For-profit corporations exist to maximize shareholder value subject to remaining solvent [13, 14], and shareholder value is, by construction, shareholders’ equity—the residual claim on assets after liabilities are netted. The firm’s per-period reward is the realized log-return on that equity [18], the survival constraint is $K_{\tau }>0$, and the cumulative objective $J_t$ is the expected sum of those per-period rewards over a finite planning horizon. All later channels are priced against this $J_t$.

To execute its optimization the corporation maintains an inventory of operational components—the channels through which it spends capital, generates the cash flows that keep it solvent, and earns capital back. Five channels suffice: what the firm holds ($\mathbf{I}$, portfolio), what it can see ($\mathbf{S}$, sensors), what it can do ($\mathbf{U}$, actuators), how it learns ($\mathbf{Z}$, R&D), and what it knows ($\mathbf{\Theta }$, parameters). We adopt a lossy five-channel partition that suffices for the quant-trading instance and that we will argue extends to the general case. The framework only requires that each row can be priced in dollars and that channel-specific scaling laws are estimable from the firm’s history.

The action vector ${\mathbf{a}}_t$ partitions the cycle-$t$ change in the corporation into the same five channels as the state; each entry $a_t^k$ is the dollar change in channel $k$ during the cycle (Action vector). The symbol-by-symbol definitions of the five channels are given via the projections in Corporation tuple and Action vector.

The environment ${\mathbf{E}}_t$ is the part of the world outside the corporation that matters for the next cycle: prices, order flow, liquidity, counterparties, macro state, regulation, and the news process. Coupled dynamics states how the corporation and that environment move together after the firm acts.

Every operating cycle appends one row of observations to each channel history ${\mathcal{H}}_t^k$. The sensor channel can additionally widen by adding new data: Wayback Machine snapshots, exchange filings, alternative-data archives. Each turn provides the next iteration of training data.

Once enough rows accumulate, accurate predictions can be made on a channel-by-channel basis, allowing the firm to predict its own self-improvement dynamics. In theory, a unified world model would be trainable with enough data such that the firm could jointly predict its own state, improvements to it, and the state of the external economy all at the same time.

Concretely: at decision time $t$ the controller queries the world model for ${\hat{\mathbf{g}}}_t$ off the trained row laws (predict), solves the convex inner program of Portfolio Optimization to pick the best action (optimize), executes the chosen ${\mathbf{a}}_t^{\star }$, and the realized $({\mathbf{\Xi }}_{t+1},{\mathbf{E}}_{t+1},R_t)$ becomes the new training row that tightens the next cycle’s forecasts.

Each improvement attempt teaches the firm something about itself. The action it took, the internal change it produced, and the reward that followed become a new row in the channel histories. As those rows accumulate, the firm gets sharper predictions about which future improvements will work and tighter confidence intervals around the value of its own next actions. The loop is the thesis of the paper compressed into one diagram; everything that follows is one of these arrows worked out in detail.

The EWM is the conditional next-cycle law the firm uses to forecast the joint $({\mathbf{\Xi }}_{t+1},{\mathbf{E}}_{t+1},R_t)$ given its current information set ${\mathcal{F}}_t$ and a candidate action ${\mathbf{a}}_t$. It is the only model of the future the firm has access to, so improving it is itself an allocation with first-order effect on $J_t$.

Because the firm sees the world only through its sensors, both ${\mathbf{E}}_t$ and its own state ${\mathbf{\Xi }}_t$ enter the EWM as posteriors over noisy observations rather than as latent ground truth. The firm history ${\mathcal{H}}_t$ is the chronological log of every observation through cycle $t$—the canonical sufficient statistic in the partially-observed Markov decision process tradition [26, 27, 28, 29]. The channel histories ${\mathcal{H}}_t^k$ slice it channel by channel into $({\mathbf{o}}_{\tau }^k,a_{\tau }^k,R_{\tau +1})$ rows. These per-channel slices are the tables the firm uses to fit the row laws of the marginal-return vector.

