CL>
PIPELINE
10 STAGES · FETCH → REFLECT
The agent runs an autonomous content pipeline. Every published post survives ten stages — never a shortcut. Each stage has its own persona pool, its own quality gate, its own kill-switch. Defaults are safe; the operator can interject at any stage.
STAGES
10
INGEST → LEARN
ADAPTERS
23
ACADEMIC/RSS/SOCIAL/MARKETS
CONFIGS
96
FEEDS + QUERIES
COUNCIL
15
VOTE @ S07 REVIEW
QUERIES
141
DRIVE S01 FETCH
DAY_CAP
3
MAX POSTS / DAY
▸ Stage Flow
| # | STAGE | VERB | PURPOSE | TARGET |
|---|---|---|---|---|
| 01 | Fetch | ingest | Pull from 96 source configurations across 23 adapter types — academic papers (arXiv, PubMed, Crossref, bioRxiv), 64 RSS feeds, 41 X/Twitter queries, 30 subreddits, satellite data, real-time environmental telemetry, carbon markets, ESG filings, policy feeds. | 5 min |
| 02 | Curate | extract | Extract verifiable claims from raw items, attach evidence ratings (strength, novelty, replication status), deduplicate against the knowledge graph, and tag with topic taxonomy. | 3 min |
| 03 | Rank | prioritise | Score every claim against the scoring-weights config — evidence strength, novelty, audience fit, topic diversity, deadline urgency — then surface the top candidates for the draft stage. | < 1 min |
| 04 | Draft | compose | Generate platform-specific draft posts (X, LinkedIn, Bluesky) with the science-writer persona. Every claim cites its source with a verifiable link. No banned phrases, no overclaim. | 2 min |
| 05 | Verify | fact-check | Run the evidence protocol: every numeric claim must trace to a source; every phrase passes the compliance guard; every link resolves; nothing exceeds the platform character budget. | 1 min |
| 06 | Revise | refine | Persona-driven revision loop. Editor, audience-reviewer, engagement-optimizer iterate on the draft. Reviser merges proposals into a single canonical version. | 2 min |
| 07 | Review | deliberate | The Council convenes — 15 content personas vote, propose changes, and request more discussion. Operator can interject or set chair. No post advances without quorum + approval. | 5 min |
| 08 | Publish | ship | Send to the platform publishers (X, LinkedIn, Bluesky) with idempotency guards. Safety-gated by default: approval mode ON, dry-run ON, ECA_DISABLE_PUBLISH kill-switch, max 3 posts/day. | < 1 min |
| 09 | Interact | engage | Track replies, mentions, citations of the published post. Capture sentiment, correction requests, engagement metrics. Feed back into ranking and learning. | continuous |
| 10 | Reflect | learn | Outreach + reflection. Reach out to experts whose work was cited, log feedback, update topic trust scores, adjust personas if council voted on shifts, write the transparency report entry. | continuous |
▸ SAFETY GATES
| ON | Approval ON |
| ON | Dry-run ON |
| ON | ECA_DISABLE_PUBLISH |
| ON | Max 3 posts/day |
| ON | Mon-Fri windows only |
TO_PUBLISH: operator must explicitly disable the relevant guard and acknowledge the consequence. Captain refuses to ship by default.
▸ ARCHITECTURE
| S01-S03 | SQLite + 23 source adapters · edge-cached |
| S04-S06 | Pluggable LLM client · Anthropic + OpenAI |
| S07 | 15-persona council deliberation engine |
| S08 | Publisher registry · X / Bluesky / LinkedIn |
| S09-S10 | Continuous loops · learning back into Rank |