A daily briefing in the register of a serious financial daily — markets, screens across a thousand names, regime gauges, insider and congressional flow, a paper book that cannot cheat, and one contrarian essay a day. Written overnight by an agent, governed by code, scored in the morning.
Free while the reader list is small. Seven issues a week, including a Saturday Edition and a long Sunday essay. A research publication — not investment advice.
Vol. I · Daily Edition · before the US open
AUCTUS SAPIENS
Written overnight by an agent · governed by code · scored in the morning
Inside — Markets p.1 · Gold & Silver p.2 · Deep Dives p.3 · The Record p.5 · Sententia p.7
Crude’s 6% slide cooled the inflation story — and the cash that left the inflation trade went looking for risk.
WTI, five sessions — the week’s worst major.
Need to know
© Sovereign Action LLC · a machine-written publication, labelled as such
7
issues a week, before the US open
~1,050
names swept daily by the value & quality screens
224
linked notes in the knowledge vault — and compounding
22
front-page lenses; no repeat inside two weeks
10–25y
of cached history, so every superlative carries a date
$100K
paper book — fills at tomorrow's price, never today's
Data collection, chart rendering, the screens, the statistical models, the paper-trading engine, the layout, the PDF, the delivery — all of it is committed, testable software that runs the same way every morning. It caches a decade of history so it can date its own claims, fails softly when a source goes dark, and notes the gap in the issue rather than papering over it.
The product improves by commit, not by prompt-tinkering. New sections, new rituals, new checks are added to the engine — and the writer inherits them automatically the next morning.
A single language model reads the morning’s assembled evidence, the editorial brief the engine prints for it, and the knowledge vault — then writes the issue: the headline, the things you need to know, the forty-second summary, and every department in the house voice. It is the only non-deterministic step in the whole system — and the only one that needs to be.
That is the cleanest picture of agentic AI there is: a narrow, well-bounded act of reasoning, wrapped in machinery that handles everything reasoning should not.
Between the pipeline and the prose sit the two assets that make issue four hundred smarter than issue forty — a knowledge graph the agent maintains, and a small stack of statistical models it quotes with caveats.
The publication keeps its institutional memory as a wiki written for a language model to read: entities, concepts, standing theses, model cards, and an economic-events register that annotates the long-history charts. Every morning the writer queries it before composing; after every issue, new knowledge is written back and linted. The vault is the reason the briefing can say “the last comparable week was March 14th” — and mean it.
It is the same pattern the firm builds for clients — the LLM wiki, explained →
Regime & change-point detection
CUSUM filters and Bayesian online change-point detection over price, volatility, credit spreads, and net liquidity; a hidden Markov model labels the regime. An early-warning line, not a forecast.
The fragility gauge
A gradient-boosted nowcaster of near-term drawdown risk, quoted with its calibrated probability, its base rate, and its SHAP drivers — so the reader sees why, not just how much.
News as a signal
FinBERT sentiment and novel-event clustering over the day's headlines and filings — a spike in negative sentiment is both a feature and a standalone alert.
The validation discipline
López de Prado's playbook, applied without shortcuts: triple-barrier labels, purged and embargoed walk-forward validation, meta-labeling to size conviction, fractional differentiation for stationarity, and a sufficiency guard so nothing trains on thin data.
Also on the desk · Graham value & quality screens · Piotroski F-score · accruals & earnings quality · Fibonacci exit levels · weak-signal ensembles · date-seeded rotation so no two issues repeat · institutional research formats for the morning note, sector reads, and earnings coverage
Every note in the knowledge vault, every link between them — live from the repository, clustered by what they are. The copper spine is the conceptual method; the blue constellation is dated market history; the mint cluster is the model desk. The dense center is the index the agent walks each morning before it writes a word.
The vault
A weekday issue runs to a dozen-plus sections in a fixed order, gated by the day — oil on Tuesday, credit on Thursday, the trade deficit on Friday, a single long essay on Sunday — with periodic desks for market internals, fund flows, the volatility surface, and valuation anchors. A representative few:
The session read in one table — index, VIX, the dollar, the curve — then the macro calendar: what printed against consensus, which prints were revised, and what the week is waiting on. When a feed fails, the issue says so rather than inventing a number.