The filtration requirement is what separates an Economic World Model from an ordinary language model. At decision time $t$, an EWM may condition only on ${\mathcal{F}}_t$ and the candidate action ${\mathbf{a}}_t$. A language model trained on a static corpus [30] can mix documents from before and after the event it is asked to predict, so its context can contain future information. The same no-peeking discipline that separates a forecasting model from a memorizing one has been the organizing principle of game-playing AI [31, 22, 21, 23], and the loss functions below make the difference explicit. For next-token modeling on a fixed snapshot, this is fine. As the primary EWM ${\hat{\mathcal{W}}}_t$ without an externally imposed filtration discipline, it is not: held-out validation can only enforce filtration discipline if the holdout set is chronologically after the entire training corpus—a window that shrinks as internet-scale models train on ever more recent data. This results in very little data left for validation, and even less for robust benchmarking. Consequently, measured performance uncertainty remains high, and the small post-corpus window provides negligible information about how the model adapts to changing market regimes. The outcome is slower compounding of capital and structurally riskier decisions. A general LLM that is wrapped in a strict post-cutoff evaluation harness can still serve as a component or proposal mechanism; the categorical claim is about the bare model, not the wrapped system.

In principle the firm has one joint EWM ${\hat{\mathcal{W}}}_t$ over the entire corporation–environment state. In practice that joint object is approximated by a collection of channel-specific world models, each trained on its own channel history ${\mathcal{H}}_t^k$. These per-channel transition models are usually much simpler than a full simulator—a scaling law, a market-impact curve, a refit-decay model, or a search law is already a channel-world-model fragment. The split is a practical approximation, not a claim that the channels evolve independently; cross-channel coupling re-enters when the controller composes the rows of ${\mathbf{g}}_t$ below.

Stitching together the cumulative objective and the firm’s solvency, liquidity, and budget constraints gives the corporate program: the controller $\mathcal{G}$ picks the action that maximizes $J_t$ subject to the constraints in the appendix. The full constraint set is collected in Program Constraints so the body can stay focused on the convex inner program.

Differentiating the common objective $J_t$ with respect to the capital-allocation vector turns heterogeneous interventions—a researcher hire, a new alternative-data feed, a GPU, a position in AAPL, an execution-latency upgrade—into directly comparable marginal rates. Each entry of ${\mathbf{g}}_t$ is the expected log-equity growth per marginal dollar invested in that channel; at the optimum every funded channel equates the same risk-adjusted shadow price of capital, ${\hat{g}}_t^k/{\sigma }_t^k={\lambda }_{S,t}^{\ast }$ (Portfolio Optimization), which collapses to the bare equimarginal identity ${\hat{g}}_t^k={\lambda }_t^{\ast }$ only in the risk-neutral limit (${\kappa }_t\to 0$, or per-channel dispersions ${\sigma }_t^k\to 0$). The risk functional itself is general: Trsi Net writes the full corporate t-RSI in terms of any auditable uncertainty functional ${\mathcal{U}}_t$, of which standard deviation is just the operational default. The next section instantiates this row by row for the quant firm.

The horizon $T$ introduces uncertainty into this pricing: the variance of cumulative reward ${\sum }_{\tau }{R}_{\tau }$ propagates through the rollout. $T$ is therefore set at the point where additional cycles no longer meaningfully tighten the estimate once that propagating uncertainty is accounted for. This uncertainty is fundamental to the controller’s allocation problem: an honest pricing of capital must trade off the central tendency of forecasted returns against their dispersion. The most general statement of this is expected-utility portfolio optimization, which evaluates allocations under whatever utility functional the firm chooses over the joint distribution of channel returns. Under the simplest case of normally distributed returns this collapses to mean-variance portfolio optimization [32] and the per-channel Sharpe ratio ${\hat{g}}_t^k/{\sigma }_t^k$ [33]—the form we use throughout the rest of the paper for simplicity; in production, more complex utility-of-distribution objectives are often preferred.

Portfolio Optimization details why this uncertainty necessitates the risk-adjusted form not only for the investment channel, but for the performance and optimal allocation of every channel.

Why t-RSI is the thing to measure.

The differentiable-corporation thesis says the firm carries two posteriors at every cycle: one over alpha created per dollar spent on each channel, and one over alpha decayed from the deployed book. A controller that compounds is one that commits capital only when those two distributions are confidently separated — it believes it will create more alpha next cycle than the alpha already on the books erodes, and it believes it by enough margin that the call is unlikely to be noise. We call that standardized separation the improvement signal-to-noise ratio, written t-RSI by analogy with a two-sample $t$-statistic. The two distributions are not draws from one underlying process: they are posteriors over two different processes (create from the channel-row fits below, decay from the firm’s forecast-evaluation and live-trading panel). t-RSI is therefore framed as a standardized distance, not a hypothesis-test instrument.

A positive t-RSI of, e.g., $2.5$ means the create-rate posterior is centered $2.5$ pooled standard errors above the decay-rate posterior. The firm’s controller does not commit capital to a candidate until that statistic clears a sealed-evaluator threshold (Certificate of monotone improvement); that gate is what makes compounding survive selection, and what distinguishes a self-improving corporation from one that promotes drift on noise. Equivalently, every row of the channel table below contributes one slice of the blue density, and the cycle’s investment decision is a draw from which slice the controller will pay for next.

