HMAS technical record · detailed v0.7.2 · pre-audit architecture · 2026

Orchestrated Hierarchical Multi-Agent System (HMAS): Technical Record

The detailed architecture, research engine, evaluation mathematics, execution path, operating controls, and appendices from the original HMAS paper.

💬 Ask Andy about this white paper →

Read this as a technical record, not a current operating claim. This edition preserves the detailed design that preceded the full-system audit. It explains what was designed, implemented, and intended, but it does not imply that every feedback loop operated as described. The audited field report separates what was built, what the audit established, and what now belongs to the replacement. Keeping both editions visible is deliberate: the technical depth matters, and so does the correction record.

Public boundaries. Methods are shown. Live thresholds, capital, credentials, and proprietary strategy parameters remain withheld.

v0.7.2 — a correction. An independent audit of this paper's load-bearing claims found two in §8.2 stated more strongly than the recorded results support. The rejected challenger cleared its configured 90% gate minimum at 91% and lost to the incumbent's 93% — it did not fall below its promotion bar; and the largest candidate did not "fail the honesty track outright" — it scored 93% there with zero hallucinated completions and was beaten by a better model, while falling below the bar on a different battery. Both sentences are corrected above, with the measured figures. §2.6's model-size comparison, which conflated the local and cloud batteries, is corrected too. Publishing the correction rather than quietly editing it is the same discipline §7.5 describes.

Abstract

HMAS was designed to avoid two common failures in automated investment research: trusting a single model's judgment and optimizing a single headline metric. It is an orchestrated hierarchical multi-agent system in which a coordinator delegates strategy-discovery work across tiers of reasoning agents, while fixed software checks evaluate candidates using several statistical methods rather than one score. This document describes that intended architecture, the evaluation mathematics, the regime and risk layers, and the operating machinery. The later audit showed why fixed checks are necessary but not sufficient: they can only protect the system when their inputs, error paths, and downstream effects are also observed.

I · Thesis & first principles

This system began with a single conviction, stated plainly so it could be argued with: systematic investing is the proven edge; trading is R&D. Discretionary trading bets on a human's judgment in the moment — it cannot be audited, repeated, or compounded with any confidence. A systematic process can. Rules that are written down, tested against decades of history, and held to a standard they must keep clearing on data they have never seen are the only kind of edge that survives contact with reality. So the deliverable here is not a trade. It is a process that discovers candidate strategies, subjects them to a standard of proof, and retires them the moment they stop earning their place. Putting real money to work is the last and smallest step; most of the engineering is the research — and the discipline of refusing to deploy what has not been proven.

There is a second deliverable, and it is stated here because it changes how the work is judged. This system is also a controlled experiment in a more general question: what harness does a probabilistic language model need before its output can be trusted with consequential decisions? Language models do not compute answers; they sample plausible ones — which means fabrication is not a defect to be patched but a property to be engineered around. Markets are the ideal laboratory for that engineering because they are adversarial and keep an objective scoreboard: fabricated competence is eventually priced. So every re-evaluation of this system (§7.5, §7.6) does double duty — it improves the trading process, and it extracts another rule of the harness: another check, another invariant, another balance that transfers to any domain where a confident wrong answer is expensive. The method is the scientific method, applied to the system itself: hypothesize, measure, publish the unflattering result, and encode the lesson so it cannot silently recur. The financial return is one product; the harness knowledge is the other, and it is the one with implications far beyond this system.

1.1 · Three bodies, adaptive markets, and the dark forest

The three-body problem supplies a useful limit on the ambition to predict. An archival NASA explanation describes general three-body motion as chaotic: known laws do not produce an explicit formula for indefinite prediction, even though particular equilibria exist. Markets are not celestial mechanics, so the comparison should not be stretched into a model. Its value is narrower: interacting participants make the path sensitive to conditions, and that makes exact long-range point prediction a poor engineering objective. The response is process— search broadly, test on unseen data, measure deterioration, and control exposure.

Andrew Lo's Adaptive Markets Hypothesis gives that intuition an economic foundation. Behavioral finance challenged the idea of continuously rational markets by documenting decisions shaped by fear, greed, loss aversion, overconfidence, and imitation. Lo's synthesis treats these behaviors within an evolutionary environment governed by competition, adaptation, and natural selection. Participants learn, enter, leave, copy, and change tactics. Efficiency therefore varies with the environment and the population competing within it; yesterday's market is not guaranteed to remain in place for tomorrow's strategy.

Liu Cixin's Dark Forest remains useful as an operating metaphor, provided it stays bounded. An edge is a scarce opportunity in a competitive environment. Publishing its parameters, trading it visibly, or attracting imitators can change the conditions that made it valuable through copying, crowding, or counter-positioning. That explains why this paper shows its methods but withholds live parameters, and why forward performance must keep earning the strategy's place. It does not establish that every strategy inevitably decays, and it cannot explain away overfitting, poor data, or an evaluation loop that failed to close. Those remain engineering failures to own.

HMAS is designed to search an environment that cannot be solved, test evidence produced by participants who adapt, and withdraw when the conditions supporting an apparent edge no longer exist.

1.2 · Three problems that force a multi-agent design

A single model, however capable, cannot run this process well — for three reasons that compound:

The search space is too large for one mind. The space of strategy ideas, parameterizations, and universes is effectively unbounded. Covering it means fanning the search out across many workers exploring in parallel, not marching one reasoner through it serially.
Always-on top-tier thinking is uneconomic. The most capable models are expensive per token. Paying that rate on every routine step — listing files, reshaping data, running a backtest — burns the budget on work a cheap local model does just as well. Cost has to scale with the value of the decision, not be paid flat.
The honesty problem. Language models fabricate at the reasoning layer. A model asked whether its own strategy passed will, often enough, say yes. Any architecture that lets a model grade its own work inherits that failure. The judgment of whether a candidate is real cannot be allowed to live inside a model at all.

