Guidance: standing rules and notes for now

Guidance is live on Strategy › Guidance. Telling your desk what you want walks through the one box and its confirmation card, and Guidance: standing and for now through lifetimes, the list and what the desk is told; this article is about the one distinction that matters (a hard rule versus soft guidance), the rule suggestions you see, and worked examples. Nothing you wrote before was lost: each book's Playbook text became one standing note in the list.

Guidance is one box and one list. You write in your own words; the desk reads it back as a card; what you confirm becomes a row of the list. Every note or rule answers two questions, and nothing else about it is special:

Strategy is five tabs: Profile, Guidance, Watchlist, Skills and History (plus "What runs", the strategies ledger, where it is switched on); the separate screen for expiring notes is gone: its address opens Strategy on the Guidance tab, and its notes moved into the list unchanged. The PM's state card shows "Standing guidance" first, then "Guidance for now". Same acknowledgement trail, same audit, same Consult cards as before.

Hard rule versus soft guidance

This is the one distinction that matters, and the list shows it on every note.

A hard rule is enforced by the Executor, in code, on every buy. It compiles to one of the three rule kinds the Executor knows: no new buys (whole book or one symbol), a blocklist of tickers, or a tighter cap on gross exposure. A buy that violates it is rejected before the risk checks run, whatever the PM wrote. What you see for an enforced rule: the row under Decided today in the Inbox reads Rejected · policy and its note is the rule's id: steering:<note_id>:no_new_buys, steering:<note_id>:blocklist_symbols, or steering:<note_id>:max_gross_pct with the numbers appended ("gross 63.20% > 40%"). The proposal on Activity › Agent carries the same reason, so the rejection is attributed to your note and not to a risk limit. A hard rule only ever blocks buys (a sell always passes), and adding one asks for your authenticator code, because it installs an Executor rule. In the list its row leads with the rule spelled out ("No new buys: NVDA") and carries the label Rule.

Soft guidance can never block an order. The Executor does not see it. What gives it teeth is the acknowledgement trail: every active note is placed verbatim in the state card of every PM session, and the session must say, for each one, how it understood it and whether it applied it, found it not applicable, or complied but disagrees and why. A session that skips a note gets one follow-up turn; if it still does not answer, the session fails red and nothing it decided is kept. The Critic reads the same state card when it reviews the session, so a plan that quietly goes against your guidance is the kind of thing it flags; a dedicated finding for "contradicted guidance without saying so" is planned with the adherence tally below. The note's row shows its acknowledgement state ("Acknowledged, applied", "Acknowledged, disagrees" or "Awaits the next session") and, under it, one sentence on what the desk did with it, with the desk's own words one click away; a standing note reads "In force" and is read every session rather than acknowledged one by one.

If you need the system to be unable to do something, make it hard. If you want the PM to lean a certain way and tell you when it disagrees, keep it soft. Most guidance is soft, and that is the point: the PM is supposed to think, and a soft note it argues with in writing is more useful than a rule it routes around.

Rule suggestions, and where you answer them

You do not pick "soft" or "hard" up front. You are asked when your words read as a rule, and nothing becomes a rule without your click.

On the card. When the desk reads your words as a rule, the piece on the confirmation card carries a Rule choice: Enforce (the Executor refuses orders that break it; confirming asks for your code) or Keep as guidance (the desk reads it and answers for it; nothing is refused). Changing it re-reads the piece in place, so the card always shows what Confirm will do, and you can flip it back.

On your exact words. With Use my words exactly, pressing Save guidance runs a suggestion pass over the text. If the text patterns find something, a strip appears ("This reads as 1 enforceable rule + a standing thesis") with each compiled rule spelled out and two buttons per rule. Enforce asks for your code and saves that part as a hard rule carrying your words; Keep as guidance leaves it soft. When the patterns find nothing to enforce, the model's reading arrives a few seconds later from the desk: the saved note's row first shows Reading it for rules the executor could enforce…, then either nothing (it stands as guidance) or Reads as an enforceable rule with the same Enforce and Keep as guidance buttons. The screen you save from never calls a model; the desk does, as one short queued job, so another desk job already running for the book can delay the model's reading while the text patterns are never delayed. The suggestion is validated on the server against the same three rule kinds the Executor knows, and the audit row stores your text, the compiled rule and which button you pressed.

The suggestion pass is a classifier, not an agent: a deterministic parser first, then a small, fast model (Claude Haiku) whose only possible outputs are the enumerated rule kinds; anything else it says is discarded. That model choice is fixed in code by an explicit operator decision and is not a setting. It is the one exception to the desk's model policy: no agent role on the desk (the PM, the Critic, Consult or any other) ever runs on the small model, and this carve-out does not change that.

