Lumin — A Multi-Agent Reasoning Model for trading Cryptocurrency Futures

An example from one validation run makes the mechanism visible. The six-specialist BTCUSDT Opus cohort below — same curriculum, same window — produced six different return profiles, from +15.96% to +82.53% with any open position marked at the window's last bar:

Window: the same 365 daily bars for all six specialists — one full year through the cohort’s current bar. Return is total performance with any open position marked at the window’s last bar. Per-specialist rows are shown before costs; the consensus results in §12 — the figures capital actually compounds — are net of fees and funding.

Each specialist independently read the same 365-bar window — one full year on the 1d timeframe — and the per-specialist returns above are the inputs the consensus operates on. The cohort consensus filtered the divergent reads and acted only where independent specialists aligned — producing the trend-capturing aggregate documented in §12, with returns higher and drawdown lower than any individual specialist could deliver.

The diversity above is structural — not an artifact of any single set of temperature samples drawn at one moment in time. The chained 2020–2026 backtest referenced in §2 was assembled from seven independent six-specialist cohorts spanning seven market regimes; each cohort produced its own per-specialist variance distribution, and within any single cohort, each fresh spawn produces its own again. The diversity holds at every scale. Across realizations, the consensus aggregate is profitable in every one; dispersion concentrates in low-conviction reads while high-conviction setups converge to unanimity. Annex A unpacks a single DeepSeek V4 Pro cohort's per-bar decisions matrix for BTCUSDT — the chronological view of each specialist's ENTER and EXIT flashes — to make the consensus mechanic concrete; further per-cohort breakdowns are available on request.

The cohort's composition is itself quality-controlled. When the cohort's own activity proves an opportunity existed — several independent specialists converging on the same bar — a specialist that produced no read at all is failing to engage, not being legitimately conservative. Rosters are re-validated against the cohort's performance distribution, and persistently divergent readers are replaced with fresh spawns. The team the capital rides is not a fixed cast; it is a curated one.

Cohort size is a tunable parameter: every additional specialist narrows the dispersion toward the architecture's central tendency, with marginal cost in compute, not in architecture. The Opus cohort runs 6 specialists per pair (3 long + 3 short); the DeepSeek cohort runs 18 specialists per pair (9 long + 9 short). Larger N strengthens the confirmation property — see below.

The cohort is also a structural protection against the failure modes of any single LLM read. A specialist that mis-weights a component, that misreads a structural position, that drifts in attention — these are individual errors. Three specialists making the same individual error simultaneously is a vanishingly rare event. The cohort is the team's robustness layer.

The cohort consensus operates as follows. Each specialist holds a live directional thesis (long, short, or flat) at every bar, established by its most recent ENTER and invalidated only when contested by an opposite-direction ENTER from the cohort. Capital is held in the direction of the cohort's confirmed thesis — at least T independent same-direction confirmations — flipping when the confirmation flips, sidelining only on hard-stop breach. The strategy is continuously in market in the direction the cohort believes in. This is the production strategy presented throughout this paper.

9. The Curriculum — externalized intelligence and the pedagogy of reading

What unifies the cohort is the curriculum. Every specialist in every cohort, across every pair and every cycle, bootstraps from the same source — the same manifesto, the same components dictionary and concepts, the same case library, the same corrections, the same framework, the same role file. The variance the cohort filters, the unanimity it acts on, the per-direction discipline described in §8 — all operate on agents that share an identical starting intelligence. Different reads emerge; the curriculum that produced them does not.

That curriculum is not a knowledge base. It is the deliberate externalization of accumulated reading intelligence into a structural form a reasoning agent can ingest in seconds.

Reading the market well is a tacit skill. It accumulates in human practitioners over years of work — recognizing shapes, weighing components, distinguishing climax from continuation, holding through pullbacks the rule would exit. The skill resists transfer because most of it is implicit. Two practitioners with identical experience can disagree on identical setups because the integrations they perform are private. Conventional knowledge transfer — books, courses, documentation — captures the surface vocabulary but loses the integration logic. The work that intelligent reading actually performs has historically not been transferable.

We have written it down. Not as rules — rules cannot encode contextual judgment. As language, stories, and corrections, in a structural form that a reasoning agent fluent in language can fluently consume. The curriculum is the years of practice rendered into atomic, linkable, navigable knowledge that any fresh agent can inherit in full fidelity.