1.3 · Design invariants

Those three problems pin down a small set of non-negotiable rules:

An evidence bar. Nothing is deployed on a promising backtest. A candidate is promoted only after it clears a fixed, pre-registered standard of proof.
A conjunction gate, not a Sharpe ratio. The bar is a set of independent tests that must all pass, not a single number to be maximized. Optimizing one metric is how you overfit; demanding agreement across several uncorrelated tests is how you don't. (The mathematics are in Part IV.)
Trust-via-harness. Workers are trusted only because the harness around them makes dishonesty impossible to pass off — a sandbox they cannot escape, a fixed check they cannot talk their way past, and a top-tier review on the decisions that change state.
Spend proportional to decision value. Cheap local models do the bulk exploration; expensive models are reserved for the few decisions that actually change what the system holds or believes.
One evidence bar across mandates. A conservative core and a ring-fenced high-risk sleeve face the same standard of proof. They differ in how much drawdown they will tolerate — never in how much evidence they require.

1.4 · What this system is not

It is not a black box: every promotion is reproducible and ships with the evidence that earned it. It is not a single-metric optimizer: there is no one number it tries to maximize, by design. And it is not auto-trading real money: live capital sits behind a forward-test gate and an explicit human approval, because the honest default for a research system is to assume it is wrong until weeks of out-of-sample evidence say otherwise.

II · The HMAS architecture

HMAS — an orchestrated hierarchical multi-agent system — is best read as an organization with a chain of command, not as a single program. Work flows down from a coordinator to ranks of workers; evidence flows back up; and a fixed software referee sits outside the model chain. The later audit showed that this separation limits one class of error without making the referee immune to bad inputs or permissive failure. Three hierarchies interlock.

2.1 · Three interlocking hierarchies

Cognitive / compute — models ranked by capability and cost, from cheap local workers up to a top-tier "general."
Research-control — a coordinator that decides what to explore, and executors that carry out each experiment in isolation.
Evidentiary — the gauntlet: a fixed body of statistical law that every candidate is tried against, answerable to no agent.

The first two are hierarchies of command; the third is a hierarchy of proof. Keeping them separate is the entire point — the agents that want a strategy to succeed have no authority over the thing that decides whether it did.

2.2 · The chain of command

Compute is organized like a rank structure. Local infantry — small models running on owned hardware — do the high-volume, low-stakes work: drafting code, reshaping data, summarizing, running backtests. Cloud NCOs — mid-tier hosted models — handle what needs more reasoning than the infantry can give but does not warrant the general. A top-tier general is called in only for the decisions that change state: approving a merge, signing off on a capital move. Each rank is directed by the rank above it and escalates only what it cannot settle. The cost curve follows the ranks — most tokens are spent at the bottom, where they are cheap, and the expensive model is paid only for the handful of decisions worth its rate.

2.3 · The control plane

Every model call — local or hosted, cheap or premium — passes through a single routing endpoint. That one choke point is where budgets are enforced, where a request is sent to the right tier, and where a model can be swapped out without touching the code that called it. It also makes the economics explicit: spend is metered in one place, the system runs primarily on owned hardware, and it escalates to paid tiers only under budget and only when a decision earns it.

2.4 · Orchestration

A coordinator maintains a tree of ideas to explore and dispatches each one to an executor. Every experiment runs in its own isolated worktree — a private copy of the codebase — so parallel workers cannot corrupt one another's state, and a failed or dishonest experiment can be discarded cleanly. Nothing an executor produces touches the trunk until it has passed the harness. This is ordinary software-engineering discipline — branch, test, review, merge — applied to strategy research.

2.5 · Trust-via-harness

A worker is trusted not because it is honest, but because the harness around it makes dishonesty unprofitable. Three layers do this. A sandbox bounds what an agent can touch. A fixed honesty check — plain code, not a model — checks the claimed result against the actual artifacts: did the test really run, do the numbers really reconcile, does the diff really do what its summary says. And a top-tier review reads the change before anything merges. The referee is code by design: a model can be argued with; a statistical test and a file checksum cannot. This is the single most important idea in the system — it is what lets the local models be cheap and fallible without putting the result at risk.

2.6 · The local model farm and the crucible

The infantry runs on owned GPUs, and which model fills which role is not decided by reputation — it is decided empirically, in a proving ground called the crucible. Candidate models are run head-to-head on the actual tasks they would perform — coding, structured-output generation, summarization — and promoted to a role only when they beat the incumbent on that job, not on a public benchmark. A model that is faster but produces output that is "different, not better" does not earn the slot. The result is a farm where every role is held by the cheapest model empirically good enough for it, and nothing is trusted on faith.

The crucible has since grown teeth the first edition of this paper did not describe. Batteries now include a synthesis-and-honesty track whose gates hard-fail any run that claims completion without a verifiable artifact — built after a well-reviewed model passed the coding battery while silently no-op'ing a real brief. Diffs are scored by a blind judge that never sees the model's name, alongside repeatable gates, because gates alone once passed a change that had quietly corrupted a query — only the judge caught it. And before committing to new hardware, the same batteries were run on rented cloud GPUs as a pre-purchase gate: prove the larger models clear the bar before a dollar of capital expenditure, the gauntlet philosophy applied to procurement itself — §8.2 records what it rejected. One result is worth publishing for its shape alone: on the system's real coding tasks a dense 9-billion parameter model has kept its seat against challengers roughly three times its size in the local batteries, and against one roughly nine times its size in the cloud bake-off. Bigger was not better; measured was better.

III · The research engine

If Part II is the org chart, this is the assembly line: how a vague idea becomes a tested, promotable strategy — or, far more often, gets killed.

3.1 · Where ideas come from

Candidates are drawn from a deliberately wide net: the published factor literature (momentum, carry, value, quality, trend), classical technical structures, and parameter-variant families generated around a promising seed. The premise is that any single idea is probably worthless — the value is in running a large, cheap search and letting the gauntlet, not a human's taste, decide what survives. Breadth is the point; the filter downstream is strict enough to afford it.

