The desk has already done the looking before you sit down. This is how a working day actually runs — the automated backbone, the handoffs, and the moments that genuinely need a person.
Not an inbox. Not a folder of PDFs. What's waiting is specific rather than generic — each item there because a condition was met overnight, not because a report ran on a timetable.
A thesis on a regional bank holding is sitting in review — written overnight, 22:04 to 22:48, while nobody was here. You approved it yesterday; it ran on its own. Beside it, the book: seven BUY, eighteen WATCH, twenty-five active positions — the whole thing, not a sample.
Then the part that actually gets read first.
Underneath, what changed at the companies actually held or watched: thirteen real filings, not headlines — a director departure, a material agreement, a couple of routine regulatory filings, a leadership change. Categorised, dated, scoped to positions that exist. Not the whole market's news.
And a research pack dated this morning, on a company the mechanical screen originally rejected — it came back through the near-miss route and produced a complete four-document pack overnight.
That's roughly fifteen seconds of reading. In that time: what's being worked on, what's come into range, what changed at the companies held, and what finished overnight. Now comes the only part that needed a person — open the position that matters, read where the analysts disagreed and who gave ground, check the conditions written down before the price moved, and decide.
The desk is a chain of specific handoffs, not a black box — each one a named node passing a specific artifact to the next.
Every activity below has a stated trigger — a market close, a completed scan, a changed shortlist, a crossed price, or a person's decision. None run on a bare timetable.
Each exchange's prices update within minutes of its own close, staggered through the day and night — not one global batch run regardless of whether a market has actually closed.
Fires only when membership actually changed — a name entered or left the Core Watchlist, the top five moved, a near-miss appeared. A quiet night creates no work.
Before anything else happens, something checks that every job that should have run, ran — a completion record, not a health guess. A failure is known before anything else is.
A stalled-job sweep, a stuck-lock monitor, and a capacity watchdog run continuously underneath. Blocked work is marked with its exact unblock condition and picked up automatically — nothing sits silently stuck.
A thesis cannot start without a person's approval, and the approval itself is recorded — a timestamped trail of every decision to begin research, not just the conclusions it reached.
An investigation into one company. A management-quality report. A competitive-intelligence report. Or a workshop — one question put to independent functions at once, disagreements preserved, one document at the end.
Not policy for its own sake — each one is a governing rule paired with the dated event that caused it, or a live case of it firing.
Thesis cycles are queued and released one at a time rather than run concurrently — because concurrency once exhausted a shared capacity limit mid-thesis and lost the work. Research now queues rather than collides, and a cycle that starts, finishes.
Research doesn't start because someone felt like it. Every decision to begin coverage on a company carries an approval record and a timestamped log — an audit trail of intent, not just outcome.
A genuinely good business was once screened out on a single-year metric distorted by an accounting artefact. Rather than special-case that one name, it was workshopped across every independent function, spec'd, signed off five times — one sign-off catching a real error and forcing a revision — and shipped as an additive tier that changed no existing threshold. It carries its own guardrails: a hard ceiling on how many names it can carry, mandatory filing citation at intake, and an automatic drop if a name fails to normalise by the next cycle. That tier is still running: it's what produced this morning's research pack, and near-miss names now carry a higher burden of proof — adversarial review happens before valuation, not after.
A stalled-job sweep, a stuck-lock monitor, and a nightly freshness check that reports what it repaired — because work once sat blocked for fourteen hours before a person noticed. Now blocked work names its own unblock condition and resumes on its own.
A real "Board Attention" ticket, two flagged filings, both tagged with the same generic regulatory label. One turned out to be a routine buyback expansion — checked against the actual filing, closed inline, no action. The other had genuinely crossed its trigger — so rather than reasoning it through inline, a narrowly-scoped child task was opened for further work, with an explicit note that a full thesis cycle still needed separate approval. Hold-or-delegate, decided by checking the primary source first — never by the label a scanner attached.
A desk that makes you go looking for what matters has already failed you before you've made a single decision. So the system takes the looking away from the person entirely — the screening, the checking, the routine review — and leaves only the part that actually needs judgement: does this change what I believe, and what do I do about it.
That's not a convenience feature. It's the whole point of running research operationally instead of manually — the fifteen seconds of reading exists because a great deal of unseen work happened first, on a stated trigger, with a stated reason, and a record of what it found.