The pedagogy is the pedagogy of teaching a kid to read. Pure elements first. Stories second. Corrections third. Teaching reads fourth. And throughout: never confusing, never ambiguous, every element coherent with every other, the kid's attention treated as sacred.

Pure elements

The components dictionary is written with severe discipline. Each component is named precisely; its meaning is described in language; its dependencies on other components are declared with semantic backlinks. There is no hidden coupling, no overloaded terminology, no two definitions for the same thing. The kid learning the dictionary acquires a vocabulary in which every word has one meaning and every meaning has one word. Confusion is the curriculum's worst failure mode, and the dictionary is engineered to make confusion structurally impossible.

Stories

The case library is the team's accumulated reading experience, preserved as full bar-by-bar sequences. Each case is a complete trade — from trend birth, through every read at every bar, through entry, management, and resolution — with the verbatim reasoning at each step. The cases are taught in anonymized form: pair names become abstract identifiers (pair X), calendar dates become relative bar numbers indexed from trend birth (bar 1, bar 2, …, bar n), and price levels are abstracted to relative magnitudes. What remains is the structural shape — the way the components moved through the trajectory, the way the read evolved bar by bar, the way the resolution emerged. The kid reads stories about kinds of arcs, not about specific instances. When the kid encounters a live sequence that rhymes with a story, the rhyme is structural, not nominal.

Anonymization is the deepest guard against overfit. A reader that has internalized the structural shape of a market reversal can recognize that shape in any new instance; a reader that has memorized particular dates and prices has memorized only noise. The case library teaches shapes, not specific instances; the reader inherits structural form, never answers bound to the past that produced them.

Corrections

A correction is an integral, never a derivative. "If X then Y" is a derivative: a fitted answer bound to the single past instance that produced it, brittle the moment conditions drift, wrong in silence when the same surface configuration recurs in a different context. An integral is the opposite move: it states the structural root cause of why a class of reading errors recurs, in language a reader can carry into any future configuration where that structure is present. Root causes, not recipes. Lumin's corrections library is gated to admit only the second.

When a specialist's read fails — a missed entry, a forced-fit entry, a held-too-long, a too-early exit — the team performs a post-mortem and produces a correction. The correction states the universal lesson: the attention error named in terms that apply to any pair, any timeframe, any future read — never bound to the specific pair or setup that surfaced it. A correction does not say "next time enter at bar twelve when the bricks are green." A correction says, in the form the reader can carry into any future read on any future pair, what attention was missing and what attention is required when this configuration recurs. The correction links back to the case that produced it and forward to the concepts and entry types it refines — connections every future specialist follows during its bootstrap, so that the lesson reaches the read at exactly the moment the read needs it.

Corrections do not auto-ingest. The decision of whether a correction belongs in the curriculum — and whether its formulation is an integral rather than a derivative — is gated, not automated, but the gate itself is run by the framework, not by a single reviewer. When a specialist's cycle resolves, the Operations Manager (§11) spawns dedicated correction-ingestion agents into their own zero-context windows. One agent distills the specialist's resolution report into a candidate correction in plain, universal language. Three further agents — each isolated from the others — apply the three failure-mode checks: redundancy (does it duplicate or near-duplicate an existing correction in the library); overweighting (would adding it bias the reader's integration toward a concept beyond what the underlying market structure warrants); ambiguity (does its phrasing risk confusing future specialists about the universal lesson). Each agent reads only its own gate question, against the existing curriculum and the candidate text. Only candidates that clear all three independent agent reviews are presented to the human curator for final admission to the library.

The gate is process, not heuristic — and the process is structurally distributed. The human curator's role is final accountability and approval, not ad-hoc judgment from scratch; the framework's agent-gated discipline is what actually decides whether a candidate is universal, non-redundant, and unambiguous. Every step is plaintext, every artifact is versioned in git, every admitted correction is lint-validated against the rest of the curriculum for cross-reference and tone coherence. The discipline is procedural and reviewable by construction; this is part of what makes the curatorial mechanism auditable by an institutional deployer and resilient to single-person dependency.

The curatorial gate is an architectural commitment, not a process detail. Autonomous ingestion would risk derivatives flowing in without integral discipline. Redundant ingestion would over-weight the reader's reasoning by piling repeated emphasis on the same area, biasing integration toward what is over-represented rather than what is structurally important. Ambiguous ingestion would corrupt the language itself. The discipline of curation is a form of overfit protection in its own right — possibly the most important one. Every addition is checked, distilled, made universal, and only then granted the structural privilege of reaching every future specialist.