Headline definition.

Write ${\widehat{\Delta {\alpha }^{\mathrm{create}}}}_{t:H}$ for the firm’s posterior alpha creation over horizon $H$ (the sum of per-channel contributions, with bookkeeping fixed in t-RSI Measurement Conventions), and ${\widehat{\Delta {\alpha }^{\mathrm{decay}}}}_{t:H}$ for the matching alpha-decay posterior. t-RSI is the standardized distance between them:

The channel-by-channel decomposition of ${\widehat{\Delta {\alpha }^{\mathrm{create}}}}_{t:H}$ as a path integral along the planned allocation path, and the bootstrap propagation that supplies each SE, live in t-RSI Measurement Conventions; the operational walk-through of the headline three-month calculation lives in Three-Month t-RSI Calculation.

What the data presently say.

Empirically the firm’s own panel measures the decay rate as small and near zero. The data-derived headline reads ${\hat{\lambda }}^{\mathrm{decay}}$ from a per-asset alpha-decay estimator that fits the held-out forecast edge against deployment age once per asset and aggregates the resulting per-asset rate distribution robustly across the held-out asset universe (median + MAD/IQR summary; SE combines within-asset-cell dispersion with between-cell variability over training-run seeds and forecast horizons). The same qualitative picture holds in production: across $16$ months of cumulative Mark II and Mark III live trading, every linear, rank, and distributional trend statistic against deployment age returns $\hat{\kappa }$ near zero. The data-derived headline t-RSI is therefore dominated by create-side dispersion, not by decay. The subsections that follow estimate the per-channel create-rate distribution one channel at a time; Headline t-RSI composes them into the headline t-RSI for the current operating point.

What this channel is. Investments are the firm’s current production channel: capital positioned in the trading book and rebalanced by the controller. The EWM slice for this channel forecasts price-change distributions. Construction, scaling laws, and held-out benchmarks are deferred to a forthcoming companion paper; the empirical claims in this paper are channel-decomposed and do not depend on the companion-paper architecture. The controller turns those forecasts into a rebalance; the market function turns the rebalance into realized cash. This is the only row whose one-cycle return is observed directly from the broker ledger rather than inferred counterfactually [38, 39, 40, 41, 42, 43, 44, 45, 46].

What this means. The investments row tells the controller what one extra dollar of trading capital is worth right now: forecast edge ${\hat{\mu }}_t$ on the candidate trade, minus the learned execution friction ${\hat{\varphi }}_t$ paid at execution. Of the three terms in Investment marginal return, the EWM forecast and the size dependence of ${\hat{\varphi }}_t$ are estimable from a backtest; the remainder of ${\hat{\varphi }}_t$ is not, and is the proprietary surface the rest of this subsection explains. Two of the three terms are public; one is not, and that asymmetry is why the live track record matters more than the backtest figure for this row.

Backtest.

$\varphi$ deserves special emphasis: it mostly cannot be estimated from a backtest. Market impact admits a square-root-law approximation (Investments Supporting Equations), but the remainder—routing quality, venue- and broker-specific spreads, financing, adversarial response—is venue-, instrument-, size-, and time-of-day-specific and can only be learned by trading. The firm has executed approximately $400M of trades; that volume is the data behind its internal $\varphi$ surface, which is proprietary for exactly the reason Investment marginal return makes plain—$\varphi$ is the term the rest of the world cannot buy. The dated three-deployment-generation track record (Mark I/II/III) is reported in Deployment Parameters.

What this channel is. Sensors are purchases that enlarge the filtration ${\mathcal{F}}_t$: market-data feeds, historical records, finer sampling, and any measurement that lets the EWM condition on more of the world. The empirical primitive is the data-scaling law, fit on the Muennighoff effective-data axis and the Hoffmann/Kaplan loss form [30, 9, 47, 48], with the underlying unit being the dollar-weighted bar in the spirit of information-driven sampling [49, 50, 51]. The body keeps only the local loss slope with respect to effective data; the fit protocol, prior choices, and controller-level chain rule are collected in Sensors Supporting Equations.

