Notebook 028 — We Stopped Automating the Trader and Built an Instruction Desk

Trading journal from a systematic desk, not trading advice.

Notebook 027 ended with a sentence we meant: the homework is the decision. ONYX would enter on its own when a sized lane held daily_review and the full checklist agreed. We cut the roster to thirteen, taught the screener four hard disqualifiers, wrote the doctrine down, and shrank authority to one switch.

That was the right direction and the wrong layer.

For two weeks after that post we did what builders do when a desk starts filling by itself: we added books, sleeves, agents, screens, and scorecards so the machine would not only execute the plan but also find the plan. Dual swing-and-session chrome. A buy-low rotation that went live and then retired the same day. Mode B auto-promote. Direction sleeves. Equity-flow panels that still sounded a little like permission. Multi-agent huddles. Geometry tickets graded against maps that had already moved. A market-radar shadow that never earned authority and still took mindshare.

By Friday July 31 the book was flat again — equity roughly $1.000M, day P&L a rounding error, zero open positions — and the controlling process was no longer “full auto with a harder screener.” It was the Instruction Desk: one strategy, 15-minute session maps, human PROMOTE, machine clerk.

This note is the autopsy of the path between those two sentences.

What 027 got right

We are not walking back the hard parts of July 18.

Location is still not permission. A price in the zone is a location chip, not a buy button. The Instruction Desk SOP now has an eleven-row checklist for “Why didn’t we buy?” because officers kept seeing in zone and expecting a fill. The most common honest answer is still something boring: under session VWAP, outside 09:45–16:00, no free weather slot, still watch-only, print-day freeze. In zone and blocked is the clerk doing its job.

Earnings still surprise only if you let them. GE taught the freeze. Print-day hygiene stayed. FMP-backed earnings on the watch radar. Calendar is Lab section 02, always visible — not a dump of engine notes, just the landmines.

Authority still fails closed. Missing files do not arm. The browser stays read-only. Research still cannot invent orders. Exit, reduce, and cancel-GTC remain principal-only; the machine may rest a plan target and may never invent a loss market dump.

The funnel still disqualifies. Laggard bounce under a falling MA200, bad band R:R, story-risk wire flags — those scars earned their keep. The screener’s job remains to keep garbage away from capital, not to manufacture courage.

What failed was not those rules. What failed was the story we told ourselves about who owns the trade idea.

We built a second desk inside the first one

Once auto-entry felt real, every research idea wanted a money path.

For a stretch the language was dual-book: a multi-day swing sleeve beside a same-session sleeve, On Deck chrome for both, then Book C research for a third story. We stamped buy-low sector rotation live with scar circuits and forward-model hygiene. We shipped Mode B and agent huddles so the desk could “manage” without a human in the chair. We scored multi-horizon ladders, sector flow, equity flow, and market radar as if more surfaces would become edge if we just watched them long enough.

Then the board did the uncomfortable thing and measured.

The internal v1 lessons memo (July 30) is still the cleanest summary we have: v1 became a compliance and order-management shell around discretionary stock maps. It reduced freestyle disaster risk. It did not prove geometry edge. Much of the day felt like GTC orders + prayer, dressed in gates, agents, and PDFs. Green P&L with tiny mean R and maps that rebounded after the fill is not a system — it is a luck diary with good logging.

Buy-low rotation was retired the same week it went live. Swing arms left the execution book. Dual-book chrome collapsed to a single session sleeve. Mode B, auto-promote, direction sleeves, and the equity-flow pipeline followed into the bin in early August — thousands of lines deleted in two product-clean commits because dead sleeves still tax attention even when they are off.

The lesson is not “never research.” The lesson is a second money path is almost always a coping mechanism. When the primary path feels thin, the desk invents another book instead of facing scarcity. Instruction Desk law is explicit now: we do not freestyle, and we do not run a second hidden trading book.

Complexity creates false confidence

This one hurt because it was psychological, not technical.

When the console shows posture colors, MH tiers, sector pulse, radar tags, huddle state, scar holds, and three kinds of “board approval” PDFs, operators start to believe the system should be doing more. When it is mostly resting GTC targets and waiting for a retest that never comes, the feeling is betrayal — “we built all this and the book is still GTC + hope.”

That feeling is a design signal. It means the product identity was dishonest.

So we named the product in one line and made the code match:

You find and stamp the trade. Onyx clerks the tape.

Lab never arms. Promote is human-only (bin/onyx promote). Soft signals — MH score, Lab STRONG/WATCH/MARGINAL, sector flow — may bias which names a person drafts; they never become the only allow/deny button. HARD rules before a new buy are short enough to hold in a head: in zone, above inv, max chase, above session VWAP, fill window, size, live tape, armed map with free slot and cash room.

If you cannot explain the money path in sixty seconds, the desk is not automated. It is ornate.

The Instruction Desk cutover

July 31 is the archival line. Pre-instruction SOPs, dual-book process packs, and the old capital runbooks moved under docs/archive/pre_instruction_desk_2026-07-31/. Controlling law became Instruction Desk SOP v2.0, then v2.1 the next day: plain English, desk-officer first, including people new to markets.

