docs,test: drop the withdrawn Granite auxiliary and name granite-4-h-small for R1 (WO-16) — VOTE CAST 2026-07-28

README.md

@@ -139,7 +139,7 @@ Our team's operational playbook for using Bob — prompt patterns, export workfl
 | **Frontend**    | Next.js 14 (App Router) + TypeScript + Tailwind + shadcn/ui + Recharts                                 | Modern React, fast static + SSR, great DX for visualizations                    |
 | **Backend**     | FastAPI + Python 3.11 + Pydantic v2 + SQLAlchemy 2 + asyncpg                                           | Async by default, type-safe, fits a Python AI pipeline natively                 |
 | **AI Pipeline** | spaCy 3.7 (`en_core_web_lg`) + sentence-transformers (`all-mpnet-base-v2`) + scikit-learn + umap-learn | Industry-standard English NLP; strong 768-dim semantic embeddings; reproducible |
-| **LLM**         | IBM Watsonx (Llama 3.3 70B + Granite 3 8B)                                                             | End-to-end IBM stack; full-IBM points; strong creative English generation       |
+| **LLM**         | IBM Watsonx (`meta-llama/llama-3-3-70b-instruct`) — sole model in use; `ibm/granite-4-h-small` declared as an unexercised fallback | Every generation runs on IBM Watsonx; strong creative English generation        |
 | **Database**    | PostgreSQL 16 + pgvector (Supabase)                                                                    | Single DB for relational + vector; HNSW index for fast RAG                      |
 | **Crypto**      | python-jose, ES256 (ECDSA P-256)                                                                       | Standard JWS; small signatures; native browser verification                     |
 | **Hosting**     | Vercel (frontend) + Railway (backend) + Supabase (DB)                                                  | Zero-ops, free or near-free tiers, push-to-deploy                               |

backend/tests/test_watsonx_client.py

@@ -96,7 +96,7 @@ def test_generate_retries_with_backoff_then_succeeds(monkeypatch):
         patch("app.services.watsonx_client.Credentials"),
         patch("app.services.watsonx_client.time.sleep", side_effect=sleeps.append),
     ):
-        text = generate("prompt", None, "ibm/granite-3-8b-instruct")
+        text = generate("prompt", None, "meta-llama/llama-3-3-70b-instruct")
 
     assert text == "ok"
     assert sleeps == [1.0]
@@ -118,7 +118,7 @@ def test_generate_exhausts_retries(monkeypatch):
         patch("app.services.watsonx_client.time.sleep", side_effect=sleeps.append),
         pytest.raises(WatsonxError, match="failed after"),
     ):
-        generate("prompt", None, "ibm/granite-3-8b-instruct")
+        generate("prompt", None, "meta-llama/llama-3-3-70b-instruct")
 
     assert sleeps == list(_RETRY_DELAYS_SECONDS)
     assert mock_model.chat.call_count == 1 + len(_RETRY_DELAYS_SECONDS)
@@ -142,7 +142,7 @@ def test_generate_hard_timeout(monkeypatch):
         future = MagicMock()
         future.result.side_effect = FuturesTimeoutError()
         executor.submit.return_value = future
-        generate("prompt", None, "ibm/granite-3-8b-instruct")
+        generate("prompt", None, "meta-llama/llama-3-3-70b-instruct")
 
     assert sleeps == list(_RETRY_DELAYS_SECONDS)
     assert future.result.call_count == 1 + len(_RETRY_DELAYS_SECONDS)
@@ -161,7 +161,7 @@ def test_generate_live_watsonx():
     text = generate(
         prompt="Reply with exactly one word: pong",
         system_prompt="You are a terse assistant. Answer with a single word only.",
-        model_id="ibm/granite-3-8b-instruct",
+        model_id="meta-llama/llama-3-3-70b-instruct",
         params={"max_tokens": 16, "temperature": 0},
     )
     assert isinstance(text, str)

docs/MVP.md

@@ -51,8 +51,9 @@ These three decisions shape everything else. They are locked; changing one is a
 
 - **LLM: IBM Watsonx, end-to-end.**
   - Primary (creative writing): `meta-llama/llama-3-3-70b-instruct`
-  - Auxiliary (structured/classification tasks): `ibm/granite-3-8b-instruct`
-  - Baseline ("vanilla"): the **same** `llama-3-3-70b` but **without** style conditioning.
+  - Auxiliary: **none**. `ibm/granite-3-8b-instruct` was originally locked here for structured/classification tasks, but it was never wired into any code path and it is **not available to this project**: it is absent from the `eu-de` catalogue and, where it was still listed (`us-south`), its lifecycle is `withdrawn` since 2026-03-31. Revised 2026-07-28 — see `docs/decision_log.md`. Every call the product makes goes to the primary model.
+  - Escalation backup (R1), **not in use**: `ibm/granite-4-h-small` — confirmed `available` in `eu-de` on 2026-07-28. Nothing calls it today; it exists only as the named fallback if the Sprint 1 voice-match gate fails (see §11 R1). Adopting it would be a separate decision.
+  - Baseline ("vanilla"): the **same** `llama-3-3-70b` but **without** style conditioning. The escalation backup never touches this comparison.
   - *Why same model on both sides:* it's an honest comparison. The only variable is our style conditioning — judges notice and respect that.
 