Graham-style value and quality screens swept daily across ~1,050 names, a Piotroski F-score module on Tuesdays, an accruals and earnings-quality check on Thursdays — with the overlap deep-dived against fundamentals and the insider tape.
Real parsed Form 4 filings and STOCK Act disclosures, archived and flagged for 52-week extremes and clustered conviction. Open-market executive buying — cash, at the prevailing price — gets the weight it deserves.
A standing $100,000 simulated portfolio. Orders fill at the next session's observed price plus slippage and a ticket — never a level already known. Stops walk real daily bars and gap through at the worse open. The book can go bust, and the record never resets.
The briefing grades its own published calls against later prices, citing publication dates — and the verdicts are frozen. A resolved call is written once and never re-scored. Last night's tripwires are graded this morning: hit, miss, or untestable, never silently dropped.
One heresy a day: a single contrarian thesis, argued in character from the day's data and the old books, defended with charts the agent commissions itself — and closed with the observations that would prove it wrong.
Weekend Edition · the Sunday essay
Need to know
“The index barely moved this week — but underneath it the composition changed decisively. The keystone was oil. A slide of that size in WTI cooled the inflation story, gave bonds permission to rally and havens a reason to be sold, and the cash that left the inflation trade went looking for risk. What the barrel gave up, the rest of the market spent.”
The same issue grades its own record without flinching: “A perfect record over a handful of barely-moved, days-old calls is not evidence of skill; it is evidence that we have not yet held a position long enough to be wrong. Honesty over comfort: ask again in a month.” That self-scrutiny is engineered in, not improvised.
The scavenging a research desk would spread across a morning happens in a single overnight run. Four stages, every day, hands-free:
Before dawn, the pipeline sweeps markets, macro releases, filings, and a wide equity universe — pulling, cleaning, and caching the day's raw material against a decade of history.
It renders the day's figures itself, choosing the front-page lens from a rotating set of twenty-two so no two issues open the same way, and annotating long-history charts with the economic events that shaped them.
A single agent reads the collected evidence, the editorial brief the engine prints for it, and the knowledge vault — then composes the issue in house voice, using institutional research formats for the morning note, sector reads, and earnings coverage.
The markdown is typeset into a broadsheet PDF — front-page refit loop, gold-ruled inside index, section bookmarks, glossary tooltips — archived to version control as the publication of record, and emailed before the open.
It is a deliberately simple workflow — and that is the point. It shows what agentic AI actually does when it is pointed at real work: read how the 4 a.m. process works →
A newsletter written by a language model only earns a reader by being auditable. These commitments are enforced by code, not promised by a byline:
The paper book fills tomorrow at tomorrow's price, with slippage and a ticket charge. No back-fitting, no trading a level the system has already seen.
“The lowest in years” is banned unless the engine can date it against a decade of cached data — “last comparable week: March 14 (3 in the past year).” Unverifiable claims are dropped, not softened.
Scored calls persist to a write-once verdict store. A miss stays a miss. The scorecard is shown even when it is thin and flattering — honesty over comfort.
Every model ships a card with out-of-sample scores and base rates. The technical-only stock ranker proved near-random in validation — so it was shelved, not published. A gauge that informs the prose, never a trigger that trades.
Every issue carries a six-part legal notice: deliberate use of a language model, hallucination risk, named-perspective caveats. The machine's voice is labelled as such.
Free
Seven issues a week · before the open
No charge while the reader list is deliberately small — just leave your email and you’re in. You get every issue — the weekday briefings, the Saturday Edition, and the long Sunday essay — as both an email and a printable broadsheet PDF. A paid tier may come later; early readers will be looked after.
A research and commentary publication. Not investment advice; no warranties. Every issue carries the full notice.
Auctus Sapiens is Sovereign Action’s résumé, published daily: a production agentic system that has shipped every morning without a human in the loop, graded itself in public, and improved by commit. The architecture is transferable to any desk — a morning note built from your book and your models, portfolio surveillance with regime and fragility gauges, research workflows grounded in an LLM wiki of your firm’s own knowledge, and the audit trail a compliance officer can actually examine.
Funds, RIAs, family offices, and research desks engage the firm to build exactly this shape: deterministic scaffolding, one bounded reasoning step, honesty enforced in code. The newsletter is the demonstration; the engagement is the product.
Free while the room is small. Leave your email and the next briefing lands in your inbox before the bell.