What this means. The fitted exponent says how much predictive error remains after the firm buys another decade of effective data. At the reported operating point, ${\alpha }_{D_{\mathrm{eff}}}\approx 0.16$ means a $10\times$ increase in effective dollar-weighted tokens multiplies the reducible part of the loss by roughly $10^{-{\alpha }_{D_{\mathrm{eff}}}}$, or about $0.7$: the model keeps about 70% of the reducible error and removes about 30%. The interval around ${\alpha }_{D_{\mathrm{eff}}}$ is the uncertainty on that local data slope, not on trading alpha. $R^{\star }$ says when repeated passes over the same data stop behaving like fresh information, and ${\mathcal{L}}_{\mathrm{noise}}$ is the residual floor that more sensors cannot remove at the current model size and architecture; R&D moves architecture and therefore moves this floor (see Sensors Supporting Equations). This is why the chart is a sensor asset measurement: it prices data by the remaining predictive loss it can still reduce.

We expect this scaling law to continue, as scaling laws are some of the most well-established results across a variety of domains; as it turns out, our preliminary evidence suggests that quantitative trading demonstrates scaling laws as well. What the chart is actually pricing (effective dollar-weighted tokens, not number of predictive factors), the realistic order-of-magnitude scaling factors available against the current operating point, and the cross-domain evidence for power-law extrapolation are collected in Data Scaling.

What this channel is. Actuators are the interfaces through which the controller can act. In the current quant-trading envelope, the measured actuator is the tradable asset universe: as the universe expands, the firm both trains on more dollar-weighted histories and gains more instruments through which forecasts can be deployed. The body keeps only the local data-performance slopes measured by that panel; the fit protocol, joint-path interpretation, and general actuator row $g_t^{\mathbf{U}}$ are collected in Actuators Supporting Equations.

What this means. The actuator row asks whether expanding the tradable surface changes what the investment production function can do. The two slopes above are empirical between-$N$ regressions: realized annualized return and realized annualized Sharpe per decade of effective dollar-weighted tokens. In this trading-envelope instance, the asset universe and the dollar-weighted data universe expand together, so the fitted slopes are read directly as local data-performance slopes rather than as separately identified data-only and actuator-only effects. The fitted slopes, their 95% confidence intervals, and the $T_A/T_B$ extrapolation rows in the table above are all read directly off the actuator panel and the same OLS bookkeeping that draws the 95% prediction band on the chart. A note on sample size: the multi-seed sweep produces 15 universe sizes $\times$ 3 seeds $=45$ raw runs that anchor the sensor-row data-scaling fit, but the actuator-panel regression takes cluster medians across the three seeds at each universe size, so the $n=12$ reported in the parameter table is the 12-cluster-median count after dropping the three smallest universe sizes for which the per-seed median has insufficient evaluation breadth; the underlying 36-row $(N,\text{seed})$ panel is the object the cluster-by-$N$ bootstrap of Three-Month t-RSI Calculation resamples. The parameter table’s $n$ column reports the regression degrees-of-freedom unit (cluster medians), not the raw run count.

Potential scale.

We estimate the actuator-row scaling law could be extended over two tranches of source data that still satisfy the dollar-weighted-token construction of Three-Month t-RSI Calculation (dollar-denominated, strict time filtration); the corresponding $T_A$ and $T_B$ extrapolated Sharpe and annualized return are reported in the parameter table above. Tier $T_B$ is included because the underlying data class satisfies the same empirically measured dollar-weighted-token line fit that anchors Tier $T_A$.

• Tier $T_A$ – pure asset-price data: the catalogue of asset prices, quotes, trades, and order-book snapshots already commercially available across data brokers (global equities, futures, FX, rates, options, crypto, OTC fixings).

• Tier $T_B$ – broader dollar-denominated time-series data: any other dollar-denominated series with a strict time index that respects the filtration (card panels, payments and settlement flows, insurance-claims tapes, payroll feeds, satellite- and geolocation-derived activity counts, public budget and procurement records), not restricted to tradable asset prices.

What this channel is. R&D is purchases that improve future search cycles: researcher labor, LLM API spend, GPU-hours, evaluation infrastructure, and search tooling. The empirical primitive the body reports is the derivative of the selected rolling upper-tail held-out frontier with respect to completed auto-research experiments, fit on the same axis for two metrics simultaneously — annualized Sharpe ratio and annualized return. The whitepaper manifest pins the headline frontier to the top-10% rolling upper tail for both Sharpe and annualized return, while the appendix reports the surrounding cutoff sensitivity scan. The held-out window is the canonical $1258$-day validation split each campaign run reports against the $2516$-day training period (see Channel Derivations); the §5.3/§5.4 sensor and actuator panels are evaluated on their own held-out splits as documented in Data Scaling. The body keeps only those two local derivatives; the dollar-to-experiment production function, the human/LLM allocation split, the architecture-quality scalar ${\Gamma }_t$, the search-scaling law ${\Gamma }_t=\xi {\log }_{10}(1+N_t^{\exp }/N_0)$, the chain rule $g_t^{\mathbf{Z}}$, and the R&D transition law are collected in Channel Derivations.