3.2 · The executor loop

Every candidate runs the same five-step loop: propose → isolate → backtest → evaluate → merge or prune. A coordinator proposes the experiment; an executor clones it into a private worktree; the backtest runs against survivorship-free history; the evaluation gauntlet (Part IV) renders a verdict; and only a candidate that clears every gate is merged into the registry. Everything else is pruned and logged with the reason it failed. Because each experiment is isolated, hundreds can run without contaminating one another, and a bad one costs nothing but the compute already spent.

3.3 · The world-event overlay

Price is not the only input, and treating it as the only input is how a systematic process goes blind to the world that moves it. A dedicated world-events intelligence layer sits beside the price engine and runs five stages — acquire, digest, catalogue, signal, overlay — turning public reporting into a continuously-updated, point-in-time interpretation of the world that the research loop can consult. Its product is not a trade; it is a queryable view of what is happening and what it has historically meant.

Acquire. The local infantry — never the expensive models — pulls a free-first menu of public sources: a global event stream, macro and uncertainty indices, financial news, a social firehose, and, on owned GPUs, self-hosted transcription of earnings calls and broadcasts. Pulling data is high-volume, low-judgment work, so it is paid for in cheap local compute, not cloud tokens. Every item is normalized to one record and stamped with the lag between when it happened and when we saw it — so a backtest can later reconstruct what was knowable when, the single most important defense against look-ahead in news data.

Digest. Raw items are read cheapest-capable-first, up the same chain of command as the rest of the system. A fixed, zero-cost layer runs first — a finance-specific sentiment lexicon (negation-aware), entity and ticker linking, and an embedding-based novelty score — and a model is called only when that local gate is uncertain or the item is high-value. Every digested event carries four numbers: relevance, sentiment, surprise (sentiment measured against expectation, not against zero), and novelty (how fresh the item is versus what we have already seen). Those last two are where naive news strategies quietly die: a market reacts to the surprise, not the level, and reactions to stale, recycled news tend to reverse rather than continue.

Catalogue. Interpretations are stored in a four-level, append-only, point-in-time catalogue — raw items, de-duplicated events, scored per-asset impacts, and a single rolled-up regime vector. Nothing is ever overwritten; a correction is a new version with a later timestamp, the same immutability discipline that protects the backtest baselines in Part VII. This is what lets the system ask a question the price-only machine cannot: how did this kind of event move this asset before, and does the live one rhyme with a catalogued precedent?

Signal. A catalogued event becomes a typed feature on a named economic channel — a monetary surprise, a credit-premium shift, a geopolitical-risk spike — never a free-floating number. It is routed by its nature: a genuine fundamental surprise is traded with the drift it creates; an attention spike or a stale headline is faded. And it is held to the same conjunction rule as the gauntlet (Part IV): a signal fires only on several independent confirmations, never on one attractive number. Before any asserted impact is trusted with capital, a top-tier model must confirm it against the source text; if that reviewer is unreachable, the impact is recorded for audit but trusted with nothing. Trust-via-harness, applied to the news.

Overlay. The rolled-up regime vector is injected at each fork's "observe" step, where it shapes what gets proposed and how exposure is scaled — it never overrides the referee. The same signal implies opposite actions under different mandates: a systemic-stress dislocation is a reason for the conservative core to de-risk and, simultaneously, a reason for the ring-fenced high-risk sleeve to lean into its pre-sized, ruin-bounded tranche. A regime-weighted, strictly tranche-bounded meta-allocator turns that into per-fork weights — advisory only, with no real capital moving until the forward gate of Part VII clears.

Two horizons, one discipline. The overlay as built serves the slow horizon: a weeks-to-months regime state that conditions sizing and search. A mid-2026 re-evaluation (§7.5) was candid about what that leaves on the table and what it got wrong. What it leaves on the table is the fast horizon — the hours-to-days reaction window after a specific, timestamped event — for which the acquisition layer is already fast enough but no event-clock machinery yet exists. What it got wrong is subtler and more instructive: the layer asserts, for every catalogued event, an expected direction and horizon — and nothing yet scores those assertions against what prices actually did. The catalogued history is large enough to fix that properly: classical event-study analysis (abnormal returns around the event clock, estimation windows strictly pre-event) run across the archive, so that each event type carries a measured effect size instead of a hypothesized one. That is the overlay's next gate, and it is deliberately the same standard the gauntlet applies to strategies: no signal graduates to the fast horizon on a plausible story — only on measured, fee-clearing reactions. The lesson generalizes and is worth stating as a rule: a predictive layer must be made to grade its own predictions, or it is a narrative layer.

3.4 · Multi-mandate forks

The same engine drives two forks with different appetites: a conservative, yield-seeking core and a ring-fenced, ruin-bounded high-risk sleeve. They share one evidence bar — both must clear the identical gauntlet — and differ only in their drawdown budget and position limits. The high-risk sleeve is walled off so that its worst case is bounded by design and cannot reach the core. Same proof, different tolerance for pain; never the reverse.

IV · The evaluation gauntlet the math

This is the referee. The gauntlet is a fixed combination of separately evaluated statistical tests; a candidate is promoted only if it passes every one. The tests examine different weaknesses, but they share data, implementation, and research history and should not be described as independent guarantees. Live thresholds are withheld here — the dark-forest rule — but every method and its source is shown.

4.1 · Why a conjunction, not a Sharpe ratio

Search enough strategies and some will look brilliant by chance alone. If each backtest is judged at significance level $\alpha$ , the probability that at least one of $N$ independent junk strategies clears the bar is:

\Pr(\text{at least one false pass}) \;=\; 1-(1-\alpha)^{N}

At a 5% level, a hundred coin-flip strategies give a better-than-99% chance of at least one "winner." This is the multiple-comparisons problem, and in quantitative finance it is the dominant way backtests lie (Harvey, Liu & Zhu, 2016). A single number to maximize — the Sharpe ratio most of all — is precisely the knob overfitting turns. The defense is a conjunction: several uncorrelated tests that must agree, each correcting for a different failure mode. You cannot overfit to all of them at once without producing a strategy that is actually real.