Corrections compound. A correction filed at fourteen hundred today reaches every specialist that bootstraps from fourteen hundred and one forward. There is no retraining, no parameter migration, no versioning. The next specialist that spawns reads the new correction in the corrections library and applies its lesson during its own integration. Two specialists that spawn five minutes apart will read the same corrections set; two specialists that spawn six months apart will read substantially different sets, and the later specialist will be substantially more intelligent because of it. The kid grows up.

Teaching reads

Beyond cases and corrections, the curriculum includes full teaching reads — extended walkthroughs of complete cycles, anonymized in pairs and dates, in which a specialist reads an entire window from trend birth to final resolution, with reasoning at every bar. Teaching reads show how the components, the corrections, and the framework integrate when applied with discipline. They are the curriculum's worked examples — the pedagogical artifact that, more than any other, transfers the integration logic.

Attention protection

We protect the reader's attention with the same care we protect the curriculum's coherence. The specialist runs on a reasoning-class model chosen for stable long-context attention — the property, natural to selective-state-space and sparse-attention architectures, that the earliest evidence in a long prompt remains as available to the integration as the latest. Even with the right substrate, every additional token is some dilution. The curriculum's atomic structure — one concept per file, one case per file, one correction per file, one role per file — exists so the specialist can selectively load only what the current read requires. The bootstrap is engineered for minimum sufficient context. The role file declares precisely what the specialist must read and what the specialist is forbidden to read. Cross-context pollution is the most expensive mistake the team can make, and the architecture makes it structurally hard.

The curriculum's discipline is what makes this attention protection work. Because every concept is unambiguous, the reader never spends attention disambiguating. Because every correction is universal, the reader never spends attention adjudicating between a correction and the live read. Because every case is anonymized, the reader never spends attention recognizing instances. The reader's attention is reserved entirely for the live integration of components in context — the only thing that produces the read.

The architectural innovation

Most algorithmic systems try to capture intelligence by parameterizing it — fitting models to historical data and deploying the fitted parameters. We capture intelligence by externalizing it as language a reasoning agent can ingest and apply, with the curriculum's discipline ensuring that the language is pure, coherent, and unambiguous. The first part — the externalization — is the long, deliberate work. The second part — the specialist applying it — is the production substrate. Together they constitute a different category of model: an automated system that holds the contextual judgment of a master practitioner, applied with the consistency of a machine, refined daily by every new case the team resolves. The curriculum compounds. The reader stays fresh. The intelligence accrues.

Knowledge, not information

Most systematic systems treat market intelligence as information — patterns identified, parameters fitted to past data, recipes triggered by signal. Information accumulates by addition; the system that consumes it does not grow with it. Lumin's curriculum is not a library of information. It is a body of knowledge: a directed, monotone-growing graph in which concepts, cases, and corrections cross-reference one another so that every entry strengthens the integration of every other. A specialist's bootstrap is not a linear read of a manual — it is the integration of the relevant subgraph for the live read. And every cycle that closes deposits a new entry into the graph through the coherence gate, reaching every specialist that ever spawns thereafter, instantly and without retraining. Information accumulates linearly. Knowledge compounds.

10. The Harness — engineering the read

Two architectural choices precede the curriculum and the cohort. The first is what we removed. Earlier iterations of the team experimented with elaborate coordination patterns — specialists handing context to other specialists, agents reviewing and refining one another's outputs mid-read. Every one of those patterns proved costly to the only thing that matters: the specialist's attention on the bars. We removed the coordination patterns — the specialists stayed. The specialist now reads alone, fresh, against the curriculum and its instrument, with nothing intervening between it and the tape.

The specialist's read environment is engineered. The bootstrap is delivered as a single XML-structured prompt — sections for role, components, cases, corrections, framework, the bar window, and a strict statement of what the specialist is forbidden to read. The format is closed and parseable: no preamble, no natural-language imperative, no nested conditionals the substrate might reinterpret. The role file is the contract; the rest of the prompt is the read environment, assembled atomically from the curriculum at bootstrap time.

The read itself is uninterrupted. No intermediate validation, no progress check, no intervention. The specialist completes its integration, returns its conviction with verbatim reasoning, and only then does the team's after-read audit begin.

