poc(5): model routing bench — same atom, four tiers, real numbers #45

Open
opened 2026-07-11 18:17:38 +02:00 by arcodange · 0 comments
Owner

Part of the AI back-office PRDPOC-5 + model-fleet. Closes decision D5 (OCR provider).

Deliverables

  • Run the SAME extraction atom across the four tiers on the full golden set: Pi SLM (Qwen3-class 1.7–4B GGUF, GBNF), M4 local (hermes/Ornith + Ollama candidates Qwen3 14B/30B-A3B, Mistral Small), Mistral API, Claude API.
  • Scanned-document subset benched across OCR contenders (Mistral OCR / local vision / Tesseract baseline) → close D5.
  • Published table: per-field accuracy × latency × cost/doc × availability per tier; dual-extraction disagreement-rate baseline.
  • Outcome written into the atoms' model_policy in the registry (#38) + the PRD agent-catalog matrix statuses updated (🧪/reassigned).

Acceptance: the table exists with real numbers; routing policy v1 committed to the registry; PRD STATUS updated.


Execution — Blocked by: #39 (golden set), #40 (the atom) · Blocks: #47 (consumes routing policy v1); closes D5 → note it via factory#22 · Human gates: M4 awake with Ollama models pulled; Mistral OCR access if benched (else Tesseract-vs-local only, documented).
Start: claude --worktree routing-bench from the erp repo root; read fleet/atoms/invoice-extract/atom.yaml (model_policy slot) + the golden-set scoring script from #39.
Done means: evidence comment here (the full table, real numbers) → PR merged with Closes #45 → PRD agent-catalog/STATUS updated in a paired factory PR.

Part of the **AI back-office PRD** — [POC-5](https://gitea.arcodange.lab/arcodange-org/factory/src/branch/main/vibe/PRD/ai-back-office/poc-plan.md) + [model-fleet](https://gitea.arcodange.lab/arcodange-org/factory/src/branch/main/vibe/PRD/ai-back-office/model-fleet.md). Closes decision **D5** (OCR provider). **Deliverables** - Run the SAME extraction atom across the four tiers on the full golden set: Pi SLM (Qwen3-class 1.7–4B GGUF, GBNF), M4 local (hermes/Ornith + Ollama candidates Qwen3 14B/30B-A3B, Mistral Small), Mistral API, Claude API. - Scanned-document subset benched across OCR contenders (Mistral OCR / local vision / Tesseract baseline) → close D5. - Published table: per-field accuracy × latency × cost/doc × availability per tier; dual-extraction disagreement-rate baseline. - Outcome written into the atoms' `model_policy` in the registry (#38) + the PRD agent-catalog matrix statuses updated (🧪→✅/reassigned). **Acceptance:** the table exists with real numbers; routing policy v1 committed to the registry; PRD STATUS updated. --- **Execution** — Blocked by: #39 (golden set), #40 (the atom) · Blocks: #47 (consumes routing policy v1); closes D5 → note it via factory#22 · Human gates: M4 awake with Ollama models pulled; Mistral OCR access if benched (else Tesseract-vs-local only, documented). **Start:** `claude --worktree routing-bench` from the erp repo root; read `fleet/atoms/invoice-extract/atom.yaml` (`model_policy` slot) + the golden-set scoring script from #39. **Done means:** evidence comment here (the full table, real numbers) → PR merged with `Closes #45` → PRD agent-catalog/STATUS updated in a paired factory PR.
arcodange added this to the P1 — Flagship pipeline milestone 2026-07-11 18:17:38 +02:00
Sign in to join this conversation.
No labels
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: arcodange-org/erp#45