 - **Language: everything in English — UI, docs, video, and the generated text.**
@@ -324,8 +325,9 @@ Each is justified so we don't reopen the discussion mid-sprint.
 
 ### LLM
 - `ibm-watsonx-ai` SDK
-- Creative: `meta-llama/llama-3-3-70b-instruct`
-- Auxiliary: `ibm/granite-3-8b-instruct`
+- Creative: `meta-llama/llama-3-3-70b-instruct` — the only model the product calls
+- Auxiliary: none (the originally locked `ibm/granite-3-8b-instruct` is unavailable; see §2)
+- Escalation backup, **not wired in**: `ibm/granite-4-h-small` (available in `eu-de`, ratified 2026-07-28). No code path, no env var, no call — only R1's named fallback
 - **SPRINT 1 TASK:** validate voice-matching quality with 5 real prompts (does the conditioned output read like the target author?); if Llama < 6/10 human eval, escalate (see R1).
 
 ### Database
@@ -509,9 +511,9 @@ The project is "done" when **all** of these are true:
 
 | # | Risk | Mitigation |
 |---|---|---|
-| **R1** | Llama-3.3-70b doesn't convincingly match a target voice | Validate Sprint 1 day 2. If <6/10, plan B: stronger conditioning / more RAG passages, `llama-3-1-405b`, or `granite-3-8b`; last resort Mistral Large via Watsonx. |
+| **R1** | Llama-3.3-70b doesn't convincingly match a target voice | Validate Sprint 1 day 2. If <6/10, plan B: stronger conditioning / more RAG passages first. **Model-swap escalation: `ibm/granite-4-h-small`** — ratified 2026-07-28, confirmed `available` in `eu-de` that day. The three backups this row used to name (`llama-3-1-405b`, `granite-3-8b`, Mistral Large) are all unreachable from `eu-de` and were removed. The escalation is **declared, not exercised**: nothing calls Granite 4 today, and it never becomes the vanilla baseline — the A/B stays `llama-3-3-70b` vs itself. |
 | **R2** | Stylistic metrics don't distinguish authors | Validate Sprint 1 with a blind A/B. If Austen ≈ Dickens (both 19thc British), add bigram/trigram distinctive features. (Poe is the easy separation.) |
-| **R3** | Generation >15s, demo doesn't flow | P3 measures latency Sprint 2 day 1. If >10s: switch parallel→sequential with optimistic loading, and/or use `granite-8b` for the baseline (faster). |
+| **R3** | Generation >15s, demo doesn't flow | P3 measures latency Sprint 2 day 1. If >10s: switch parallel→sequential with optimistic loading, shorten `max_tokens`, or fall back to the pre-recorded generations of R5. **R3 has no model-swap plan B and no replacement for one.** The old "use a faster `granite-8b` for the baseline" is void twice over: that model does not exist for us (see §2), *and* swapping only the baseline would break the A/B whatever model replaced it — the comparison is only honest while both columns run the same model. R1's `ibm/granite-4-h-small` is **not** an R3 mitigation for the same reason. The three levers above are the whole of R3. |
 | **R4** | Live demo fails during recording | Sprint 3: record each step separately as an editable fallback. Final video may mix live + pre-recorded. |
 | **R5** | Watsonx rate-limit / load spike | Exponential backoff (Sprint 2). For the demo, pre-record two famous generations as a local cache fallback. |
 | **R6** | A teammate is sick for a week | Each role has a designated backup; pair on critical pieces; the async daily keeps the backup current. |

docs/ONBOARDING.md

@@ -229,7 +229,7 @@ Lives at **`docs/decision_log.md`**. 4-column table:
 Read them in `docs/decision_log.md`. The key ones to keep in your head:
 