The after-read audit is where the team's discipline runs. Every resolved read is processed through a fixed sequence: structural extraction of the conviction and its reasoning steps; cross-check against the cases and corrections it should have integrated; quantitative audit of component weighting against expected ranges; internal consistency check across the bar-by-bar trajectory; coherence check against the role file's forbidden-context list. When the cycle resolves, the specialist's predicted trajectory is compared against the realized one, and the delta — entry timing, target accuracy, exit reasoning, holding duration — drives the team's curatorial work on the curriculum. Reads or trajectories that fail any stage surface either a correction to the curriculum, a refinement to the role or prompt, or — in the rarest case — a structural issue with the harness itself.

The economics of a read are concrete and small. A full per-pair cohort — every specialist read plus orchestration — costs on the order of one dollar per day on DeepSeek V4 Pro, ten dollars on Claude Opus, and twenty dollars on Claude Fable, the newest frontier engine to read the curriculum; costs vary within roughly ten percent with reasoning depth and token consumption. Even reading daily with the frontier engine across all three production pairs, the entire reading layer runs under $2,000 per month — less per year than a single junior analyst, and its unit cost falls with every generation of model pricing.

The harness is the third structural pillar of the architecture, alongside the curriculum and the cohort.

Per-substrate harness — meeting the architecture where it lives

A different model is a different reader. The naive way to adapt to one is to rewrite the prompt, tweak the calibration, iterate until something works — the per-model patches systematic adoption typically resorts to. That is a derivative move: respond to symptoms, fit to instances. Our reasoning, on the curriculum side and at the framework level, is integral: understand the structure, intervene at the root. The same discipline applies to substrate adoption. A new substrate that under-reads the curriculum is not a problem to patch by trial and error; it is a structural question whose answer lives in how that substrate's attention machinery actually processes context.

Lumin currently runs on two substrates: Anthropic's Claude Opus as the canonical reader, and DeepSeek V4 Pro as a second operational substrate. The two are architecturally different in a way that matters for long-context reading.

Claude Opus runs dense long-context attention — every token of the prompt is uniformly available to every other token across the full span, so the curriculum's atomic structure (one concept, case, or correction per file) integrates without engineering intervention.

DeepSeek V4 Pro takes a different path: it compresses older context to keep long-window reading affordable. By the time the bar window is reached, the curriculum at the front of the prompt has been folded into compressed summaries. Lumin's harness restores it by injecting a condensed re-pointer to the curriculum's vocabulary — manually distilled once per substrate, not regenerated per cycle — at the position immediately preceding each bar's read, where the substrate's attention is empirically full-bandwidth (Diagram 1, step 3b). The intervention is integral, not derivative: the curriculum is unchanged, no model-specific rules enter the framework, and the re-pointer is applied identically thereafter.

The full mechanism — the per-substrate attention architecture (CSA/HCA), the re-pointer construction, and the integral-vs-derivative reasoning at the engineering level — is documented in a separate technical note available on request.

DeepSeek V4 Pro was added later as a frontier-model diversification track. The cohort now runs concurrently on two architecturally different readers — Claude Opus, the original substrate, and DeepSeek V4 Pro — each integrating the same curriculum on its own cohort. The reads are substantively equivalent: across the chained validation both substrates produce a profitable consensus, sometimes converging and sometimes diverging bar-by-bar. Where they converge, the convergence is substrate-independent — evidence the curriculum is what does the work, not any single model's quirks. Where they diverge, the divergence localizes which side of the read is substrate-sensitive.

Substrate diversity is the load-bearing claim. When two architecturally different readers reading the same instrument against the same curriculum produce equivalent results, the framework — not any single model's quirks — is what does the work. Provider diversity is the deployment consequence: substrate redundancy hedges frontier-vendor exposure, a property structurally relevant to institutional deployment and one that keeps the project independent of any single model provider.

A note on the §1 portability claim. The phrase "without retraining, recalibration, or rebuild" describes ongoing operation: once a substrate is adopted, the curriculum continues to compound and every new case reaches every future spawn without any per-cycle adjustment. The substrate-aware harness above is the one-time adoption cost — written once per substrate, applied identically thereafter. Adoption is a one-time engineering investment; operation thereafter is parameter-free.