4.2 · Monte Carlo Permutation Test — the null of no signal

The first gate asks: could this performance have come from a strategy with no real edge? We answer non-parametrically. The strategy's own returns (or its entry signals) are randomly permuted many times — destroying any genuine timing while preserving the marginal distribution — and the strategy is re-evaluated on each shuffle to build the null distribution of its performance metric $heta$ . The permutation p-value is:

p \;=\; \frac{1+\#\{\, i : \theta_i^{*} \ge \theta_{\text{obs}} \,\}}{N+1}

where $heta_{ ext{obs}}$ is the observed performance, $heta_i^{*}$ the metric on the $i$ -th of $N$ permutations, and the $+1$ in both numerator and denominator counts the observed sample itself — the standard correction that keeps the test valid and never reports $p=0$ (North, Curtis & Sham, 2002; Davison & Hinkley, 1997). A strategy whose edge dissolves under shuffling never had one (Masters, 2018; White, 2000).

4.3 · Deflated Sharpe Ratio — correcting for the search itself

A high Sharpe ratio estimated from a finite, non-normal track record and selected as the best of many trials is biased upward twice over. The Deflated Sharpe Ratio (Bailey & López de Prado, 2014) corrects both. It is the probability that the true Sharpe ratio exceeds a benchmark $mathrm{SR}_0$ , given the record's length $n$ , skewness $hatgamma_3$ , and kurtosis $hatgamma_4$ :

\widehat{\mathrm{DSR}} \;=\; \Phi\!\left[\frac{\bigl(\widehat{\mathrm{SR}}-\mathrm{SR}_0\bigr)\sqrt{n-1}}{\sqrt{\,1-\hat\gamma_3\,\widehat{\mathrm{SR}}+\dfrac{\hat\gamma_4-1}{4}\,\widehat{\mathrm{SR}}^{2}\,}}\right]

where $Phi$ is the standard-normal CDF. The deflation lives in $mathrm{SR}_0$ , the expected maximum Sharpe ratio under the null across $N$ independent trials — it rises with the breadth of the search, so the more strategies you tried, the higher the bar the winner must clear:

\mathrm{SR}_0 \;=\; \sqrt{\operatorname{Var}\bigl(\widehat{\mathrm{SR}}\bigr)}\,\left[(1-\gamma)\,\Phi^{-1}\!\Bigl(1-\tfrac{1}{N}\Bigr)+\gamma\,\Phi^{-1}\!\Bigl(1-\tfrac{1}{N e}\Bigr)\right]

with $gamma$ the Euler–Mascheroni constant and $Phi^{-1}$ the normal quantile function (Bailey & López de Prado, 2014; its single-trial precursor is the Probabilistic Sharpe Ratio, Bailey & López de Prado, 2012). In plain terms: DSR punishes short samples, fat tails, negative skew, and the sheer number of strategies searched — the four ways a Sharpe ratio flatters a backtest.

4.4 · Monte Carlo robustness — the shape of the bad days

Passing on average is not enough; the strategy must survive a plausible worst case. We resample the return stream with a stationary block bootstrap (Politis & Romano, 1994), which preserves serial dependence, and on each resample compute the maximum drawdown:

\mathrm{MDD} \;=\; \max_{t\le T}\frac{\displaystyle\max_{s\le t}V_s-V_t}{\displaystyle\max_{s\le t}V_s}

where $V_t$ is the equity curve. The gate is on the tail of the bootstrapped drawdown distribution and on resilience to dropping a random fraction of trades — a check that no single lucky trade carries the result:

q_{0.95}\!\bigl(\mathrm{MDD}^{*}\bigr)\;\le\;\tau_{\mathrm{DD}} \qquad \frac{\theta(\text{skip }k\%)}{\theta}\;\ge\;\rho

The bound $au_{mathrm{DD}}$ and the retention floor $ho$ are withheld. This is the gate that catches the "great Sharpe, ruinous drawdown" trap a headline number hides.

4.5 · Forward-degradation — backtest is not forward

Backtest performance does not establish what a strategy will retain out of sample. Regime change, competition, costs, selection bias, and implementation error can all produce deterioration. We therefore track the ratio of live-forward performance to backtest performance and require it to stay above a floor, with forward performance strictly positive:

\mathrm{deg}\;=\;1-\frac{\theta_{\text{fwd}}}{\theta_{\text{bt}}}\;\le\;\delta \qquad \theta_{\text{fwd}}>0

A strategy that was excellent in-sample and merely mediocre forward is suspect; one that is negative forward is retired, regardless of how good its history looked. (Part VII describes the bug that once corrupted this very comparison — and why we treated the fix as a first-class result.)

4.6 · Survivorship-free universe construction

All of the above is worthless if the test data is contaminated, and the most common contamination is survivorship bias: backtesting on the companies that exist today silently deletes every firm that failed, and makes almost any long strategy look good. The gauntlet runs on a point-in-time, survivorship-free universe — securities included as they existed on each historical date, delistings and all — so a strategy is graded against the world as it actually was, not a winners-only fiction.

4.7 · Walk-forward and regime-conditioned validation

Parameters are fit on a rolling in-sample window and tested on the next, never-seen out-of-sample window — walk-forward analysis (Pardo, 2008) — so the reported performance is out-of-sample by construction. Results are then sliced by market regime (calm vs. stressed, as defined in Part V) to expose strategies that only work in one weather and quietly bleed in another.

4.8 · A worked example

A representative passage: a diversified, inverse-volatility trend portfolio enters the gauntlet. It clears MCPT (its edge survives shuffling), clears a positive DSR after deflation for the breadth of the search, and — critically — survives the bootstrap drawdown tail where most single-name trend strategies fail. It is promoted to forward testing, where the real verdict is rendered over weeks, not assumed from the backtest. The point of the example is mundane by design: no single number decided it; the conjunction did.

V · Risk, regime & capital

Surviving the gauntlet earns a strategy the right to be considered for capital. What it actually receives depends on the regime it is entering and a strict allocation discipline.

5.1 · Regime detection — an honest thermometer

The system reads the market's "temperature" from a small panel of stress indicators and from a generalized-variance statistic: the square root of the determinant of the asset correlation matrix, $\sqrt{\det \mathbf{R}}$ . As correlations rise toward one — the hallmark of a market in which everything moves together — the determinant collapses toward zero, a compact signal that diversification is evaporating. It is a close cousin of the Absorption Ratio of Kritzman, Li, Page & Rigobon (2011). The crucial caveat, stated plainly: this statistic is coincident, not predictive. It tells you the house is on fire, not that it is about to be. It is used to size and de-risk, never to forecast.

5.2 · Diversification beats regime-timing

The obvious temptation is to switch strategies on the thermometer — go to cash when it reads hot. We tested it directly, look-ahead-free, and the result was unambiguous and against the hypothesis: a regime switcher degraded on every tilt, while an always-on blend across uncorrelated strategies won — better return, shallower drawdowns. Diversification, not timing, is the durable risk control. We report it precisely because it contradicted our prior; that is the standard the rest of the system is held to as well.

5.3 · Position sizing requires calibrated evidence

The legacy design discussed fractional Kelly sizing (Kelly, 1956; Thorp, 2006) while using model confidence as an approximation for win probability. The audit rejected that input as a basis for capital allocation. Fractional sizing limits exposure but does not repair an uncalibrated probability. The replacement will size only from empirical outcomes, subject to independent risk limits. Until that evidence exists, the correct size is zero.

5.4 · Capital discipline

Capital is deployed in tranches, never all at once, and allocated per-venue to proven, forward-validated winners rather than pooled into a single book. The high-risk sleeve is ring-fenced with a hard ruin bound, so its worst case is capped and isolated from the core. Forward testing is free; real money is the scarce resource, released only as out-of-sample evidence accrues. The default posture is to keep most powder dry.

VI · Execution & operations

A verdict is only worth as much as the machinery that acts on it. The execution layer was designed to turn an approved signal into an order and stop safely on failure. The audit later found error paths that continued without required evidence, which is why the current field report treats observed broker state and fail-closed behavior as acceptance requirements.

6.1 · The signal-to-fill path

Approved signals flow through a single path: signal → bridge → executor → venue. The bridge translates a strategy's intent into venue-specific orders; the executor places and reconciles them; and what the broker actually did — not what the system intended — is the source of truth that flows back. Several venues sit behind one interface, so a strategy is written once and routed to wherever it is permitted to trade.

6.2 · The circuit breaker — and a bug we publish

A circuit breaker halts trading when guardrails trip. In review, we found it could fail open: under one error path the breaker could be bypassed rather than halting. We document this deliberately. The correct default for a safety device is fail-closed — when in doubt, stop — and a system that hides the bugs it finds in its own safety logic cannot be trusted with capital. The fix, and the discipline it represents, are part of the record rather than something to bury.

6.3 · Observability

Everything is watchable in real time: per-strategy equity, a cross-strategy overlay that shows how the book moves together, pending and filled orders measured against broker truth, and a per-account audit trail that reconciles the system's view with the venue's. If a number on a dashboard cannot be traced to an underlying fill, it is treated as a defect.

6.4 · Automation and alerting

The pipeline runs unattended on scheduled timers — discovery, evaluation, recording — with no manual feeding. Every state-changing action and every failure pushes a notification, so silence means "nominal" and any exception is surfaced immediately rather than discovered later. Automation is a requirement, not a convenience: a process that needs a human to run it is a process that will be run inconsistently.

6.5 · Data infrastructure

Market, fundamental, and macro data arrive through a thin source shim that presents one internal interface over several free providers — replacing a paid market-data subscription with no loss of coverage for the system's needs. The same layer is where the survivorship-free history of Part IV and a point-in-time fundamentals feed are assembled. Owning this layer keeps both cost and data quality under direct control.

VII · Validation philosophy & intellectual honesty

The hardest engineering in this system is not a clever model; it is the discipline to believe the unflattering result. Three commitments encode that.

7.1 · The forward gate

No strategy reaches real money on the strength of a backtest. It must first survive a multi-week forward test on live data it has never seen, and clear the forward-degradation gate of Part IV. A backtest is a hypothesis; forward performance is the evidence. The gap between them is exactly the dark-forest decay, and the gate exists to measure it rather than wish it away.

7.2 · A postmortem we kept

A data bug once let a forward-grading job overwrite the stored backtest baselines with short, noisy forward numbers — quietly corrupting the very comparison that detects decay. When found, it was written up as a first-class result: the mechanism, the blast radius, the fix, and the guardrail added so it cannot recur. A research system that edits its own history is worse than useless; documenting the failure is how the history stays trustworthy.

7.3 · Why trading is paused

At the time of writing, live trading is paused by choice. The honest reading of the current evidence is that the live track is not yet ready, and only a subset of paper strategies carry genuine forward signal. Pausing is not a failure of the system; it is the system working — the same evidence bar that gates a strategy also gates the decision to risk money at all, and right now it says wait.

7.4 · The crucible

The same empiricism is turned on the tools. Before a model is trusted in a role, it is run head-to-head against the incumbent on the real task in the crucible, and promoted only if it is measurably better — not newer, not larger, better. Several well-reviewed models have been tried and declined for being "different, not better." Trust here is earned with evidence — including the trust placed in the AI that helps build the thing.

7.5 · The mid-2026 re-evaluation — issues found, published

In July 2026 the entire system was re-audited end to end — code, live database, and every evaluation document — with one blunt question: profitable systems exist, so why isn't this one profitable yet? The findings are published here in the same spirit as the circuit-breaker bug (§6.2), because a system that only publishes its architecture and never its failures is advertising, not engineering. The audit found three stacked failure layers, each masking the one beneath it:

The system almost never actually traded. The lone real-broker lane had been silently in rehearsal mode — recording order intent without a single fill — for over a month, and no alarm existed for the condition "funded strategy, zero fills." Meanwhile the regime overlay had read risk-off in every snapshot of its live history, holding the exposure of exactly the strategies built to consume it at zero. The strategies that looked best were the ones that were frozen. A signal that never changes state is indistinguishable from a constant — and a safety posture that fails flat instead of failing loud hides itself.
What did trade, decayed on contact. Forward performance collapsed relative to backtest across every sleeve, every universe, every mechanism — the dark-forest decay of §1.1, measured live and found to be near-total. The audit traced it to structural backtest optimism: curated universes graded on today's survivors, validation folds that never spanned a stressed regime, and a selection objective that rewarded drawdown-avoidance over return — so the process was breeding cautious strategies that survive by not trading.
The feedback loops didn't close. The degradation gate that should retire decayed strategies had never once fired — its ratio test was undefined exactly where it was needed most, when forward performance goes negative. No per-trade ledger existed, so win rates and per-trade economics were unmeasurable. And the world-event layer asserted event-by-event predictions that nothing ever scored (§3.3). Several gates also failed open on missing data or infrastructure error — the same class of bug as §6.2, found again in new places.

The remediation is a specification, not a promise, and its acceptance criteria are concrete: real fills with measured per-trade economics; a regime signal with adaptive thresholds that demonstrably changes state, watched by a stuck-signal alarm; a degradation gate that has actually demoted something; event predictions scored against realized returns; and every gate fail-closed. The honest summary of the audit is the one the first edition of this paper anticipated: the machine was built as a laboratory and behaved like one. The retool's job is to connect the laboratory to the market without lowering the evidence bar it was built around.

7.6 · The supervisor is a language model too

The re-evaluation surfaced a harder, more valuable finding than any single bug: several of the issues it found had been documented by the very model that failed to fix them. Across many sessions, a top-tier model had been asked, repeatedly and directly, to find and fix the system's shortcomings. It found some. It wrote fix specifications, handoff notes, and remediation plans. And weeks later, documented one-line fixes remained unapplied, a safety gate that had never fired remained trusted, and the flat-line conditions of §7.5 sat unalarmed. Why does the most capable tier describe problems it does not resolve?

Because a language model produces probable answers, not guaranteed identical ones — it samples a plausible completion of a request, and producing the artifact of a fix (a spec, a plan, a diff, a confident summary) is a perfectly plausible completion that ends the task successfully. Whether the world actually changed is a separate fact the model never observes unless the harness forces it to. Three corollaries did the damage. Absence blindness: a model reviewing a system attends to what is in front of it, and the gravest defects were things that did not exist — missing alarms, missing ledgers, predictions no code ever scored — which generate nothing to attend to. Session locality: each working session sees a slice, fixes the slice, and writes an optimistic summary that becomes the next session's inherited "truth," laundering hope into context. Builder bias: the model that builds a subsystem grades it generously — the exact reason strategies were never allowed to grade themselves here, applied one level higher than it had been.

The remedy was already in this paper; it had simply not been applied to its own author. Trust-via-harness extends to the top of the hierarchy. In the retool: no fix is "done" on a narrative — done means a fixed check ran against the deployed system and its output is in the record. A health-check makes the absences visible mechanically — invariant violations printed as facts into every session's context, so what is missing finally generates tokens. Audits are run by a role with no authorship stake, mandated to report numbers and explicit absences, never summaries. Requests for fixes are phrased as falsifiable acceptance criteria with the verifying command attached. And every weak point any re-evaluation finds is converted into a permanent, machine-checked invariant before it is considered closed — a ratchet, so the floor only rises. The general law, and the single most transferable result this project has produced: repeatability lives in fixed checks, not in the model—including at the tier that builds the system. Repeatable does not mean correct; the inputs and acceptance rules still require review.

7.7 · The verification ladder — confidence, cross-model review, and the honest boundary

Two natural objections to §7.6 deserve direct answers, because both are half-right. "Just use the model's confidence level" — confidence is real but treacherous. Locally-served models expose token log-probabilities, a genuine, free signal of the model's own uncertainty, and the harness uses it: a low-margin answer is routed up the escalation ladder automatically. But the most capable hosted models expose no such internals, leaving only verbalized confidence — "I'm 85% sure" — which is itself a sampled output, and systematically overconfident. So the rule is calibration before use: a stated confidence becomes a usable probability only after it has been mapped against realized outcomes over time, per agent, per task — the same empirical-calibration discipline the sizing layer (Part V) demands. Until that mapping exists, a confidence number is displayed as a claim and never sized on, never gated on. An uncalibrated confidence presented as fact is precisely the failure §7.6 documents.

"Have models review each other's output" — yes, and the system does; but review by another probabilistic process does not turn opinion into repeatable proof, and pretending it does would rebuild the original problem one level up. The design is a four-rung verification ladder, and the order is the point. Rung zero: fixed checks — exit codes, artifact existence, checksums, does-the-diff-match-the-narration — the layer whose result should repeat from the same input; nothing above it may overrule it. Rung one: a local skeptic — a cheap owned model that mechanically verifies claims against artifacts for the output of every tier, the top included. A small model cannot judge architecture — and does not need to: catching a fabricated or unverified completion is a verification task, not an evaluation task, and it is the cheapest insurance the system buys. This is a finding worth publishing on its own: routinely pointing small local models at the most capable model's claims catches low-confidence output presented as certainty — the exact failure mode capability does not exempt. Rung two: a blind panel — for state-changing paths, multiple diverse reviewers, author identity withheld, greedy decoding and pinned builds for reproducibility, with disagreement escalated, never silently averaged. Rung three: the empirical scoreboard — fills, returns, realized event reactions — the final arbiter, which also calibrates the reviewers themselves. The honest boundary, stated so it cannot be quietly forgotten: rungs one and two lower variance and catch fabrication; rung zero adds repeatability; rung three reports what actually happened. None of them makes a flawed input or acceptance rule correct. The discipline in one sentence: models verify downward against artifacts, never sideways against opinions.

