What Sera Learns From Correction, and How to Make Her Forget

Sera gets things wrong. What matters is what happens next. Every time a reviewer fixes a draft or rejects one with a reason, that correction is captured, and once the same correction has been made consistently, Sera stops making that mistake for this organization.

The correction is the edit you already make

There is nothing extra to do. The loop runs on the review work that is already happening:

• You edit before approving. Sera keeps the version she wrote. If the values you approve differ from the values she drafted, the difference is the correction. Approving something unchanged is silence, not a signal.
• You reject with a reason. A rejection reason is the single richest piece of feedback in the system, because it says what was wrong rather than only that something was.

Capture is deterministic and immediate. No model call happens at the moment you click, so correcting a record never slows the interface down.

One correction is a note. Three is a preference.

A single edit could be a one-off. So a correction starts life as a candidate, becomes emerging on a second consistent observation, and becomes established on a third. Only established preferences, and ones a person has explicitly confirmed, are ever applied to how Sera writes.

If a later correction contradicts an established preference, it is marked contested and stops being applied until a person resolves it. Preferences that stop being reinforced eventually age out. Nothing is deleted; the row and its history stay.

What Sera can and cannot learn

This is the boundary that matters most, and it is structural rather than a promise. Sera can only ever learn output preferences:

• Formatting, such as how a summary should be shaped.
• Level of detail, such as shorter task descriptions.
• Ordering, such as which section comes first.
• Terminology, such as which word this organization uses for a thing.

She cannot learn anything about a person. The learning system has no category for personality, psychology, health, character, or protected characteristics, so an observation of that kind cannot be stored even if a reviewer's note contained one. It is discarded as not generalizable.

The transparency panel

Everything Sera has learned about how this organization wants things written is visible at /transparency, reachable from the dashboard as "How Sera Works With Me". For each preference you see:

• What she learned, in plain language.
• What kind of preference it is.
• Its status, applied or still learning.
• Its evidence, how many times it was observed and since when.

Any signed-in person can see the whole list. That visibility is the point: a system that quietly adapts to you without telling you what it adapted is not transparent, however well it behaves.

Forgetting

An admin can click Forget on any learned preference or name alias. It stops being applied immediately. The history of the row stays, because honesty about what was once learned is worth more than a clean-looking list.

Use it whenever a preference has outlived its context: a reviewer's personal style that the circle does not share, a terminology choice that changed, or anything you simply do not want shaping her output.

Why this is the interesting half of the product

An assistant that makes the same mistake forever is a tool you work around. An assistant that learns silently is one you cannot audit. This design takes the third path: correct her in the ordinary course of review, watch what she learned on one page, and remove anything you disagree with in one click. The learning is real, and it stays governed.

Key points

Reviewer edits and rejections are captured deterministically at the moment they happen: an edit made before clicking Approve is the correction, and a rejection reason is the richest signal of all. Corrections are classified nightly and move through a lifecycle of candidate, emerging, established; only established preferences (three consistent observations) or ones a person has confirmed explicitly are ever applied to a prompt. Sera can only ever learn output preferences: formatting, level of detail, ordering, and terminology. The learning system has no category for a person's personality, psychology, health, or character, by design. Open /transparency from the dashboard to see every learned preference with its evidence trail and count; any signed-in user can see the full list, and an admin can forget any item with one click, which stops it applying while the history stays for honesty. A preference that later gets contradicted is marked contested and stops applying until a person resolves it, and unused preferences age out on their own.