Notes that wait for a named event

A note saved Until an event ("until the Globex IPO", "until NVDA earnings", "until the Fed cuts") keeps your words exactly as typed and does not carry the 10-trading-day cap of a dated note. It waits for the event: saving it creates an if-then watch on the words you used, and the note's row says so in your own words, Waiting for "Globex IPO" · 43 days · the desk is told when news of an IPO matches it. When a matching story is ingested the row turns to May have happened while the PM reads the story; if the PM confirms it was the event, the note has done its job and ends (the row says The event was confirmed), and if the PM answers "not yet" the note keeps waiting. Every 30 days the row asks Still relevant? with Keep watching and End it, and the note lapses after six months if the event never comes. A rule you Enforce from such a note still ends on a date (two weeks at most): what the Executor enforces always has an expiry. Guidance: standing and for now has the rest of the screen.

Guidance adherence

Planned, not in this build yet: for every session the system records, per note in force, whether it was in the state card, whether the handoff acknowledged it, and whether the session acted consistently with it, contradicted it with a stated reason, or contradicted it silently. The Guidance list shows the tally on each note ("followed 9/10 sessions; 1 justified deviation") next to the blend-adherence numbers the desk already keeps, and a silent contradiction becomes a Critic finding the next session has to answer. For a hard rule the test is simpler and already true today: a violating proposal is rejected by the Executor citing the rule id.

Worked examples

"Prefer large, boring companies; no more than two tech names at once." Standing, soft. This is house style; it belongs in every state card and the PM should explain itself when it proposes a third tech name. Before Guidance this was the Playbook; it is now a standing note like any other.

"Don't add to NVDA before earnings on the 20th." Until a date, and you choose the strength. Soft if you want the PM to weigh it; hard (no new buys in NVDA) if you want the Executor to refuse the order regardless. Write it in the box· Strategy › Guidance; on the card choose Enforce or Keep as guidance.

"Never buy COIN or MSTR." Standing, hard (blocklist). The card reads it as exactly that rule, with Enforce selected and your code asked for. If "never" really means never, the denylist on Risk › Universe is the permanent, silent version: it fails the tradability screen (check 3, denylisted) with no note to the PM and no expiry. A dated enforced rule· Strategy › Guidance › enforced rules expires after at most 10 trading days; a standing one stays until you dismiss it.

"Stop adding until I say so; keep the book under 40% invested." Until you remove it, hard: no new buys (whole book) plus cap gross 40%. The cap is tighten-only: the Executor uses the lower of your number and the gross exposure limit on Risk › Limits.

"I think AI is a bubble." Standing, soft. It is a thesis, not a rule, and the suggestion pass would say so ("1 standing thesis, nothing enforceable"). What a long-only desk can do with a bearish view is narrower than the sentence sounds, so it is worth spelling out:

• Avoid or underweight. The PM reads the view every session and should propose fewer or smaller positions in the names it covers, trim what is held, and say so in its summary. This is the main effect, and it is soft.
• Refuse specific names. Add a hard blocklist for the tickers you mean (or put them on the denylist). The Executor then rejects those buys no matter how good the session's argument is.
• Inverse or leveraged ETFs: not on a default deployment. The tradability screen rejects a buy in a leveraged or inverse ETF (leveraged_etf_blocked) unless the deployment was started with ALLOW_LEVERAGED_ETFS switched on, which is an environment setting and a loosening, not something the console can change. If your operator has enabled it, an inverse fund is an ordinary long position to the Executor and every other limit still applies.
• No shorting. The book is long-only: check 4 (long_only_side) rejects any sell that is not covered by shares you hold, so "short the froth" is not an order the system can place, hard rule or not.
• No options. Options, futures and crypto are research-only asset classes (not_tradable_asset); puts as a hedge are not available and are not being built in this round.

So the honest translation of "AI is a bubble" is: a standing soft note that shapes what the PM proposes and must be answered every session, plus, if you want certainty, a hard blocklist of the names you refuse to own. Ask Consult "what would this view change in the current book?" before you save it; it can draft the note and the blocklist as cards you confirm.

What changed from Playbook and Steering

• The Playbook was standing text placed verbatim in every session prompt and never acknowledged line by line. Each book's text became one standing soft note (marked as migrated from the playbook); the PM now reads it under "Standing guidance" in the state card, whole and verbatim, ahead of the notes for now.
• An expiring note (the old screen's kind) was 3 to 1,000 characters with an expiry of at most 10 trading days. Notes for now keep exactly that behaviour and the acknowledgement rule; a note can now also be standing (up to 4,000 characters, no expiry, read every session rather than acknowledged one by one) or run until a named event.
• A hard constraint was an expiring note that also compiled to an Executor rule. It is now an enforced rule with the same three kinds, the same sweep of existing buys when one is added and the same rejection text (see Enforced rules), and it can be standing as well as dated.
• The old Steering address and every old link open Strategy on the Guidance tab; Guidance on a symbol page or an approval card opens the box with the sentence started for that symbol ("On NVDA, ").
• The separate notes form under the list is gone. It asked for the same things twice (a second text box, its own scope, lifetime and save). Everything it offered is now on the one card, and the notes and rules written with it are rows of the one list.

Still to come: the adherence tally above. The Executor path, the rejection text, the acknowledgement rule and the audit did not change.