the market writes — Lumin reads.
White paper · Multi-agent reasoning
A new language for reading the markets.

LUMIN — a new language for reading the markets

Begin
I  ·  The premise

Machines learned
to read.

A large language model does something genuinely new: it reads. Hand it a paragraph and it holds the whole of it at once — every clause in the light of every other — and arrives at meaning.

Not by matching templates. By understanding context. It is the first machine that reasons over language the way a careful human does.

The leap

The market is
already speaking.

Every chart is a text. Price writes it, one bar at a time — pushes and failures, arrivals and exhaustions. Everything worth knowing is already there, in what the market has expressed about itself.

It had only never been written in a language a reasoning mind could read.

So we wrote one.

A language, though, is a strange thing to build. It cannot be a list of rules — the market would outgrow them by morning. It has to be something a mind can read. And a mind never reads marks one by one — it reads what forms between them. Ask a painter.

II  ·  The language

A painter doesn't work in red, green, blue.

They work in the spectrum — the ambers and magentas that appear only when the primaries meet in context. The primaries are the material; the spectrum is what the eye reads.

Reading a market is the same. The measurements are the primaries. The meaning lives in the spectrum between them — and that is where the reader works.

We mixed that spectrum deliberately. Three layers — each one built from the one beneath it.

From measurement to meaning,
in three layers.

Dimensions
~65 · the pixels

The raw axes of a single bar — its range, its volume, where it closed, its momentum, its place in the structure. Pure measurement, no interpretation.

rangevolumeclose-positionmomentumstructurecluster-density
Components
~170 · the letters

Each measurement read against what came before — a spread breakout, a volume extreme, a new fifty-bar high, a bar that engulfs the last. The vocabulary the bar uses to say what happened.

CHRHCLRLvr↑H50insideengulfxhxl
Concepts
~47 · the words & phrases

The patterns those words form across many bars — the moment buyers arrive, the first pullback that confirms a trend, the exhaustion that ends one. Where the reading actually lives.

weakness-stackrecovery-candlefirst-pullbackseller-arrivalclimax-then-pauseconvergence

Pixels, then letters, then words. A reasoning model fluent in this language reads the chart the way it reads a page. Here is one line of the page.

II · The language  —  one bar, written

This is what a bar says.

April 21, 2025. One daily bar of Bitcoin — the same five measurements every screen in the world displayed. But no bar speaks alone. First, the bars before it, whispering:

xlxvVR2xsTE_CHOP
uxvVR2xsHC7
xhxvVR3sHC7

— then the bar itself:

CH NR h CHW3 VR8 SR7 VB7 SB7 H14 HC21 TE_UP SA_DOM RSI3 90

one bar  ·  the dictionary holds ~170 words — these thirteen spoke loudest here  ·  verbatim from the production engine

A human translation — the way a read sounds: a higher high is an intention — this bar wanted higher. A close pinned at that high is the verdict — it succeeded, and kept it: the fourteen-day ceiling taken, nothing handed back into the close. Participation returns in force after a week of absence — strong, not climactic: conviction without panic. The reach expands with the volume; effort and result agree. First close above everything the last twenty-one days printed. Short momentum ignites; the longer pulses are only beginning to turn. The pause has resolved — and the move is early, not late.

We translate the meaning, not the words. The dictionary stays ours.

Then the third layer assembles — not computed, read. From these thirteen words, in the light of the six quiet bars before them, the reader hears the concepts form:

the coil — resolving expansion bar buyer arrival acceleration the energy arc — turning convergence — defined later in this story

The words are deterministic — any machine can count them. The concepts are not. The same thirteen words beneath a stretched high, after a long climb, assemble into a different sentence entirely — and no buyer arrives. Components are measurement. Concepts are meaning. That is why Lumin is not a recipe.

The translation took a paragraph — and stopped short. The reader doesn't translate — it is fluent. It hears the thirteen words, and the concepts assembling above them, at once — bar after bar, three hundred and sixty-five times in a single read.

hold on to this bar — you will meet it again

Two of those words ask to be heard in full: what effort and result confess when they disagree — and why, when they agree as violently as they just did, the break is built so no one gets back in.