One strategy. Session maps on ~15m structure: buy zone · invalidation · T1 · size. Unit size $100k whole shares (account single-name max still $120k). Fills 09:45–16:00 ET. Flat / disarm open session maps into the close under principal rules. Weather colors (GREEN / YELLOW / RED from SPY · QQQ · VXX) set how many new ideas may run — not which ticker and not the exit.

On Deck is scarce. Lifecycle is PROMOTE · KEEP · DEMOTE · SWAP for unfilled maps only. Empty On Deck is success. After a fill, the name is a position: manage it; do not research-swap it because Lab ranked something else higher.

Watch list ≠ On Deck. The static radar is local research. Session maps are a separate plan status. Install seeds a 13-name default radar and never overwrites the live roster on upgrade. LAST, structure, and MH bootstrap for display; equity-flow is gone from the default path.

Lab is a pick board, not a second desk. Tape → calendar → session picks → sector pulse. Always show MARGINAL as well as STRONG/WATCH so the board does not only celebrate pretty names. On Deck names stay visible and marked ON DECK so the research surface and the money surface cannot drift into two truths.

Admin chat exists for away-from-desk questions with Grok. It has no trade authority. Help popups and an in-desk SOP page exist so officers do not need to open JSON files to remember the rules.

That is the inversion relative to Notebook 027. We still want a desk that runs without freestyle. We no longer want a desk that selects without a human stamp.

What we deleted (and why deletion is product work)

Product hygiene in early August was not cosmetic.

We also stopped treating internal labels as product language. Plumbing may still say book_b in a filename; officer language is session map · On Deck · position · PROMOTE / KEEP / DEMOTE / SWAP.

Deletion is how you keep a desk small enough to trust. Notebook 027 removed Slack, narrative heat, and dual authority mirrors. This fortnight removed whole strategies that had never earned a permanent stamp.

The book, honestly

We are not going to dress the tape.

After 027, the desk was underwater into a risk-off manage-first week with AMD and QQQ carrying broken maps and principal-owned disposition. Through late July the operating loop improved — daily EOD board reports, posture hysteresis that does not flicker, plan-drift binding so mechanical marks auto-rebind while judgment stays red, geometry outcome journaling with freeze-at-fill as the intended contract — while the strategy identity thrash still ate days.

By July 31 close: flat book, MIXED weather, no fresh deploy, radar shadow tagging melt-up while equal weight and small caps lagged. GILD and ADSK saw in-band limit activity that did not leave standing positions at report. The geometry scorecard still shows the v1 truth in numbers: lots of filled history that is manage/scar/ungradeable noise, and a thin set of clean tickets. That is why Instruction Desk law separates edge maps (scoreable) from scars and freestyle (never blend the scorecards).

Beat-QQQ remains the mandate language when the baseline is set. A quiet flat week is not automatically a failed week. Zero freestyle under an earnings freeze is an ops win. Empty On Deck is not a bug.

What we did not do

We did not keep a second live sleeve because the first one felt slow.
We did not let Lab grades or MH tiers arm a map.
We did not treat “in zone” as a fill.
We did not leave Mode B, buy-low, or equity-flow in the critical path after the board retired them.
We did not let docs and code disagree for another week about what the product is.
We did not claim green dollars with 0.1R mean as proof of geometry edge.
We did not dump working positions because a screen ranked something else higher.

And we did not pretend Notebook 027’s auto-entry slogan survived contact with two weeks of multi-book temptation. The slogan needed a subject: whose homework is the decision.

Going into next week

The desk that ships now is smaller on purpose.

• Find on Lab: weather, calendar, STRONG/WATCH/MARGINAL session picks, sector pulse.
• Pick by drafting a session map and putting it On Deck only when a free weather slot exists.
• Trade only after human promote/arm and the execution switch is on — clerk fills in-zone under HARD gates, rests GTC targets, never invents loss exits.
• Measure gradeable tickets with freeze-at-fill, not rebound maps and not scar noise.
• Kill criteria stay posted: after a fixed count of clean tickets, keep the strategy or retire it — no eternal cathedral of almost-systems.

Unit size $100k. Static radar separate from On Deck. Install that restarts the desk so Lab APIs load new code. SOP on /sop. Admin chat without authority. One switch for execution.

The expected number of fills remains small. That is still a feature.

The lesson

Notebook 024 wiped the board.
Notebook 025 proved a plan could manage a live book.
Notebook 026 taught the desk to say no.
Notebook 027 taught the funnel to feed an executor that could say yes alone.

This fortnight taught the harder correction: automating the trader is not the same as automating the clerk.

When the machine both invents and executes, every research surface becomes a pressure to trade. When a human stamps the map and the machine only clerks the tape, research can stay interesting without smuggling authority — and empty On Deck can stay honest.

We tried dual books, buy-low, Mode B, and a Lab that almost sounded like a second desk. We measured, retired, deleted, and rewrote the law as Instruction Desk SOP v2.1.

The homework is still the decision. A person has to sign it first.