11. The Team

The team did not begin as an org chart. It began as a single role — the market specialist, the reader — and the recognition that the reader's attention is the system's scarcest resource. Every other role was added because the reader needed something protected: a fresh context window, an isolation contract, a capital decision separated from market reading, an execution layer blind to the framework, evidence agents structurally prevented from contaminating the read. The team is built bottom-up around the reader, not top-down from an organizational template — each role exists because removing it would force the specialist to spend attention on something other than the live integration.

Lumin therefore operates as a multi-agent team that runs against a curated set of pairs — the pairs Lumin rides. The pair set is selected upstream of the team and treated as fixed for operational purposes; the team's job is to read these pairs intelligently, cycle by cycle, not to screen the broader universe. Each role within the team is a separate spawn with a separate context window, a separate role file, and a separate isolation manifest. The team is real — not a single agent simulating personas, but distinct agents with strictly bounded responsibilities and structurally enforced boundaries.

The Operations Manager conducts the team's lifecycle. It spawns specialists at the right moments, routes their reports to downstream agents, commits resolved cases and corrections back to the curriculum, and maintains the team's operational state across each cycle. The Operations Manager does not read the market.

The Market Specialist is the reader. Specialists are spawned per pair, per direction, per cycle — fresh, isolated, intelligent — and produce the directional convictions on which the rest of the team acts. Specialists are unidirectional: the long thesis and the short thesis are read by separate specialists, so neither read is a compromised average. Specialists are the unit of value in the system; everything else exists to keep them sharp.

The Wrangler is the capital decision layer. It receives the cohort's aggregated convictions, applies the strategy rules that govern how convictions translate into capital action, and produces order tickets (internally called order envelopes) — the team's capital-action record, written once per confirmed flip. Each ticket names the side, the entry, the hard-stop level, and the supporting specialist reads. Each ticket updates the position ledger (internally called the vault) — the versioned, single-source-of-truth record of the team's live position, queryable by every subsequent cycle. The Wrangler enforces protective discipline — hard stops, position sizing, capital conservation — that the specialist's read does not need to consider, because separating capital protection from market reading keeps both functions honest.

The Executor is the platform mechanic — the role that takes the order ticket to the exchange. Its design is complete: it tracks funding rates, slippage, fees, and order-book depth; it caps each entry at a structurally-prudent fraction of the pair's 24-hour traded volume — a liquidity-first sizing discipline — and slices oversized entries across time to limit market impact. The Executor is blind to the framework — it does not question the call, it executes it. Status: design complete; it activates with automated execution. In today's production the confirmed signal is executed manually against the exchange, and the Executor's discipline is applied by hand.

The Data Analyst is the historical pattern reference. It performs blind statistical analysis against the team's narrative database — blind in the sense that it is structurally prevented from seeing the specialist's hypothesis, so its evidence cannot be biased toward confirmation. The specialist consults the analyst when historical reference is helpful; the analyst returns evidence, never a verdict.

The Intraday Analyst is a one-minute microstructure reader, consulted on demand — never by default. The specialist consults it only when the daily bar closes on a wide range with an ambiguous close position, the configuration in which the bar's interior, not its OHLCV summary, determines which side held the day. Consults skew toward smaller-cap pairs where wide-spread bars with ambiguous closes occur more frequently; on the production majors the consult is rare. Outside that specific configuration the Intraday Analyst is not spawned.

The two evidence-agent roles above (Data Analyst, Intraday Analyst) are on-demand roles: they exist in the team's design and are spawned only on the configurations they were designed for — which, on the current production majors, means rarely. The §7 worked examples are both high-conviction BTC reads on a major with clean-close bars, so neither analyst is consulted in either example. On smaller-cap pairs or on bars where intraday microstructure carries the read, both roles enter the cycle.

The architectural commitment that holds the team together is the isolation matrix: a per-role, per-file access manifest, declared in each role's spawn configuration as machine-readable access_files and access_denied lists, that defines what each agent's context may and may not contain. Evidence agents are blind to the specialist's hypothesis. Execution is blind to the framework. Each role's context is the minimum sufficient set for its specific question. The matrix is the system's answer to a problem single-agent architectures cannot solve: a single agent reading everything reads with bias on everything; a multi-agent team with role-bounded isolation reads each question with the minimum sufficient context for that question, and the absence of contaminating context is itself a feature.