II · The language  —  its integrals · a first reading

Every word reports a struggle.

A reading is an integral — the why beneath the words, the mechanics that make them mean. Take one line from the page and read it back. A bar's size is the result of fuel. When buyers and sellers meet at the same price, their orders absorb each other and the bar stays tight. When one side is thin, price must travel far to find a taker.

So the combinations speak. Heavy volume on a narrow bar is absorption — great effort, no progress, the tension before a break. Light volume on a wide bar is fragility — a thin market pushed easily, ready to snap back. That is the language while the struggle holds. The second reading is what it writes when one side wins.

volume is the fuel  ·  spread is the result  ·  the close names who held

II · The language  —  its integrals · a second reading

When a move is meant to lock you out.

A level holds because one side defends it — orders absorbing every attack, sometimes for weeks. When the defense finally empties, the break is sharp because nothing is left to absorb it. The bar says three words at once: volume expands, range expands, the close lands at the extreme.

The sharpness is deliberate. Price gaps past the level so the traders caught on the wrong side never see their entry again. And no retest returns to offer mercy — the level has been swept clean, and nothing remains to pull price back. That is one-sided flow, written on a single bar.

And the flow shows even before the break. When the up-bars run quiet while the down-bars run heavy — across a flat range, before anything gives way — the book is already tilting. Buyers withdrawing as sellers arrive is distribution; the reverse is accumulation. The reader sees the tilt before the chart admits it.

So the language gives the reader its words. But words alone are not yet meaning — and here the market plays its oldest trick.

III  ·  Why reading wins

The same words can mean opposite things.

exhausted base

Buyers arriving. A wide bar, a surge of volume, a strong close — at the bottom of a long, exhausted decline.

stretched high

Sellers arriving. The very same words — the surge, the spread, the close — near a stretched high, a climb gone far on thinning fuel.

Same volume. Same spread. Same close. Opposite meaning. The difference lives outside the bar — in the level it reclaimed or never broke, in the closes that held or weakened on the way in. Only the story that led there tells you which.

An up-bar after an exhausted base, reclaiming its level, is an arrival. The same up-bar under a stretched high, each close weaker than the last, is the last buyer. A fixed rule sees the measurement and fires either way — at the bottom and at the top, with the same confidence. A reader sees the story — and refrains by default. Silence is the modal answer — and silence is the one thing a rule cannot give.

If the same words can carry opposite meanings, then everything depends on who — or what — is reading them.

III · Why reading wins  —  the limit of rules

A rule cannot refrain.

Inside the market it was built for, a rule is honest. But when the same measurement returns in a world it never saw, it keeps firing — unaware the meaning has changed.

It cannot pause, because pausing would mean reading the trajectory that led here — and a rule has no representation of the trajectory. That is the ceiling of every fitted model. A reader holds the whole path in mind, and refrains by default.

If the whole of the edge is reading the story, then everything turns on the reader. So who, exactly, reads?

IV  ·  The reader

A fresh mind,
every time.

Each read is performed by a reasoning agent spawned with no memory and no bias — only the language it has just learned. It reads one full year of the chart, bar by bar, to the present moment, and forms a conviction.

Most often, that conviction is to do nothing. Silence is its discipline — action is the exception, never the habit.

IV · The reader  —  the read

One year, read to the present.

On every spawn it first reads its curriculum — the language, the cases, the corrections — in seconds. Then it reads the instrument itself: three hundred and sixty-five bars, one full year, in order, to the current bar and not one step beyond.

It cannot see the future; the window stops at now by construction — a property the architecture treats as law, not convention. Whatever conviction it reaches, it reached the way you would: forward, in the dark, one bar at a time.

IV · The reader  —  blind by design

No names. No dates.
No déjà vu.

A reasoning model arrives with a memory of the world — including every famous chart in it. So the reader is blindfolded. It is never told which market it is reading, nor the calendar: bars arrive numbered, not dated — and prices are rescaled, so even the chart's magnitude betrays nothing.