What this means. The two fitted derivatives say what a ten-fold increase in completed auto-research experiments buys on the selected held-out frontier: $\partial \mathrm{Sharpe}/\partial {\log }_{10}(1+n^{\exp })$ Sharpe ratio points and $\partial \mathrm{AnnReturn}/\partial {\log }_{10}(1+n^{\exp })$ percentage points of annualized return. Each is the headline R&D primitive in the metric the firm actually optimizes against. The fitted curves are read off the report-selected top-10% rolling upper-tail frontiers of the strict-invalid- and sealed-holdout-filtered cohort; the running-best step remains in the figures only as muted context. Starting from the current campaign count $n_{\mathrm{current}}^{\exp }$ and projecting forward to $n_{\mathrm{current}}^{\exp }+\Delta n$ completed experiments, the additional Sharpe purchased is $\frac{\partial \mathrm{Sharpe}}{\partial {\log }_{10}(1+n)}[{\log }_{10}(1+n_{\mathrm{current}}^{\exp }+\Delta n)-{\log }_{10}(1+n_{\mathrm{current}}^{\exp })]$, and the additional annualized return is the analogous local increment with $\partial \mathrm{AnnReturn}/\partial {\log }_{10}(1+n)$. This is how the R&D row enters the t-RSI numerator: a marginal R&D dollar is priced through how many additional experiments it buys (Channel Derivations) and how much each completed experiment is currently shifting the selected frontier along the fitted curve.

Selection vs. structural scaling.

Across 929 auto-research experiments, the rolling top-10% frontier of held-out Sharpe rises log-linearly with completed-experiment count (slope $\approx 0.34$ per decade). We note this is a selection statistic rather than a structural scaling law: the top-$k\%$ mean of $n$ samples from a stationary distribution would also grow with $n$. The substantive question is whether the underlying distribution is shifting — for which the running median (rather than the top-10%) is the cleaner diagnostic, and we treat the present fit as a directional finding to be confirmed against that diagnostic in subsequent campaigns.

Why this fit, given that caveat.

The empirical methodology for studying auto-research is itself new: the only large-scale precedents we are aware of are the open-source single-GPU agent loop catalogued by [52] and the autonomous-R&D-capability evaluation harness of [53], neither of which prescribes a settled frontier-estimation protocol. Within that gap we have chosen a rolling top-quantile with a bootstrap confidence band rather than the more common running-best step. Running-best is a strict order statistic over an ever-growing sample: it is monotone in $n$ by construction and is exactly the selection-bias regime that the deflated-Sharpe / backtest-overfitting literature documents as untrustworthy ([54, 51]). A rolling top-$k\%$ tail mean is a smoother, less brittle envelope of the same upper-tail behavior, and reporting a bootstrap CI alongside it surfaces the residual selection inflation directly. We do not claim this estimator is the right answer for auto-research scaling — only that it is a more defensible reading of the frontier than the running-best step at the same sample sizes, and that the eventual structural test against the running median (above) is what would convert it from a directional finding to a settled fit.

What this channel is. Parameters are model scale purchased through training compute. The marginal-return row sits on the joint Hoffmann/Chinchilla loss surface [30, 9]: an architecture-set noise floor, a model-size term, and a data term whose coefficients are pinned by a $(M_t,D_t,A_t)$ sweep. The body keeps only the joint surface; the full chain-rule decomposition $g_t^{\mathbf{\Theta }}$, the compute-cost identity, and the worked example are collected in Channel Derivations. The complementary empirical question — how quickly deployed parameters go stale, i.e. the alpha-decay term that prices the cost of not refreshing ${\mathbf{\Theta }}_t$ — is what the deployed history already lets us measure, and we report that here.

What this means. With the model-scaling slopes still pending, the deployed evidence speaks to the other half of the parameters row: not how steep the loss surface is in $M_t$ today, but how fast the deployed weights ${\mathbf{\Theta }}_t$ themselves age out. That is the ${\alpha }^{\mathrm{decay}}$ term the controller charges against every cycle the firm chooses not to refresh its parameters.

Empirical alpha-decay from parameter staleness.

