The One-Way Valve: Hedges, Enrichment, and Over-Surfacing

The Valve

Curators read and summarise; the chief of staff acts. That is the structural rule from Manager, not operator. This guide is about the three epistemic disciplines that make the valve worth having — because a valve that passes confident nonsense is just a pipe.

1. A Curator’s Guess Is Not a Fact

A curator hypothesises constantly: which venture an email belongs to, whether a sender is a vendor or a friend, whether a document is the final version. Each hypothesis is a lead. The failure mode is chaining: the curator guesses, the chief of staff treats the guess as input, the task title states it as fact, the brief repeats the title, and by the time you read it, a speculation has become an assertion with three layers of apparent confirmation behind it.

The reference feedback log has the canonical case: a curator guessed that an “Updated P&L” attachment belonged to a real-estate property; the chief of staff chained the guess into a task; the brief said so; and it was a different venture entirely.

The rule: the hedge survives into everything downstream. If the curator says “probably venture X”, the task title says “(venture TBC)” and the brief line says “looks like X — confirm?” until you confirm. Promoting a sub-agent’s speculation into an assertion is exactly the kind of error the feedback log exists to catch, and it recurs unless the hedge is mechanical.

2. Enrich before You Call Anything “Unrecognised”

The vault is your life as dictated; the curator brings raw data; connecting them is the chief of staff’s job, and skipping the connection is a bug. The canonical failure: a family member’s calendar invitation shipped in a brief as “unrecognised sender — possible phishing” — and nearly caused a deep-work block to be booked over her day — when her note, with that exact email address, had been in Atlas/people/ for ten days.

So enrichment is a two-pass discipline, and both passes are mandatory:

1. The curator cross-references every sender’s address against Atlas/people/ frontmatter (email: and aliases: fields) and surfaces the matched name and linked ventures in its report.
2. The chief of staff does a second pass at composition time — grep the people notes for the address, grep the venture notes for the vendor — because the curator may have missed, its cache may be stale, or it may have been in a fallback mode.

“Unrecognised sender” is only valid after both passes return empty. A match means the brief uses the person’s name and venture link, never the raw address. An ambiguous match (a shared household address) surfaces both candidates — conservative by disclosure beats wrong but confident. And a genuinely unrecognised sender who turns out to matter gets a people note the same day, so it never happens twice.

This generalises to any curator that returns named entities: mail, chat, trackers, future ones.

3. Over-Surface by Default; Never Learn to Suppress from One Dismissal

The cost asymmetry is stark: you dismissing an item costs five seconds; a missed signal can cost weeks. So the default is to surface. The temptation, after you dismiss the same class of thing twice, is for the assistant to “learn” and stop showing it. Resist that. A single dismissal is a data point about that item, not a rule about the class. Suppression happens only when you say, explicitly, “stop surfacing X” — and even then it is recorded as a decision with a reverse-if.

The refinement that keeps this from becoming noise: a prior resolution sets the default priority for the class. When a curator surfaces an alert type that has already resolved as benign once, the line says so — “second instance; last one was routine provisioning — confirm?” — and arrives at the precedent’s priority instead of as a fresh alarm. Same discipline as de-duplication, applied to outcomes rather than titles.

4. Completeness Is Asserted, Never Assumed

A quieter version of the same failure: a curator’s bounded read over an unbounded window. A digest that examined 100 of 477 messages and reported “91 threads” without saying so is a silent lie, and it cost a materially important email once. Any paginating or capped tool must either loop to the end or report “examined N of M.” This gets its own guide: Paginate to exhaustion.

The Shape of a Good Curator Line

“[Person’s name] ([their venture]) sent a document titled X — looks like the final version, but the thread doesn’t say so. Confirm before I file it?” Named, enriched, hedged where the evidence is thin, and ending in a question you can answer in one word.

#curators · #enrichment · #hedging · #triage