How Sera Works

How Sera Works

Sera is a background worker, not a chat bot. She does not wait for you to open a window and type. She runs continuously and quietly, checking your capture inbox on a fixed cadence and processing whatever has arrived.

---

The polling loop

Every 3 minutes, Sera runs a full cycle:

1. Poll the capture inbox for recent messages.
2. Classify each email to decide what kind of thing it is.
3. Record it as a Source Email in Notion, deduplicated by Message-ID so the same email is never processed twice.
4. Extract the substance using Claude.
5. Write candidate records into the relevant Notion databases.

Because every message becomes a Source Email keyed by its Message-ID, Sera can re-scan the same window repeatedly without creating duplicates. Dedup happens before any extraction runs.

---

What Sera processes

Sera recognizes several distinct kinds of input and handles each appropriately:

• Google Meet Recording — access is checked so the file is reachable, but no extraction runs on a raw recording. The meeting is marked partial.
• Meeting Transcript and AI Notes — the full text is exported and run through extraction. This is where the real memory comes from.
• Operational Email — a substantive email is extracted into messages, tasks, decisions, and other records as warranted.
• Forwarded Thread — the full history in a forwarded chain is read as context.
• Governance Agenda request — an email whose subject signals a "[Governance Agenda]" request is answered directly: Sera compiles a governance agenda from recent memory and emails it back.

---

The model chain

Extraction is resilient by design. Sera tries the best model first and degrades gracefully rather than failing:

1. Primary model — a capable Claude Sonnet model does the main extraction.
2. Haiku fallback — if the primary call fails, a faster fallback model attempts the same work.
3. JSON repair — if the returned structure is malformed, Sera makes a repair attempt before giving up.

The result is that transient hiccups rarely turn into lost content.

---

What runs while nobody is watching

The 3-minute loop is not the only thing happening. A handful of jobs run on their own schedule, and all of them stop at proposing:

• The contradiction sweep, nightly. Compares records against each other and files genuine conflicts as Open Unresolved Contradictions, with sources and the smallest reconciling decision. It cannot resolve one.
• Vital Signs health, nightly. Pure arithmetic against each metric's baseline, target, and direction. No model forms an opinion about a number.
• Correction learning, nightly. Classifies the edits and rejections reviewers made that day, so consistent corrections become applied preferences. Visible at /transparency.
• The Attention Brief, daily. A short, capped, already-ranked list of what deserves a person today.
• The Monthly Story, monthly. One narrative drafted into KB Drafts, sent to nobody until a human publishes it.

Purpose awareness

If your organization has configured a Governing Purpose Statement, Sera becomes purpose-aware. Every Decision Candidate she extracts is scored for Purpose Alignment, one of: Aligned, Neutral, Misaligned, or Unclear. This lets leadership see, at a glance, whether the decisions being made actually serve the organization's stated purpose, without anyone having to grade them by hand.

---

What Sera never does

Sera's boundaries are as important as her capabilities. By design, she:

• Never approves anything. Everything she creates is a draft, candidate, or pending item.
• Never applies canon. Proposed changes to governance canon are filed as pending requests only.
• Never sends unsolicited email. She replies only to explicit requests (like a governance agenda) and to necessary access notifications.
• Never deletes. She adds and proposes; she does not remove your records.

---

How fast is it

Most emails are picked up and fully processed within about five minutes of arriving, given the 3-minute poll cadence plus a short extraction time. You send content to the capture inbox, and shortly after, structured candidate records are waiting for review in Notion.

Key points

Sera runs as a background worker polling the capture inbox every 3 minutes: classify, create a Source Email record (dedup by Message-ID), run Claude extraction, write candidate records to Notion. She handles Google Meet recordings (access-check only), transcripts and AI notes (full extraction), operational emails, forwarded threads, and a Governance Agenda subject request. Model chain is primary Claude Sonnet, then a Haiku fallback, then a JSON repair attempt. If a Governing Purpose Statement is configured, every Decision Candidate gets a Purpose Alignment score. Sera never approves, never applies canon, never sends unsolicited email, and never deletes. Most emails are processed within about five minutes.