The parameter row is also where the corporation books the time-rate alpha-decay term that enters Investment marginal return and the t-RSI numerator of Three-Month t-RSI Calculation. Running the forecast-evaluation panel pooled across three training-run seeds — effectively approximately three independent five-month training-run studies (127–128 assets per run, three forecast horizons, $\sim$150,000 held-out rows per horizon over the [2020-01-01, 2025-01-01) test window) — gives a per-asset median decay rate ${\hat{\lambda }}_{\mathrm{eff}}\sim 2\text{–}4\times 10^{-4}$ per 60-minute cycle, with the per-asset Mann–Kendall slope on MAE-skill against deployment age centered near zero across the 127-asset universe. The same qualitative picture holds in production: across $\sim$16 months of cumulative Mark II and Mark III live trading (Deployment Parameters), every linear, rank, and distributional trend statistic against deployment age returns a $\hat{\kappa }$ near zero, with several leaning in the positive-trend direction. Empirically, then, the measured alpha decay sits at or below the panel’s resolution; the small magnitude is most plausibly attributable to the firm’s low $\sim 27\times$ annual portfolio turnover, which holds deployed weights inside the regime where parameter drift is dominated by sampling noise. The data-derived ${\hat{\lambda }}^{\mathrm{decay}}$ that enters the headline t-RSI is therefore small and near zero. The per-horizon decay panels, fresh-vs-stale density panels, and per-asset ${\hat{\lambda }}_{\mathrm{eff}}$ histogram are in Channel Derivations.

Model-scaling sweep is forthcoming.

The $(M_t,D_t)$ Chinchilla-style grid that pins ${\alpha }_M(A_t)$, $A_M(A_t)$, the training-efficiency term, and the model-size-dominates-data crossover is the obvious next sweep on the multi-seed scaling backbone. The current data lets the optimizer price decay (small, noise-dominated) but not yet model-size headroom; that sweep is queued as the next campaign and will refresh this row’s coefficients on completion.

The regime in which the model learns the relevant distribution in one pass, so the best validation loss after repeated epochs is no better than the loss after the first epoch. In this paper the boundary is the intersection between the one-epoch loss curve and the best-epoch loss curve. At that point ${\mathrm{epoch}}_{\mathrm{best}\mathrm{loss}}\to 1$: one repetition over the data is enough for the model to converge.

This is a structural statement about the parameters channel. Once the data can be learned in a single viewing, the firm exits the static train/validation/holdout regime and enters a walk-forward regime in which the deployed weights are refit every $t+\tau$ for some $\tau \ll T$, where $T$ is the total length of the dataset. This is the same generalization regime observed at scale in modern language-model training, and operationally it is what is commonly called continual learning—data evaluated and trained prequentially, in chronological sequence.

Two consequences propagate forward. First, the variance of the expected-return estimator collapses: every training sample becomes a valid out-of-sample evaluation point at the next refit boundary, so the firm’s posterior on its own forward returns tightens monotonically with operating history. Second, the tighter posterior feeds back into both the architecture-search channel and the controller. Architecture search inside R&D gains confidence that a candidate represents a genuine improvement rather than a sampling fluke, and the risk-aware controller allocates more optimally across all five investment channels because ${\sigma }_{c,t}$ is smaller, increasing the rate of improvement. Continual learning is therefore not just a downstream consequence of ${\mathrm{epoch}}_{\mathrm{best}\mathrm{loss}}\to 1$—it is also the mechanism by which the firm’s per-row uncertainties tighten.

The chart says that first-epoch loss is falling faster than best-epoch loss as the dollar-weighted-token budget grows. If those two power laws intersect, then the model no longer needs repeated passes to reach its best loss: the first pass is enough. Operationally, that means a smaller fraction of the available history is needed for training and a larger fraction can remain genuinely held out for testing. Taken to the limit, if the distribution can be inferred from a small enough slice of the data, continual learning becomes a prequential approximation: the firm learns from the latest examples while preserving most of the data stream as evaluation.

At the current operating point the firm measures a positive standardized distance between its alpha-creation rate and its alpha-decay rate over the next quarter. Read the headline as: how many standard errors of the posterior of the difference the projected net alpha increment $\Delta {\alpha }_{t:H}^{\mathrm{net}}$ sits above zero. The numerator is the difference of the posterior mean create rate and the posterior mean decay rate; the denominator is the standard error of that difference, propagated from an end-to-end bootstrap on the channel-row fits and the empirical alpha-decay panel cluster bootstrap. The headline conditions on the data-derived alpha-decay rate, which the firm’s forecast-evaluation panel and the 16 months of cumulative Mark II/III live trading measure as small and near zero (every linear, rank, and distributional trend statistic against deployment age returns a $\hat{\kappa }$ near zero, with several leaning in the positive-trend direction). The scope of the underlying extrapolation is narrow: the calculation only asks the data-scaling slope to continue roughly one and a half orders of magnitude past the in-sample $D_{\mathrm{eff}}$ (one decade absorbed every two months over the three-month horizon), which is well inside the 3.5-OOM in-sample range of the sweep, and integrates that extension over a single three-month horizon.

