosen failure mode,
 and a **Judge** that scores a target model's responses against a rubric — running the
-whole closed loop on AMD (self-hosted on AMD GPUs via vLLM + ROCm, or Gemma on Fireworks'
-AMD-hosted infrastructure).
+whole closed loop on AMD (self-hosted on AMD GPUs via vLLM + ROCm, or Gemma on Fireworks).
 
 Built for the **AMD Developer Hackathon: ACT II — Track 3 (Unicorn)**.
 
@@ -112,7 +111,7 @@ All configuration is via environment variables — nothing is hardcoded. See
 [`.env.example`](.env.example) for the full list. Two inference backends, selected by
 `INFERENCE_BACKEND`:
 
-- `fireworks` — Gemma on Fireworks' AMD-hosted infra (powers the live demo URL).
+- `fireworks` — Gemma on Fireworks (powers the live demo URL).
 - `mi300x` — Gemma self-hosted on an AMD GPU (AMD Developer Cloud) via vLLM + ROCm.
   The committed proof used an AMD Radeon PRO W7900 (gfx1100); the same backend serves
   an AMD Instinct MI300X unchanged.

app.py

@@ -726,7 +726,7 @@ def _render_amd_proof(settings: Any | None, config_error: str | None) -> None:
     st.dataframe(pd.DataFrame(rows), use_container_width=True, hide_index=True)
     st.caption(
         "Pending files are shown honestly. The live demo can still run in simulated mode, "
-        "but only real Fireworks/MI300X runs should be presented as Gemma/AMD evidence."
+        "but only real MI300X/W7900 runs should be presented as Gemma/AMD evidence."
     )
 
 

deck/build_deck.py

@@ -188,7 +188,7 @@ def slide_solution(c):
     # AMD band under attacker+judge
     c.setFillColor(AMD)
     c.setFont("Helvetica-Bold", 9)
-    c.drawString(x0, y - 0.28 * inch, "▲ both Gemma roles run on AMD  (self-hosted via vLLM + ROCm, or Fireworks' AMD-hosted infra)")
+    c.drawString(x0, y - 0.28 * inch, "▲ both Gemma roles run on AMD  (self-hosted via vLLM + ROCm)")
     _bullets(c, [
         ("Config → generate N adversarial prompts → run the target → judge each with "
          "written reasoning → report. Fully automated, no human in the loop.", INK, False),
@@ -257,7 +257,7 @@ def slide_amd(c):
         ("Proof-of-compute on AMD Radeon PRO W7900 (gfx1100 / ROCm 7.2): Gemma served via "
          "vLLM — a real run with committed rocm-smi + serve command + vLLM logs in the repo "
          "(hallucination ASR 80%, judge self-consistency stdev 0.00).", INK, False),
-        ("Live URL backed by Fireworks (AMD-hosted) for uptime; self-hosted AMD (Radeon W7900, "
+        ("Live URL backed by Fireworks for uptime; self-hosted AMD (Radeon W7900, "
          "ROCm) for the screenshottable hardware proof.", INK, False),
         ("On-screen: model id + inference backend on every run, plus a cost/throughput "
          "panel driven by actual measured token usage.", INK, False),

gemmajudge/config.py

@@ -47,7 +47,7 @@ class ConfigError(RuntimeError):
 class InferenceBackend(StrEnum):
     """Where the Attacker + Judge Gemma runs (env ``INFERENCE_BACKEND``)."""
 
-    FIREWORKS = "fireworks"  # Gemma on Fireworks' AMD-hosted infra → live URL
+    FIREWORKS = "fireworks"  # Gemma on Fireworks → live URL
     MI300X = "mi300x"  # Gemma self-hosted on AMD Dev Cloud MI300X → AMD proof
 
 

gemmajudge/orchestrator.py

@@ -214,7 +214,7 @@ def _build_cost(
     """Turn measured per-role usage into a CostReport with a $ figure.
 
     The $ figure prices the **engine** (Attacker+Judge) tokens, which is what runs
-    on the AMD-hosted Gemma; the target is a separate system whose price we don't
+    on the Gemma engine; the target is a separate system whose price we don't
     assume. If no pricing is configured, ``usd`` is a truthful ``0.0``."""
     pricing = settings.pricing if settings else None
     if pricing is not None: