What Happens When Sera Reads a File

What Happens When Sera Reads a File

Sera is a background worker, not a chat assistant. She reads every email sent to your capture address and turns it into structured institutional memory. Here is exactly what happens, in order, each time a file or email reaches her.

---

Step 1 — A Source Email record is always created first

The very first thing Sera does with any incoming email is write a Source Email record in Notion. This happens before classification, before extraction, before anything else. The Source Email record is the permanent audit anchor.

This ordering is deliberate. If extraction later fails for any reason, the Source Email record still exists and shows the failure reason in its error log. Nothing an email ever contained is silently lost. If you cannot find something Sera should have captured, the Source Email log is always the first place to look.

Step 2 — Classification

Sera reads the email and decides what kind of message it is: a meeting recording notification, a transcript, an AI notes export, an operational email, a forwarded thread, or an agenda request. Classification determines which processing branch runs next.

Step 3 — Attachment processing

Any attachments or linked assets are identified and prepared. For meeting assets this includes checking access to the underlying file so the content can be exported to text for extraction.

Step 4 — Language detection and translation

Before extraction, Sera detects the language of the content. If the incoming language differs from the output language configured for your organization, the full content is translated first, and only the translated text is passed to extraction.

The Source Email record shows both the detected language and whether translation was applied. This means everyone can write and meet in whichever language they like. Sera reads all of it, but the records she creates normalize to the single configured output language so your memory stays consistent and searchable.

Step 5 — Claude extraction to Draft, Candidate, or Pending

Sera then uses Claude to extract structured records: tasks, decisions, risks, profiles, memory candidates, and more. Every record she creates is a Draft, Candidate, or Pending item. Sera never creates a Published, Approved, or Active record. She proposes; humans dispose. This is the core safety property of the whole system.

Step 6 — Routing to the right review queue

Extracted records are routed to where a human will see them:

• Governance and institutional-knowledge candidates go to the Memory Review Queue.
• Anything flagged sensitive goes to Sensitive Review, which is admin-only.

Ordinary operational records land in their normal databases as candidates for confirmation. Depending on the review policy the organization has set, some low-stakes candidates can promote themselves after an objection window rather than waiting on a click. The trust dial article covers exactly which ones, and which can never do that.

---

When something fails

Extraction and access are not always successful on the first try. A Drive file might not be shared yet, or a transient error might interrupt a step. Sera handles this with an automatic retry schedule:

1. Immediate first attempt
2. Retry after 30 minutes
3. Retry after 2 hours
4. Retry after 24 hours

If all four attempts fail, the item moves to Manual Review so a person can step in. Most failures are transient and resolve themselves well before the fourth attempt, so a single failure log entry is rarely cause for concern.

---

Why this design matters

Because the Source Email record is written first and survives everything downstream, you always have proof that an email arrived and a record of what happened to it. Because translation happens before extraction, multilingual teams get one clean, normalized memory. And because Sera only ever writes candidates, no record enters your canon without a human choosing to put it there.

Key points

Every email produces a Source Email record first, before any processing, so the audit trail survives even a total extraction failure and records the reason. Sera then classifies the email, processes attachments, detects the language, and translates the full content ahead of extraction whenever the incoming language differs from your configured output language. Claude extraction only ever produces Draft, Candidate, or Pending records, never Published, Approved, or Active ones, and each result is routed to the correct queue: governance items to the Memory Review Queue, sensitive items to Sensitive Review. Failures retry four times on an exponential backoff of immediate, 30 minutes, 2 hours, and 24 hours, and then move to Manual Review for a human.