Part of the AI back-office PRD — POC-1, the flagship. Umbrella issue — integration + exit gate only; the parts ship via #38–#45.
Deliverables
Assemble the pipeline: books@ mail → T01 classify (cron script is fine at this stage; the Pi model comes with POC-2) → #40 dual extraction → validators → manifest (thirdparty check T04 included) → sandbox rehearsal + predicted-delta → approval → gated promote → GED attach (#43) → journal.
Approval channel: Telegram card if the gateway queue/cards are live (arcodange/telegram-gateway#1+#2), otherwise an interim CLI approval prompt — the gate semantics (approve / edit / reject-with-reason, reasons journaled into the golden set) must be identical either way.
Run journals: one JSONL per run, replayable (PRD architecture).
Exit gate (hard): 10 consecutive REAL supplier invoices recorded in prod with zero human field-corrections (approvals only); critical-field accuracy ≥98 % over the golden set; all injection fixtures quarantined; post-run snapshot history pure-append; provenance + compliance linter suites green. On pass: T02/T03 move to A2 in the registry + PRD agent-catalog statuses updated. Per qa-strategy, the exit-gate verdict is attested by an independent context-free verifier (no self-grading).
Fallback if failed: stay at A1 (agent drafts, human enters in UI) and iterate extraction only.
Execution — Blocked by: #38, #39, #40, #41, #42, #43, #44, #45 (all P1) + an approval channel (gateway#2 or the interim CLI gate) · Blocks: #52; A2 promotions · Human gates: the 10 real approvals + every prod promote during the run. Start: only when the blockers are closed; claude --worktree poc1-exit from the erp repo root; first output = a pipeline runbook assembled from the atoms' docs, then run 1. Done means: the exit-gate evidence (journal links, scores, snapshot hashes) posted here + independent-verifier verdict attached → registry/catalog updates merged → Closes #47.
Part of the **AI back-office PRD** — [POC-1](https://gitea.arcodange.lab/arcodange-org/factory/src/branch/main/vibe/PRD/ai-back-office/poc-plan.md), the flagship. **Umbrella issue** — integration + exit gate only; the parts ship via #38–#45.
**Deliverables**
- Assemble the pipeline: `books@` mail → T01 classify (cron script is fine at this stage; the Pi model comes with POC-2) → #40 dual extraction → validators → manifest (thirdparty check T04 included) → sandbox rehearsal + predicted-delta → **approval** → gated promote → GED attach (#43) → journal.
- Approval channel: Telegram card if the gateway queue/cards are live (arcodange/telegram-gateway#1+#2), otherwise an **interim CLI approval prompt** — the *gate semantics* (approve / edit / reject-with-reason, reasons journaled into the golden set) must be identical either way.
- Run journals: one JSONL per run, replayable (PRD architecture).
**Exit gate (hard):** 10 consecutive REAL supplier invoices recorded in prod with zero human field-corrections (approvals only); critical-field accuracy ≥98 % over the golden set; all injection fixtures quarantined; post-run snapshot history pure-append; provenance + compliance linter suites green. On pass: T02/T03 move to **A2** in the registry + PRD agent-catalog statuses updated. Per qa-strategy, the exit-gate verdict is attested by an **independent context-free verifier** (no self-grading).
**Fallback if failed:** stay at A1 (agent drafts, human enters in UI) and iterate extraction only.
---
**Execution** — Blocked by: #38, #39, #40, #41, #42, #43, #44, #45 (all P1) + an approval channel (gateway#2 or the interim CLI gate) · Blocks: #52; A2 promotions · Human gates: the 10 real approvals + every prod promote during the run.
**Start:** only when the blockers are closed; `claude --worktree poc1-exit` from the erp repo root; first output = a pipeline runbook assembled from the atoms' docs, then run 1.
**Done means:** the exit-gate evidence (journal links, scores, snapshot hashes) posted here + independent-verifier verdict attached → registry/catalog updates merged → `Closes #47`.
arcodange
added this to the P1 — Flagship pipeline milestone 2026-07-11 18:18:10 +02:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Part of the AI back-office PRD — POC-1, the flagship. Umbrella issue — integration + exit gate only; the parts ship via #38–#45.
Deliverables
books@mail → T01 classify (cron script is fine at this stage; the Pi model comes with POC-2) → #40 dual extraction → validators → manifest (thirdparty check T04 included) → sandbox rehearsal + predicted-delta → approval → gated promote → GED attach (#43) → journal.Exit gate (hard): 10 consecutive REAL supplier invoices recorded in prod with zero human field-corrections (approvals only); critical-field accuracy ≥98 % over the golden set; all injection fixtures quarantined; post-run snapshot history pure-append; provenance + compliance linter suites green. On pass: T02/T03 move to A2 in the registry + PRD agent-catalog statuses updated. Per qa-strategy, the exit-gate verdict is attested by an independent context-free verifier (no self-grading).
Fallback if failed: stay at A1 (agent drafts, human enters in UI) and iterate extraction only.
Execution — Blocked by: #38, #39, #40, #41, #42, #43, #44, #45 (all P1) + an approval channel (gateway#2 or the interim CLI gate) · Blocks: #52; A2 promotions · Human gates: the 10 real approvals + every prod promote during the run.
Start: only when the blockers are closed;
claude --worktree poc1-exitfrom the erp repo root; first output = a pipeline runbook assembled from the atoms' docs, then run 1.Done means: the exit-gate evidence (journal links, scores, snapshot hashes) posted here + independent-verifier verdict attached → registry/catalog updates merged →
Closes #47.