Whatever the model once memorized about any market's history cannot help it here — and cannot bias it. The read can only come from the bars. Every narrative is scanned for leaked years, names and events; the scans come back empty.

no symbol  ·  no calendar  ·  rescaled prices  ·  pretraining sealed off

A reader who says nothing by default needs a reason to break the silence — only one, but an exacting one.

V  ·  Conviction

It acts only when three things agree.

What

The event the bar made — a push that closed at its high, a breakout that failed, a flush met by recovery. The verb.

How

The credibility behind it — the volume, the spread, where it closed. Conviction, or merely theatre.

Where

The place in the story — a base, a trend, an exhaustion. The same event means opposite things in different places.

When all three converge — a coherent paragraph, not a checklist — conviction forms. Lumin specifies exactly one rule, and this is it.

V · Conviction  —  inside a read

Six bars of indecision.

April 2025. Bitcoin coils in a narrow band. Bar after bar the range contracts, the volume drains, the closes turn indecisive — neither side willing to commit. To a rule, nothing is happening.

To the reader, a sentence is forming: a market wound tight, sellers spent, buyers not yet arrived. Then a bar closes high — and the volume that had been absent returns. You have read this bar before — thirteen words, ending in a young move. What: a reclaim. How: on conviction. Where: the floor of an exhausted base. The three agree.

$80K $100K $120K the coil ENTER LONG · $87,466 FLIP SHORT · $118,242 · +35.2%
BTCUSDT · daily closes · Feb – Sep 2025 · the +35.2% price leg, held four months through every pullback
swipe to explore →

The cohort went long at $87,466 — the leg ran +35% over four months.

And yet one reader, however gifted, is still one reader. A brilliant read and a confident mistake can feel identical from the inside.

VI  ·  Robustness

Not one reader.
A cohort.

Several specialists read the same year independently and in parallel, each blind to the others. Capital moves only when enough of them, alone, arrive at the same direction.

One mind can be fooled. An independent consensus is far harder to.

VI · Robustness  —  consensus

Two minds, arrived at alone.

Reasoning is diverse by nature — two readers may weigh the same bar differently. That is not noise to suppress; it is the raw material of consensus. Capital moves only when at least two specialists, from separate spawns that never saw each other's work, reach the same direction independently.

It is the defense against a single mind's two failure modes — a conviction the evidence doesn't support, and a lone reader drifting where none of the others would follow. For a false signal to pass, many independent minds would have to make the same mistake at the same moment. They rarely do.

net P&L +0% +25% +50% +75% +100% +125% L1 L2 L3 S1 S2 S3 the consensus · +123.5% net
dots · six specialists' individual reads of one validation year, before costs  ·  diamond · the production consensus over the full 14-month window, net of all costs
swipe to explore →

the consensus compounded to +123.5% net — higher than any single reader achieved alone

VI · Robustness  —  the team

A team, each kept blind.

Around the reader stands a team, and its defining principle is what each member is not allowed to see.

The specialist reads the market and nothing else. A separate layer turns convictions into disciplined orders — stops, sizing, capital protection — so the reader never has to. The execution layer places those orders without questioning the thesis. And the evidence agents are structurally blind to the specialist's hypothesis, so their data can never be bent toward agreement.

Operations ManagerSpecialistWranglerExecutorData AnalystIntraday Analyst

each question answered with the minimum context it needs  ·  confirmation bias made structurally impossible  ·  execution today is manual; the Executor activates with automated execution

All of this guards the read. None of it explains where the reader's intelligence comes from.

VII  ·  The compounding asset

The language is the asset. The model is the substrate.

Everything Lumin knows is written in plain language — not buried in weights, not retrained. Every lesson the team learns reaches every future reader instantly.

The wiki, the cases, the corrections — every one of them is a page any future reader inherits whole. It gets more intelligent with time.

