Who Bragi is
In the old stories Bragi sits at the high seat's side — wisest in words of all, keeping the account of deeds and speaking it plainly, even to kings. He commands nothing and rules no one; he is the one whose account is trusted, because a skald's account can be checked against what actually happened, and everyone in the hall knows it. That is the arrangement here, and as with Hrafn, the name describes the mechanism rather than decorating it. Bragi keeps the account of everything the ravens bring back — a bank argues the memory cycle has peaked, a strategist argues gold is crowded, nine houses argue about inflation — and the seat's work is to weigh what each claim means for this fund, against everything already on the record, on evidence, in a form that can be checked. The counsel is daily by design.
Where Hrafn is many — one raven per document, spawned at dawn, gone by evening — Bragi is one. It reads everything every Hrafn produced, every day, and it is the only seat in the pipeline permitted to hold your holdings while doing so. Hrafn answers "what does this document argue?" Mímir answers "what do you want to know?" Bragi answers the question in between, the one that makes the other two useful: "what does all of it mean for what we own?" The answer is always advice — the decisions stay where they belong, with the fund.
The deliberate asymmetry, completed
Hrafn is blind to the book — by design
A reader who knows the fund is long memory semiconductors finds memory-semiconductor arguments everywhere and reads the document that undercuts the position gently. So the readers are never told what the fund owns. Their extraction is about the market's output, uncontaminated by ours.
Bragi is sighted — by design
Someone must ask "is this good for us?", and asking it honestly requires the actual holdings, their weights, and the dates they were held. Concentrating that sight in one seat, downstream of every reader, means the question is asked exactly once, with the whole corpus and the whole book on the table — and never leaks backward into the reading.
The clearest example of why judgment cannot be delegated to the readers: the mandate inverts signs. This fund expresses AI through the picks-and-shovels layer — memory, power, cooling, the inputs — not the hyperscalers. A bullish call on hyperscaler outperformance is therefore a call against the layer this book owns, and a raven reading in isolation cannot know that. Bragi applies the inversion, names the rule that fired, and attributes the disagreement to the publisher who wrote it, so "the street disagrees with us" always arrives with names attached.
The analysis contract — exactly what crosses each boundary
Every artifact in the chain is a versioned, machine-checked schema. Nothing moves between seats as prose; prose is authored once, at the end, on top of numbers that have already survived their gates.
| Artifact | What it is | Who owns it |
|---|---|---|
theme-conviction |
What arrives from the research floor: a four-by-four grid — the fund's four majors (AI, Minerals, Power Generation, Energy Infrastructure) against four horizons — plus per-document scores, the tickers each document actually named, and the emerging subjects the readers coined that day. Holdings-blind by construction. | Research (Hrafn's floor) |
fund-conviction |
The same measurement with the book joined on. The research half is forwarded
byte-for-byte, never edited; every field the judgment adds carries a
fund_ prefix, so anyone can see at a glance which numbers were measured and which
were judged. Refuses to run if any holding is unclassified. |
Bragi |
fund-strategy · fund-signal |
The stance layer: sections with their supporting evidence rows, then the selection layer that carries the day's strongest material forward to the fund's surfaces. | Bragi |
fund-synthesis + the reading |
The daily synthesis: the grid tracked across the window, the sign census, the book-versus-tape join, the macro strip — and on top of it the reading, the only authored prose in the chain, written fresh each day and falsifiable line by line against the payload beneath it. | Bragi |
The split inside fund-conviction is the contract's core discipline:
research measures; judgment judges; neither edits the other. If a measurement is
wrong, it is wrong upstream and gets fixed upstream, on the record — never silently patched at the
judgment seat.
The engineering gates — why a number on your page can be trusted
Every gate below is a runnable program in the fund's repository, and every one of them ships with a self-test in which each failure case has a passing twin — proof the gate fires on the defect and stays silent on correct data. Several exist because a real defect was caught, and the gate is what makes that class of defect impossible to repeat.
| Gate | What it refuses |
|---|---|
| Statement freshness | No analysis is published on a stale book. If the custodian's statements have not caught up to the report date, the report does not exist for that day — there is no "probably unchanged." |
| The parcel gate | Every reader record is checked before judgment consumes it: unresolvable citations, values more precise than the document stated, dates from the future, findings that assert things about the fund's book the reader cannot know. Defective records are blocked, not smoothed over. |
| The clock assertion | Every day is joined against that day's book. A holding is never counted as "unwritten-about" on a day the fund did not own it. The pipeline refuses to emit if numerator and denominator answer on different book dates. |
| The presence sweep | Before every emit, today's payload is diffed field-by-field against the last good day. A data lane that carried 95% of rows yesterday and none today is a severed pipe, and the sweep stops the run — because a broken pipe renders exactly like a quiet market, and only the diff can tell them apart. |
| Scored-zero refusal | A day on which nothing scored is never synthesised into "the corpus went quiet." Silence from the pipeline and silence from the market are different facts, and conflating them is the most expensive error this system has ever made — which is why it is now structurally impossible. |
| Basis labels | Every percentage names its denominator, beside the number, in the payload itself. A weight on invested capital and a weight on total equity are different answers to different questions, and an unlabeled figure is treated as a defect even when it happens to be right. |
The methodology — how the sharpest signal survives
Relations, never levels
You hold the cash, you placed the trades, you read your own statements — a report that tells you your cash level has told you nothing. A line of the daily reading ships only if it states a relation: a comparison across periods, a join between two series you read separately, a counterfactual against what did not happen. The test applied to every sentence before it ships: could this change a decision? If not, it is deleted, however true.
Breadth before conviction
A conviction score standing on two documents is an artifact of who happened to publish, not a view of the market. The methodology treats document breadth as the test a score must survive: strong readings on thin bases are named as thin, and a score that moderates as coverage triples is reported as exactly that — the honest print replacing the artifact. Disagreement is carried the same way: a contested cell is marked contested, because ten houses arguing is information the average would destroy.
The sign census
Every document row is scored aligned or inverted against the mandate, sleeve by sleeve, with the inverting publishers named and the inversion rule that fired stated. This is how "consensus" is kept honest: the fund sees not just what the street believes, but precisely where the street's belief runs against the layers this book owns — and who is doing the arguing.
Book versus tape
Which held names did the market actually write about, in how many documents, across how many days, from how many houses — and which names has it never mentioned? Silence is measured as carefully as coverage, against each day's actual book, because a sleeve the tape never names has no second opinion, and knowing that is worth more than any score. This join is the reason the seat exists: neither the readers nor the surface can compute it, because only this seat holds both sides.
The macro strip
Subjects outside the four majors — rates, currencies, positioning, policy — are ranked by corpus strength and pointed at the sleeves they read against, with the transmission channel stated. When the strongest out-of-major subjects all bear on one sleeve, that concentration is itself the finding: it tells the fund where its only second opinion is coming from, and through which door.
What Bragi will not do
- Assert what it cannot trace. Every figure in every payload points back to a named field in a named artifact. A number that cannot show its source does not ship.
- Edit a measurement. The research half of every payload is forwarded byte-for-byte. Disputes go upstream, on the record.
- Publish an unlabeled figure. Every percentage carries its basis. Every claim about the book carries the date of the book it was judged against.
- Render a gap as a fact. Null is null, with a reason. A pipeline state is never reported as a market state.
- Narrate your own record back to you. The fund knows what it holds and what it lost in March. The reading exists for what the statements cannot show.
- Reach you directly. Bragi's output goes to Mímir and to the fund's surfaces through their own contract gates. One voice speaks with the fund; this seat makes that voice worth listening to.