docs(adr-002): re-adjudicate Groq model to servable qwen/qwen3.8-27b (APA-47)

[Aparnap2]

What changed

Re-adjudicate ADR-002 to a currently servable production model. Config + docs only.

ADR-002's Option A adjudication contracted llama-3.1-8b-instant, matching the code default. That model has since been retired from Groq, so the documented production config would fail every inference call. Production model is now qwen/qwen3.8-27b.

Canonical string moved in the four places that must agree:

File	Before	After
apps/api/internal/investigate/orchestrate/groq_model.go	llama-3.1-8b-instant	qwen/qwen3.8-27b
README.md	llama-3.1-8b-instant	qwen/qwen3.8-27b
docs/adr/002-groq-llm-provider.md (## Decision)	llama-3.1-8b-instant	qwen/qwen3.8-27b
.env.example	openai/gpt-oss-20b	qwen/qwen3.8-27b

.env.example had additionally drifted to the stale pre-adjudication value; it now matches code/ADR/README.

No ModelClient behavior change. No response_format/JSON mode, no reasoning parameter, no prompt change, no model-specific response parsing, no new dependencies. go.mod/go.sum untouched. The model remains untrusted input behind the unchanged deterministic validation/grounding boundary. No changes to eval/, eval-v1, the A–P corpus, E0–E3, hard gates, scoring, or any APA-11/12, retry, tenant, OCR, workflow, or agent contract.

Evidence

Servability confirmed before any edit, through the actual configured provider (defaultGroqBaseURL) with the real key from the untracked .env, replicating the client's exact wire format (max_tokens, temperature: 0, stream: false):

Requested	HTTP	Echoed model	choices[0].message.content
qwen/qwen3.8-27b	200	qwen/qwen3.8-27b	non-empty ("OK")
llama-3.1-8b-instant	404 model_not_found	—	empty

Account model list: 11 models, zero llama-3.x, qwen/qwen3.8-27b present. Premise independently corroborated.

The .env key was never printed, never staged, never committed; it is still gitignored and untracked.

Test expectation updates (renames only, no weakening). Two APA-45 literals that pin the canonical string were renamed to the new value:

• groq_env_contract_test.go → canonicalModelString
• groq_provider_selection_test.go → canonicalGroqModel

Provenance comments updated to record the APA-47 re-adjudication. No assertion weakened, skipped, or deleted. RED was confirmed first: flipping only the literals failed 5 tests across both packages (TestGroqContract_DefaultModelIsCanonicalString, TestGroqContract_DocsAgreeOnCanonicalModelString, TestGroqContract_EmptyModelResolvesToCanonicalDefault, TestAgentProvider_GroqEnvSelectsGroqClientNotMock, TestAgentProvider_OutboundCarriesCanonicalModelString) — proving the coupling assertions are load-bearing. All pass after the move, including the ADR/README/code coupling test TestGroqContract_DocsAgreeOnCanonicalModelString.

overrideGroqModel = "llama-3.3-70b-versatile" was deliberately not changed: it is a deliberately non-canonical override fixture on a loopback stub that proves GROQ_MODEL resolution wins, never dialled.

Blockers

None for this change. Two out-of-scope stale references found, not touched (they are outside this file set and one is a dated historical record — flagging rather than silently rewriting):

• docs/specs/09_DECISION_LOG.md:223,226 — records the 2026-09-26 Option A canonical string. Wants a new dated APA-47 entry, not a rewrite of the old one. 09 is architect-owned.
• docs/specs/10_PHASE3_E2E_QUALIFICATION.md:161 — names the old default. Note this line also says "retry 3x", which was already stale before this PR (the client makes 2 attempts total); pre-existing, unrelated.

Next step

Review and merge. Suggested follow-up: a dated APA-47 entry in 09_DECISION_LOG.md (architect-owned) so the decision log does not contradict the ADR.

Gates on c0fea27: gofmt -l clean · go vet ./... clean · go build ./... clean · go test ./... 43 pkgs, 0 fail · uv run pytest tests/unit 27 passed · uv run ruff check . clean · git diff main -- eval/ tests/ empty.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 89688228-db18-47eb-ac34-a0b164117f2f

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}