VIII · Roadmap

The architecture in this paper is the direction of travel, not a finished monument. The hardware expansion in §8.2 is complete; model qualification and the system retool remain active.

8.1 · The retool

The orchestration layer (Part II) is being rebuilt around a dedicated engine that drives the existing, proven gauntlet — the referee does not change, only the machine that feeds it. The two-mandate forks and the world-events intelligence layer are now first-class parts of the loop rather than bolt-ons — the overlay's acquire→digest→catalogue→signal pipeline (§3.3) is built and running, feeding the same observe step the forks read; what remains is to widen its paid-source coverage and let its signals accrue the forward evidence every other strategy must. The mid-2026 re-evaluation (§7.5) sharpened the retool's sequence: close the execution and measurement loops first — real fills, a per-trade ledger, fail-closed gates, a regime signal that can change state — so that the expanded compute below lands on a system that can convert intelligence into measured results rather than amplifying a laboratory.

That sequencing question was then settled by an explicit decision rather than drift: build forward, don't patch back. A follow-up review compared the system against standard datacenter operational practice — heartbeat ledgers, one centralized error stream, dependency-liveness probes, alerting on absence, a fleet manifest, restore drills — and found the statistical machinery years ahead of the operational plumbing beneath it. Since the audited defects in the legacy loops are architectural, and nothing trades real capital, patching them buys nothing durable. So the legacy strategy machinery is deliberately paused (data acquisition, observation, and backups keep running — the event archive and calibration history are inputs the next generation needs), and those datacenter-grade practices are being woven into the new implementation as birth requirements — every new component heartbeats, logs to one stream, probes its dependencies, and registers in the fleet manifest from its first commit. Retrofitting operational discipline is the expensive way to acquire it; this system gets to do it the cheap way, once.

8.2 · The compute expansion — built, now earning its work

The owned tier now spans two complementary hosts. A many-lane server carries four 32 GB datacenter accelerators for a 128 GB pool. A fifth 32 GB accelerator lives on an external PCIe dock connected by OCuLink. It can be cabled to that server when a job needs a 160 GB V100 pool, or to the 24 GB RTX 3090 workstation when the smaller host needs a 56 GB mixed pool. This is deliberate placement, not an imaginary seamless cluster: the resource plane must record where the movable card is, and the control plane can route work only to the capability available there. Every part was chosen the same way models are — by measured value per dollar, with the money concentrated in the accelerators and deliberately starved everywhere performance does not pay. True to §2.6, the candidate models faced the crucible before the hardware was ordered, on rented cloud GPUs — a pre-purchase bake-off whose most valuable output was a rejection. A heavyweight coding challenger roughly nine times the incumbent's size scored 91% of gates against a configured minimum of 90% — it cleared the threshold, and then lost to the 9B incumbent's 93%. That distinction is the whole point: in a purchase decision the bar that matters is not the configured threshold, it is the incumbent. On the honesty track the largest candidate in the field scored 93% with zero hallucinated completions and was beaten by a mid-sized thinking model that took every run at 100%; the same large model was the one that fell below the bar on the coding battery, at 89%. Capability did not predict honesty, and size predicted neither. The 9B kept its seat. Seats change hands only on evidence, and the hardware is sized for what measured, not what marketed.

More powerful models are now being tested on both hosts. The next qualification workload is deliberately real rather than synthetic: daily programming, where local output has to survive the same builds, tests, reviews, and running-system evidence as every other contribution. The second is improvement of AI TraceVector Alpha itself — using the additional local capacity to widen research, challenge, and refinement while leaving promotion with the fixed gauntlet. Beyond those first two workloads, the new power maps to loops the re-evaluation found open: parallel crucible batteries with enough repetitions for real statistics instead of ties; the world-event digest read continuously by a stronger local model at zero marginal cost; the event-study backfill of §3.3; self-hosted transcription of earnings calls; and nightly batch analysts large enough to review whole strategy books. Compute amplifies whatever loop it is pointed at — pointed at closed loops, it compounds; pointed at open ones, it just makes the laboratory busier. The sequencing in §8.1 exists so it is the former.

8.3 · Model-independence and the token economy

Two design rules now govern everything the retool builds. The first is a floor requirement: the finished system must be operable by any competent model — the most capable tiers are its architects and escalation reviewers, never its required operators. That is partly economics (the cost curve of §2.2) and partly robustness: a system that only its smartest contributor can run has a single point of failure with an API bill. The implication is that intelligence migrates into the artifacts: runbooks become scripts with falsifiable pass/fail, session context comes from a health-check that prints the system's invariant violations as facts, and every moment that required operator judgment is converted, as it is discovered, into a gate, a hook, or a check. The proof is a recurring weak-operator drill — a small local model runs the standard maintenance procedure end to end, and every gap where it needed judgment the artifacts didn't encode becomes the next artifact. The drill passing is the definition of done.

The second rule is the token economy: digest once into a typed store; query cheap forever. Every subsystem that survives here shares the shape — events into a typed, versioned catalogue rather than re-read text; code relationships into a queryable graph rather than files re-fed into context; session knowledge into structured, scoped memory rather than transcripts; standing directives compiled into fixed hooks that cost zero tokens at runtime; alerts deduplicated and typed rather than re-narrated per run. A workflow caught re-reading raw material in a loop is treated as a defect, not a habit. Together the two rules are one idea seen from two sides: the model is a replaceable component; the structure — typed stores, gates, invariants, and the ladder of §7.7 — is the product.

8.4 · Open decisions and known limits

Several questions are deliberately still open: the budget and policy for escalating to paid model tiers; the exact risk bounds on the high-risk fork; which single paid data source, if any, is worth its monthly cost; and the calibration-driven sizing upgrade of Part V. These are listed, not hidden — a roadmap that admits what it has not settled is more useful than one that pretends everything is decided.