If the intelligence lives in the language, the language must live somewhere a fresh mind can find it. It lives in a book the team writes together — a living curriculum every new reader inherits at birth. We call it the wiki.

VII · The compounding asset  —  the wiki

Reading is a skill that never transferred.

It accumulates in a practitioner over years — recognizing shapes, weighing evidence, holding through the pullback a rule would flee. Most of it is tacit, and tacit skill resists being written down. So we wrote it down anyway.

Not as rules. As a living wiki a reasoning mind inherits in seconds — taught the way you teach a child to read: pure elements first, where every word has exactly one meaning; then stories, real trades stripped of names and dates so only their shape remains; then corrections; then full worked reads.

manifestocomponents dictionaryconceptscase librarycorrectionsrole files
VII · The compounding asset  —  corrections

Every mistake becomes a lesson.

When a read fails, the team writes a correction — but never "next time, at bar twelve, do this." That is a recipe, bound to the past that produced it, and it misfires on the next variant.

A correction states the root cause instead — the structural reason a kind of misread recurs — in language a reader can carry into any future chart. It is checked, distilled and made universal before it is admitted. And once admitted, it reaches every future reader instantly — no retraining, no rebuild.

Every reader born tomorrow is wiser than every reader born today.

VII · The compounding asset  —  knowledge

Information adds.
Knowledge compounds.

Most systems treat the market as information — patterns found, parameters fitted, recipes fired. Information piles up; the system that holds it does not grow.

Lumin's wiki is a body of knowledge — concepts, cases and corrections that reference one another, so every entry sharpens the reading of every other. Two readers spawned six months apart read different wikis, and the later one is wiser for it.

VII · The compounding asset  —  the substrate

The same language, read by different minds.

Lumin runs on architecturally different AI models at once — Claude Opus and DeepSeek V4 Pro in production, joined in June 2026 by Claude Fable, the newest frontier generation, which read the same curriculum out of the box. Where they agree, the agreement belongs to the framework — not to any one model's quirks. Where they differ, it shows exactly which part of a read is delicate.

As reasoning models advance, Lumin inherits the gain — the same language, read with a sharper eye, no rebuild. And depending on no single provider is, for an institution, a feature in itself.

Claude OpusDeepSeek V4 ProClaude Fableone curriculumsubstrate-portable

That is the idea, entire — the language, the reader, the consensus, the wiki that compounds. A quiet, beautiful machine. One question remains.

VIII  ·  The evidence

Read across every market
it has seen.

2020–26

A chained back-test across the COVID crash, the manias, the 2022 capitulation and the recovery returned a positive result in every calendar year.

$224M

What $1M compounds to across that chained validation, net of every fee and every funding payment — versus $9.3M for holding Bitcoin.

$1M $10M $100M 2020 2021 2022 2023 2024 2025 2026 Claude Opus · $224M net DeepSeek · $421M net BTC buy & hold · $9.3M
$1M compounded across the chained validation, net of taker fees and funding · log scale · two engines, one curriculum
swipe to explore →
The live sleeve · since March 2026 · net of all costs
+38.8%

BTC, live, net — alongside +9.0% SOL and +16.5% ETH, through July 5, 2026. The chart above is the proof; this is its live continuation, visible in real time on the terminal.

In 2022, as buy-and-hold lost 65%, the cohort stayed green — the year reading is meant to survive.

And two architecturally different AI models, reading the same language, both produced the result. The framework does the work.

Reading is cheap: a full per-pair cohort costs about a dollar a day on DeepSeek, ten on Claude Opus, twenty on Claude Fable — varying within ~10% with reasoning depth.

Even reading daily with the newest frontier engine across all three pairs, the entire reading layer costs under $2,000 a month.

All figures net of exchange taker fees and funding, computed from Binance's published history · slippage not modelled · a proof of edge, not a capacity claim · full methodology in the white paper.

Everything the market has to say, it has already said. Lumin is only the first reader fluent enough to listen.

A new language for
reading the markets.

Lumin

LUMIN  ·  multi-agent reasoning model

Houssem Ben Salem, MSc