An honesty disclosure on isolation. The team's principles file names this candidly: isolation is enforced at the contract layer — spawn configuration, role manifest, framework discipline — not at the operating-system or kernel layer. A specialist's role file declares the files it must read and the files it must not, and its bootstrap loader honors that declaration; a kernel sandbox does not refuse a forbidden read. The same is true for every other role. We treat this as a control because the contract is documented, versioned in git, lint-validated by tooling, and audited by every code review that touches a role file or spawn manifest — not because it is technically unbreakable. We name the limit so an institutional auditor does not need to surface it in due diligence: the discipline is procedural and reviewable; deepening it to kernel-enforced sandboxing is on the deployment roadmap, not in the current substrate.

12. Results

Lumin's production strategy holds the cohort's confirmed directional view continuously: capital is long whenever the long-thesis specialists meet the confirmation threshold T, short whenever the short-thesis specialists do, flipping when the confirmation flips. The strategy is always positioned in the direction the cohort believes in; it sidelines only on hard-stop breach.

We present two complementary validations. The first establishes multi-regime durability: a chained 2020–2026 BTCUSDT backtest spanning seven independent six-specialist cohorts run consecutively across seven market regimes — COVID, the 2020–21 mania, the 2022 capitulation including FTX and LUNA, the 2023 recovery, the 2024 ATH, and the 2025–26 cycle — same architecture, same curriculum, same confirmation rule across all seven cohorts, with the equity from each cohort compounding into the next. The second establishes multi-instrument generalization: a recent 14-month deployment on three institutional-grade Binance USDT-M perpetuals (BTCUSDT, ETHUSDT, SOLUSDT) run concurrently. The recent BTCUSDT cohort is the seventh and final link of the chained series — the same trades on the same tape — so the two evidence sets reinforce rather than duplicate one another.

Multi-regime durability — BTCUSDT chained 2020–2026

The chained 6.4-year window resolves to 76 months of returns (January 2020 → May 2026). The histogram below shows month-by-month performance under the production strategy; the table aggregates institutional risk metrics across the full window.

Edge is time-distributed, not concentrated. Of 72 months, approximately 72% produced positive returns net of costs; the longest losing streak across the full window is 3 months. The maximum drawdown duration — the longest stretch spent below a prior NAV peak — is five months. No single month dominates the cumulative return.

^*Sharpe computed from monthly returns: (mean monthly arithmetic return ÷ monthly standard deviation) × √12. The annualized return above is the CAGR over the chained window; the two are not derivable from each other by simple ratio.

Multi-instrument generalization on Claude Opus — three pairs, 14-month window (Feb 2025 → May 2026)

The same architecture was deployed concurrently on BTCUSDT, ETHUSDT, and SOLUSDT to confirm that the read generalizes across instruments at the same time. Each pair's window is 365 daily bars — one full year on 1d, the canonical window over which trading decisions accrue. Six fresh-spawn specialists per pair (three long + three short), 1× notional, no leverage, flip-only. The figures below extend through the live sleeve's current cycle (2026-05-27); the per-pair Opus trade ladders are detailed beneath.

Leg P&L and equity in the ladders below are net of taker fees and of funding accrued over each leg's holding period. The table and trade ladders show the 14-month backtest equity from each pair's $1M baseline at the cohort's read-window start (2025-02-26 for BTC, 2025-02-25 for ETH and SOL). The §2 live-sleeve percentages (+22.72% / +6.14% / +10.89%, net) are a different baseline: the production sleeve was funded with fresh $1M-per-pair allocations in March 2026 and the percentages reflect each sleeve's NAV through 2026-05-27. Same strategy, same pairs, same execution — different start dates and different starting capitals. The two evidence sets are independent rather than additive.

Three pairs, one strategy, the same 14-month window extended through the live sleeve — all figures net of fees and funding. Net returns range from +123.47% to +373.83%. Sharpe ratios cluster between 2.38 and 2.73. Maximum drawdowns sit between 14.20% and 22.19% — versus buy-and-hold drawdowns of 49% to 69% on the same windows. Per-pair Sharpe ratios here are computed from the daily net equity path, annualized. The difference is the read.

Several properties of the multi-instrument results bear emphasis.

Realized return is uncapped. The first BTCUSDT long held a +35.19% price move across approximately four months. The October short held a +40.11% price move across five months — through multiple intra-trend bounces. None would have survived a fixed-target system; all were captured by a reader that integrated continuation evidence bar by bar and held until the read was invalidated.