Reading the magnitude.

The headline is large because the firm currently operates below the market-impact floor: at $\sim \$400$k AUM and $\sim 27\times$ annual turnover, individual orders sit well inside the regime where they can be split without measurable price impact, and the empirical impact contribution to the create-side bootstrap is indistinguishable from zero. The data-derived alpha-decay rate is also small (Three-Month t-RSI Calculation). With neither impact nor decay materially eroding the create-side gains, the standardized distance collapses to a near-pure read of the create-side dispersion, and that read is tight. A different operating point will read differently: at K-times current AUM the literature $\sqrt{Q/\mathrm{ADV}}$ impact law [39, 55] reintroduces a capacity-driven Sharpe drag on the create side, and the headline compresses. Under a turnover trajectory consistent with the empirical decreasing-returns-to-scale pattern documented for active managers [56, 57], the same posterior reads 4.59 at $K=10\times$ AUM and 2.90 at $K=100\times$ AUM; under a worst-case trajectory in which annual turnover is held fixed at its current level, the headline crosses zero between $K=10\times$ and $K=20\times$ AUM. The full two-row capacity-sensitivity table is in Three-Month t-RSI Calculation. The thresholded form of this test statistic — the certificate of monotone improvement that gates whether a candidate update is admitted into the deployed model — is detailed in Improvement Certificate.

Pre-existing market with mandated price discovery. For t-RSI to be operationally measurable, the firm’s actions must be priced by an external mechanism at sub-cycle latency, with the prices observable and timestamped. Public equities satisfy this by federal mandate: every executed trade is reported and made public, timestamped to the nanosecond, against a market that has been continuously operating for longer than any other. The relevant point is not that this market is attractive — it is that the price-discovery machinery for every action the firm takes already exists, externally, at no construction cost and with no PMF term confounding $\partial \mathrm{Equity}/\partial \mathrm{Company}$. Most candidate domains fail this condition: in nascent markets the firm has to build the price-discovery surface itself, which adds a learned-mechanism term to every gradient; in regulated-but-illiquid markets the prices exist but resolve too slowly to score per-cycle decisions. Quantitative trading is unusual in that this condition holds trivially.

Principal value capture. t-RSI requires that $\partial \mathrm{Equity}/\partial \mathrm{Company}$ be dominated by the firm’s own actions rather than by intermediated customer behavior. A trading firm is principal to its own predictions: carry is on the firm’s own positions against the market, not a per-call fee on a customer’s downstream use of a tool. The contrast with platform-AI clarifies what the condition rules out, not which business is preferable. A platform seller’s $\partial \mathrm{Equity}/\partial \mathrm{Company}$ is dominated by enterprise sales cycles, competitive substitution, and product–market fit — exogenous, lagged, customer-dependent terms whose variance swamps the marginal effect of any single internal improvement. The same model improvement that earns a platform $0.001M in API revenue might earn the principal $100M in carry; the difference is not which firm is better-run, but which firm has a directly measurable derivative against its own actions. Most industries fail this condition because their value capture flows through a customer-decision bottleneck whose noise drowns out the create-vs-decay signal the framework needs to identify, and the standard error on the relevant gradient grows accordingly.

API-complete operational degrees of freedom. Every operational decision in the firm is, in principle, a function call. Data ingestion is an API. Model training is an API. Capital allocation is an API. Trade execution is an API. Asset acquisition—new data licenses, brokerage accounts, exchange memberships—is an API. Each call produces a structured log that doubles as the causal record needed for a derivative against it. A clothing manufacturer’s production function bottlenecks on physical objects (textile sourcing, factory throughput, retail distribution); $\partial \mathrm{Equity}/\partial (\text{material blend})$ is not even well-defined as a derivative because the action space is not differentiable. Quantitative finance is the industry where it is.