 - Idea = **AutorIA** (auditable authorship layer)
-- LLM = **IBM Watsonx end-to-end** (Llama 3.3 70B + Granite 3 8B auxiliary)
+- LLM = **IBM Watsonx end-to-end** (`meta-llama/llama-3-3-70b-instruct` is the only model we call; the Granite 3 8B auxiliary was dropped 2026-07-28 — withdrawn upstream, absent from `eu-de`. `ibm/granite-4-h-small` is the *declared, unused* R1 escalation backup, not part of the active path)
 - **English everywhere** — app UI, repo, README, video, commits, and the generated text. Only the **internal team chat** stays Spanish.
 - Multi-author with **3 preloaded**: Jane Austen, Charles Dickens, Edgar Allan Poe
 - Stack: Next.js + FastAPI + Postgres+pgvector (Supabase) + JWS ES256
@@ -326,7 +326,7 @@ Full document at **`docs/MVP.md`**. Executive summary:
 
 | Decision | Value | Why |
 |---|---|---|
-| LLM | **IBM Watsonx end-to-end**. Primary: `meta-llama/llama-3-3-70b-instruct`. Auxiliary: `ibm/granite-3-8b-instruct` | IBM challenge = points for full-IBM stack |
+| LLM | **IBM Watsonx end-to-end**. Primary and only model in use: `meta-llama/llama-3-3-70b-instruct`. No auxiliary model (`ibm/granite-3-8b-instruct` was revised out on 2026-07-28 — unavailable in `eu-de`, withdrawn upstream). `ibm/granite-4-h-small` is the ratified R1 escalation backup — declared, never called | IBM challenge = points for running the whole product on Watsonx |
 | Comparison baseline | **Same model WITHOUT style conditioning** | Honest comparison = judges notice |
 | App UI + video + generated text | **English** | International IBM jury; judges can read the voice match directly |
 | Authors | **Austen, Dickens, Poe** (English, public domain) | Iconic, maximally distinct voices; instantly recognizable |
@@ -370,7 +370,8 @@ tiktoken (approx chunking)
 ```
 ibm-watsonx-ai SDK
 Creative model: meta-llama/llama-3-3-70b-instruct
-Auxiliary model: ibm/granite-3-8b-instruct
+Auxiliary model: none (Granite 3 8B revised out 2026-07-28 — unavailable in eu-de)
+R1 escalation backup (declared, not called): ibm/granite-4-h-small
 Sprint 1 task: validate voice-matching quality with 5 prompts (gate)
 ```
 
@@ -961,7 +962,7 @@ On Sunday, unfinished issues either move to "Sprint N+1" or back to "Backlog" if
 Your role has a backup designated (P1→P3, P2→P3, P3→P2). The daily async ensures your backup knows the state. Pair sessions ensure >1 person has touched critical pieces. Not ideal but the project doesn't collapse.
 
 ### "What if Watsonx doesn't match the author's voice well?"
-Plan B in MVP: stronger conditioning / more RAG passages, then `llama-3-1-405b-instruct` or `granite-3-8b-instruct` via Watsonx. If all fail, fallback to Mistral Large via Watsonx. **But**: we validate this Sprint 1 day 2, not Sprint 4.
+Plan B in MVP: stronger conditioning / more RAG passages first, then swap the model to **`ibm/granite-4-h-small`** (ratified 2026-07-28, confirmed available in `eu-de`). The escalation chain this FAQ used to name — `llama-3-1-405b-instruct`, `granite-3-8b-instruct`, Mistral Large — was fiction: none of the three is in the `eu-de` catalogue we run against, and Granite 3 8B is withdrawn upstream everywhere. Granite 4 H Small is a *declared* backup: we do not call it today, and it never replaces the vanilla baseline (that would break the A/B). **But**: we validate this Sprint 1 day 2, not Sprint 4.
 
 ### "What if a seemingly essential feature emerges that's not in the MVP?"
 5-min call with all 3, 2/3 vote. If it passes → Decision Log entry + add. If not → `docs/roadmap.md` for "post-July".

docs/architecture.md

@@ -36,7 +36,7 @@ graph TB
         sys[["AutorIA<br/>Style DNA · Conditioned Generation ·<br/>Authorship Passport"]]
     end
 
-    watsonx[("IBM Watsonx<br/>Llama 3.3 70B + Granite 3 8B<br/>[external LLM]")]
+    watsonx[("IBM Watsonx<br/>Llama 3.3 70B<br/>[external LLM]")]
     gutenberg[("Project Gutenberg<br/>public-domain corpus<br/>[external source]")]
 
     creator -->|"selects author, prompts,<br/>views side-by-side, downloads Passport"| sys
@@ -55,8 +55,11 @@ graph TB
   audience). Only needs the token + public JWKS.
 
 **External dependencies**
-- **IBM Watsonx** — the LLM provider (creative + auxiliary models). The only
-  hard runtime external dependency for generation.
+- **IBM Watsonx** — the LLM provider. A single model, `meta-llama/llama-3-3-70b-instruct`,
+  serves both branches of the side-by-side (conditioned and vanilla); there is no
+  auxiliary model. `ibm/granite-4-h-small` is the declared R1 escalation backup and is
+  never called on the active path (MVP §2 / §11). The only hard runtime external
+  dependency for generation.
 - **Project Gutenberg** — source of the three preloaded corpora; used **offline**
   during seeding, not at request time.