docs,test: drop the withdrawn Granite auxiliary and name granite-4-h-small for R1 (WO-16) — VOTE CAST 2026-07-28

[sergi-torres]

✅ VOTE CAST — 2026-07-28

This touches docs/MVP.md §2, marked LOCKED, and docs/MVP.md:4 requires 2/3 ratification. That vote was cast on 2026-07-28 by the repo owner (Sergi Torres, P1). The decision-log row is no longer pending: it is ratified, and it now also settles the escalation model that #102 left open.

What was voted: the R1 escalation model is ibm/granite-4-h-small — chosen as the successor the Watsonx catalogue itself points to for the withdrawn Granite, and because it keeps the full-IBM angle the challenge rewards. It is a backup only: nothing calls it, and it never enters the A/B. The side-by-side keeps comparing meta-llama/llama-3-3-70b-instruct with and without style conditioning — that does not change.

R3 gets no replacement. Its old plan B ("use the faster granite-8b for the baseline") is void twice over: the model does not exist for us, and swapping only the baseline would break the A/B whichever model replaced it. R3's mitigation is now the three non-model levers it already had — sequential generation with optimistic loading, shorter max_tokens, and the pre-recorded generations of R5 — and the docs say so instead of inventing a substitute.

What

1. Removes the Granite 3 8B auxiliary from the declared stack and points the integration test at the model actually in use.
2. Ratifies that removal and names ibm/granite-4-h-small as R1's escalation backup, declared but not exercised.

The finding is stronger than the issue's premise. It is not that ibm/granite-3-8b-instruct is missing from eu-de: it was withdrawn upstream on 2026-03-31 and exists in no region. Switching region does not recover it.

Measurement

Public Watsonx foundation-model catalogue — no credentials required. Re-measured 2026-07-28 for this update, HTTP 200, 18 models listed for eu-de:

GET https://eu-de.ml.cloud.ibm.com/ml/v1/foundation_model_specs?version=2024-05-01&limit=200

ibm/granite-4-h-small              -> available   (lifecycle: available since 2025-10-02, provider IBM)
meta-llama/llama-3-3-70b-instruct  -> available   (lifecycle: available since 2024-12-06)
ibm/granite-3-8b-instruct          -> NOT PRESENT in eu-de

Earlier six-region sweep (eu-de, eu-gb, jp-tok, au-syd, ca-tor, us-south), all HTTP 200:

Region	ibm/granite-3-8b-instruct
eu-de (ours)	absent
eu-gb, jp-tok, au-syd, ca-tor	absent
us-south	listed: available 2024-10-21 → deprecated 2025-11-24 → withdrawn 2026-03-31

Runtime evidence 2026-07-27 in eu-de:

meta-llama/llama-3-3-70b-instruct  -> OK: 'pong'
ibm/granite-3-8b-instruct          -> 404 model_not_supported

Why

Closes #98. Settles the open question of #102 (#102 is not closed by this PR — the escalation choice is recorded here, the issue stays for the owner to close).

In front of a jury that scores the stack, correcting the claim is safer than defending one that cannot be demonstrated live. README.md, docs/MVP.md §2 and the docs/architecture.md diagram all advertised "Llama 3.3 70B + Granite 3 8B", and half of it does not respond. Equally, an escalation chain whose three exits are all closed is worse than one honest, verified exit.

Changes

• docs/decision_log.md — the 2026-07-28 row is ratified (it was never merged, so it is rewritten in place rather than superseded by a second row; append-only integrity of main is untouched). It names Granite 4 H Small, carries the measurement as evidence, and states R3's lack of a plan B.
• docs/MVP.md §2 / §6 — active path stays single-model; the escalation backup is listed as declared, not wired in.
• docs/MVP.md §11 — R1 names Granite 4 H Small and marks it unexercised; R3 states it has no model-swap mitigation at all.
• README.md, docs/ONBOARDING.md — same distinction: model in use vs declared fallback.
• docs/architecture.md — the external-dependency bullet still read "creative + auxiliary models", a present-tense falsehood that no granite grep could catch because it names neither model. Fixed.
• backend/tests/test_watsonx_client.py — all four model_id literals target meta-llama/llama-3-3-70b-instruct (unchanged since the first push).

Rebased onto main (not merged) — the branch was behind. The docs/decision_log.md conflict was resolved keeping all rows from both sides.

How IBM Bob helped

• Custom Mode used: none
• Bob session export: n/a

Claude Code agents. On the first pass the agent stopped on a different finding instead of fixing it, per the dispatch rule: the entire R1 escalation chain did not exist — llama-3-1-405b and Mistral Large are also absent from eu-de. It filed #102 rather than picking a replacement on its own, because that is a separate decision with its own vote. This second pass records the vote.

Checklist

• Tests pass locally (make test)
• Lint passes (ruff check . + black --check . — green; no Python touched in the ratification commit)
• If schema changed (StyleProfile / Passport) → bumped version + updated docs/
• If new env var → added to .env.example
• If new/changed endpoint → updated docs/api_contract.yaml
• Documentation in the relevant README.md updated
• At least one teammate reviewed

This branch's tests have not been run. The executor worked in a worktree with pytest forbidden (the editable-install trap). Before merging, run pytest backend/tests/test_watsonx_client.py -q and above all confirm that test_generate_live_watsonx actually executes rather than self-skipping on missing credentials: a silent skip here would look like a pass.

Expected conflict in docs/decision_log.md: append-only, four writers. Keep all rows.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
author-ai	Ready	Preview, Comment	Jul 28, 2026 9:17pm