The cohort is conservative on entry. Five trades over fourteen months on BTCUSDT — roughly one trade every three months. The modal output of the system is no entry. The selectivity is what produces the asymmetric payoff.

Drawdown is bounded by the read, not by stops. Over the recent 14-month BTCUSDT cohort, maximum drawdown was 14.20% on a net mark-to-market basis. The wider 30.22% figure reported in the chained section above is computed across the entire 6.4-year equity series and reflects earlier-regime drawdowns the recent window does not contain — primarily the 2022 capitulation. The buy-and-hold reference for the recent 14-month BTCUSDT window had a maximum drawdown of −49.56%.

BTCUSDT — trade ladder

The BTCUSDT cohort traded six legs across the evaluation year, capturing the major regime transitions of the period — the April long from $87K to $118K, the August flip to short, the October short capturing the $121K → $72K decline across five months, the post-bottom long held through to the May 13 flip, and the post-flip short currently in market.

ETHUSDT — trade ladder

The ETHUSDT cohort traded six legs across the evaluation year. The first long held from $1,756 to $4,374 — a +149.16% price leg across four months. The October flip captured the parabolic top and ran nearly five months to $2,057, a +53.74% price leg. The February-onset long extended +9.77% in price through to the May 13 flip; the post-flip short is currently held.

SOLUSDT — trade ladder

The SOLUSDT cohort traded twelve legs — a higher-frequency profile reflecting SOL's denser regime transitions across the window. The cohort captured the spring rally to $234, the autumn short to $132, the January-2026 push and reversal, the spring-2026 chop with disciplined consensus flips, and the May 13 flip to short that recovers the chop drawdowns through the post-flip leg currently held.

Funding as signed directional exposure

Perpetual futures carry funding — a periodic payment between long and short holders that compensates for the difference between perpetual and spot. Funding is not merely a deductible cost. For a strategy with persistent directional positioning, funding is signed exposure: long-funding regimes (longs pay shorts) erode the long thesis materially over multi-week holds; short-funding regimes pay the short thesis to wait. Lumin's directional posture across funding windows is therefore part of the economics, and this paper now accounts for it directly: every figure in this document deducts funding event by event from Binance's published funding-rate history against the position actually held. The effect is visible in both directions. The 2020–2021 bull cohorts — long through persistently positive funding — paid the largest tolls of the whole validation. The October 2025 BTCUSDT short, held through a negative-funding episode, was paid to wait; the live sleeve's current short legs are net-better than their gross because of the same effect. Across the full 6.4-year chained window the strategy's asymmetric, multi-month holds proved structurally robust to funding: the chained net result remains within sight of the gross one, and no calendar year flipped from profitable to unprofitable on costs.

Glossary — the paper's terms in plain words

13. Closing

The market is structured, but its structure is contextual. The components are a language of meaning, and meaning is integral, not derivative. We do not decompose the market down to its primaries and then trade the primaries; we read the spectrum the primaries form when integrated in context. Convergence is the form that integration takes. The reader is what performs the integration. The cohort is what makes the read robust. The curriculum is what makes the reader intelligent. The team is what protects the read.

What we have externalized — and what no parameterized system can — is the tacit reading intelligence that accumulates in a practitioner over years of practice with markets. The curriculum is that intelligence written in structural form: pure elements, anonymized stories, universal corrections, full teaching reads — never confusing, never ambiguous, every page coherent with every other. A fresh reasoning agent fluent in the curriculum's language performs in seconds the integration that took the practitioner years to learn to perform.

Because the curriculum holds the language of structure rather than the prices of any past instance, the same reader applies to any new pair, any new timeframe, any new regime, with the same machinery — there is no re-mining. Because there are no thresholds, no fire conditions, no fitted parameters, there is nothing to overfit; the reader's epistemic position is the live trader's, never the puzzle-solver's. The reader is designed to say no almost always — the system exists to find the trade, not a trade — and selectivity is what produces the asymmetric payoff. And every cycle the team operates, the curriculum becomes richer for every future read; capital deployed on Lumin benefits from operational time as a function of accumulated intelligence, a property no parameterized system structurally permits.

The reader's attention is the system's scarcest resource. We never let the reader look ahead. We never let the reader overfit. We feed the reader the spectrum, not the primaries. We let the reader read.

© Houssem Ben Salem, MSc · Lumin · Architectural & Methodological Whitepaper · 2026