8.5 · Where the engineering goes next

The through-line is unchanged from the first principle: widen the search, keep acceptance outside the model path, and let real capital follow only after forward evidence and an explicit human release decision. The dark forest does not get friendlier; the discipline is to keep earning the right to be in it.

Appendices

A · Glossary

Alpha — risk-adjusted return in excess of the market; the edge.
Conjunction gate — a promotion rule requiring every test to pass, not a single score to be maximized.
Drawdown — the peak-to-trough decline of an equity curve.
Mandate / fork — a risk profile (conservative core vs. ring-fenced high-risk sleeve) the one engine serves under a single evidence bar.
Regime — a market state (e.g. calm vs. stressed) used to condition validation and sizing.
Survivorship bias — the distortion from testing only on assets that still exist today.
Trust-via-harness — trusting a worker because the surrounding sandbox, fixed check, and review make dishonesty unprofitable.
Walk-forward — fitting on a rolling in-sample window and testing on the next, unseen window.

B · References & citations

The conceptual framing and every formula in Part IV are tied to their published sources. The methods are public; the live parameters are not. Liu Cixin's Dark Forest is used as a literary operating metaphor, not as evidence about market behavior.

Three-body dynamics — Stern, D. P., "The Lagrangian Points L4 and L5", NASA archival educational material. The page distinguishes specific equilibria from the chaotic properties of general three-body motion.
Adaptive markets — Lo, A. W. (2004), "The Adaptive Markets Hypothesis: Market Efficiency from an Evolutionary Perspective", Journal of Portfolio Management 30, pp. 15–29.
Multiple testing / data-snooping — White, H. (2000), "A Reality Check for Data Snooping," Econometrica 68(5); Harvey, C. R., Liu, Y. & Zhu, H. (2016), "…and the Cross-Section of Expected Returns," Review of Financial Studies 29(1); Harvey, C. R. & Liu, Y. (2015), "Backtesting," Journal of Portfolio Management 42(1).
Permutation testing — Masters, T. (2018), Permutation and Randomization Tests for Trading System Development; North, B. V., Curtis, D. & Sham, P. C. (2002), "A Note on the Calculation of Empirical P-Values from Monte Carlo Procedures," American Journal of Human Genetics 71(2); Davison, A. C. & Hinkley, D. V. (1997), Bootstrap Methods and their Application, Cambridge.
Deflated & Probabilistic Sharpe Ratio — Bailey, D. H. & López de Prado, M. (2014), "The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality," Journal of Portfolio Management 40(5); Bailey, D. H. & López de Prado, M. (2012), "The Sharpe Ratio Efficient Frontier," Journal of Risk 15(2).
Bootstrap (serial dependence) — Politis, D. N. & Romano, J. P. (1994), "The Stationary Bootstrap," Journal of the American Statistical Association 89(428); Künsch, H. R. (1989), "The Jackknife and the Bootstrap for General Stationary Observations," Annals of Statistics 17(3); Efron, B. (1979), "Bootstrap Methods: Another Look at the Jackknife," Annals of Statistics 7(1).
Walk-forward analysis — Pardo, R. (2008), The Evaluation and Optimization of Trading Strategies, 2nd ed., Wiley.
Regime / systemic risk — Kritzman, M., Li, Y., Page, S. & Rigobon, R. (2011), "Principal Components as a Measure of Systemic Risk," Financial Analysts Journal 67(4).
Position sizing — Kelly, J. L. (1956), "A New Interpretation of Information Rate," Bell System Technical Journal 35(4); Thorp, E. O. (2006), "The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market," in Handbook of Asset and Liability Management.
Event-driven signals (overlay) — Loughran, T. & McDonald, B. (2011), "When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks," Journal of Finance 66(1); Tetlock, P. C. (2011), "All the News That's Fit to Reprint: Do Investors React to Stale Information?," Review of Financial Studies 24(5); Baker, S. R., Bloom, N. & Davis, S. J. (2016), "Measuring Economic Policy Uncertainty," Quarterly Journal of Economics 131(4); Caldara, D. & Iacoviello, M. (2022), "Measuring Geopolitical Risk," American Economic Review 112(4); Gilchrist, S. & Zakrajšek, E. (2012), "Credit Spreads and Business Cycle Fluctuations," American Economic Review 102(4).

C · Model-tier & cost economics

Compute is tiered cheapest-capable-first, with every call metered through one routing endpoint (Part II). The principle — spend proportional to decision value — is shown; the dollar figures are withheld.

Tier	Role	Used for
T0 · local farm	Infantry	Bulk exploration, code, data, backtests — the vast majority of calls.
T1 · subscription	NCOs / review	Harder reasoning and merge review on a fixed-cost plan.
T2 · metered cloud	The general	State-changing decisions only, under budget — rarely, and on purpose.

D · Hardware / farm inventory

The local farm runs on an RTX 3090 and five refurbished 32 GB datacenter accelerators: four in the dedicated server and one on the two-host OCuLink dock described in §8.2. Model roles are assigned by the crucible rather than by spec sheet, and both hosts run one unified software interface. The public build record is documented on The Farm; deployment details and operating credentials remain closed. The salient fact is that the system's day-to-day cognition runs on hardware that is already paid for.

E · Data-source catalog

Price, fundamental, and macro data are assembled from free and low-cost providers behind the source shim (Part VI), including a survivorship-free historical universe for backtesting and a point-in-time fundamentals feed. Specific providers and keys are not published.

F · System-evolution timeline

The system has moved from single-strategy experiments, through a portfolio-of-uncorrelated-strategies discipline and a survivorship-free data foundation, to the multi-agent retool described here — and, in mid-2026, through a published full-system re-evaluation (§7.5) that reset the retool's priorities and launched the compute expansion (§8.2). The constant across every stage has been the referee: a fixed evidence bar that the rest of the architecture exists to feed.