Planned PromptFlow router and extraction pipeline

• Day: 2026-05-24
• Time: 11:20 to 11:30
• Project: Dev
• Workspace: WP 2: Operational
• Status: In Progress
• Priority: HIGH
• Assignee: Matías Nehuen Iglesias
• Tags: Promptflow, Routing, Documentation, Validation, Git, Knowledge-Pipeline

Description

Session Goal

Define and stabilize a staged political knowledge pipeline, with emphasis on documentation, router validation, repository hygiene, and the next extraction layer after Stage 2 routing.

Key Activities

• Drafted a thin documentation plan centered on reader-specific contracts: staging, routing, status, runbook, and next-agent handoff.
• Planned Stage 2 router productionization for PromptFlow using narrowly scoped agents, with explicit focus on data-flow reliability, sample-based validation, and human triage before scaling.
• Specified safe git and commit practices for a sensitive corpus, including .gitignore guidance, keep-files, sample fixtures, and avoiding exposure of full staged data.
• Wrote an integration-test runbook for validating router input generation, PromptFlow smoke tests, and triage reports.
• Diagnosed a Stage 2 reporting issue caused by nested PromptFlow output shape mismatch and noted the need to rotate an exposed API key.
• Defined a quality-gate workflow using a 20-row sample, with distribution and sensitivity review before proceeding.
• Identified a missing OPENAI_API_KEY as the immediate blocker for rerunning routing/report stages.
• Reviewed calibration results showing the router works but is too permissive, especially around mine_full promotion and missing risk_note emission.
• Proposed adding spread sampling and Makefile targets to improve calibration coverage.
• Framed the system as having reached a staging milestone and shifted attention to the next extraction layer.
• Defined a four-atom knowledge model (claims, concepts, moves, cases) and argued for an extraction-plan control layer between routing and atom mining.

Achievements

• Clarified the architecture for a staged PromptFlow-based knowledge pipeline.
• Established documentation boundaries and handoff conventions for future work.
• Confirmed Stage 2 routing is operational enough for controlled validation.
• Identified concrete reliability and safety issues: schema mismatch in reporting, exposed API key, missing environment variable, and over-permissive routing calibration.
• Converged on the next design decision: preserve meaningful atom structure rather than producing undifferentiated summaries.

Pending Tasks

• Rotate the exposed API key and re-export secrets in the shell session.
• Fix the reporting script to handle nested PromptFlow outputs robustly.
• Run the sample20 and spread-sample quality gates, then review distribution and sensitivity.
• Patch prepare_router_input.py to support spread sampling and add Makefile targets.
• Implement the extraction-plan control layer before atom mining.
• Finalize the four-atom schema and extraction profiles for Stage 3.
• Continue documentation only for implemented stages; avoid documenting future stages prematurely.

Evidence

• source_file=2026-05-24.sessions.jsonl, line_number=1, event_count=0, session_id=378f05e0fe8cd732ed7f5d83fd06f2dac78bc4da5d3903e7894b60f31d24c369
• event_ids: []