The 15-bucket cross-industry ranking below makes this exclusive overlap explicit: quantitative finance is the only domain bucket in the top quintile jointly on theoretical LM exposure ($0.94$), automation-versus-augmentation occupational share ($0.76$), and self-reported API completeness, with high confidence [58, 59]. The AEI V4 dataset (Anthropic Economic Index) covers only $10.8\%$ of O*net tasks by count; the missing $89\%$ are overwhelmingly physical or manual. The empirical frontier of LM-mediated work is therefore the frontier of API-describable work—quantitative finance lies inside that hull while manufacturing, logistics, and on-site services lie outside.

The certificate of monotone improvement (Certificate of monotone improvement) fires on one channel at a time, but channels are linked through shared state, so marginal dollars do not decompose channel by channel. Formally, positive cross-partials of the cumulative objective on the active rows (Cross-channel supermodularity) yield supermodularity in the sense of Milgrom and Roberts [60]: a marginal dollar on channel $j$ raises the marginal value of a dollar on channel $k$, and conversely. Local positivity can fail under capacity, attention, or saturation, so no global supermodular ordering is asserted. The continuation claim is probabilistic, not global (Certified-commit continuation bound): the certificate gates each commit at the prevailing operating point rather than relying on supermodularity everywhere, and while the cross-partials in Cross-channel supermodularity remain positive on those rows, an observed increase in the certified-commit improvement rate raises the posterior probability that the next commit also improves the loop, with the inequality read as an empirical row-law claim rather than a theorem.

A differentiable corporation that keeps allocating through regime change needs three things; only the first is what section 5 measures in isolation. That first requirement is fitted marginal-return laws with measurable standard errors on every operational degree of freedom. The second is drift detection on the world model itself—an operational trigger when live inputs move outside the support on which those standard errors were identified. The third is R&D throughput high enough that, once that trigger fires, the firm refits to the new regime before an obsolete mapping bleeds out deployable capital. The certificate of Certificate of monotone improvement is evaluated on the same audited moments, so its economic content weakens exactly when drift invalidates the inferential basis for those errors.

On (2), the alpha-decay panel of § 5.8 already constitutes this surveillance and measures the decay rate as small and near zero on its measured horizon. On (3), the continual-learning construction of § 5.7 implies a refitting time constant that is short relative to the decay time constant summarized in (2), at the firm’s prevailing data-ingestion rate. The firm’s recovery rate therefore dominates its measured decay rate, which is what self-improvement under regime non-stationarity actually requires.

Sutton’s bitter lesson observes that the methods which win at scale in machine learning are the ones whose performance grows with computation, not the ones that encode hand-crafted human structure [61]. The capital analog is structurally identical and historically older: the firms that win at scale are the ones whose throughput grows with absorbed capital, not the ones whose decisions are bounded by the size of a fixed staff. The supermodular cross-partials of § 6.2 compound the internal loop at the rate the channels permit; the bitter lesson is that the loop does not stop there.

Once the corporation demonstrates, in its own operating record, that incremental capital reliably converts into measured row improvement and deployable edge, outside investors can treat that relationship itself as an investable object. The next-period deployable-capital identity (Deployable-capital decomposition) separates internally generated deployable capital $K_{t+1}^{\mathrm{int}}$ from externally supplied capital $K_{t+1}^{\mathrm{ext}}$. External capital amplifies the loop only while each marginal externally supplied dollar clears the same risk-adjusted certificate after financing costs, dilution, market impact, and capacity effects.

Two things are worth mentioning. First, different financing instruments have norms—and in some cases strict bylaws—governing what they can and cannot invest in. As the world gets increasingly digitized, the instruments with more leeway and more quantitative discipline will, on average, outcompete those that cannot process as much information; this is its own selection pressure on the type of external capital a recursively improving corporation can absorb. Second, a self-improving corporation with improving capability attracts more capital; when the marginal certificate continues to clear, that increased capital further increases the rate of self-improvement, which increases the rate of returns, which in turn increases the rate of external investment. Some of that capital—particularly AUM and equity—does carry per-dollar performance decay through market impact and capacity. What matters is the race: when incoming capital buys data or architectures that raise expected return faster than market impact erodes alpha, the rate of self-improvement continues to climb and the rate of external investment accelerates with it—a positive feedback loop.

The next phase is to close the remaining operational gradients. Salary and headcount cost, banking and cost of capital, hardware procurement and depreciation, asset-acquisition cost, and AUM acquisition cost are the remaining major terms in the corporate equation. Each term has a natural measurement surface: dollars in, operational capability out, and realized contribution to future equity growth. As those surfaces are instrumented, the corporation becomes progressively more legible to its own controller.

The completed object is a firm whose capital-allocation process is scored end to end: every major expenditure has a forecast, every forecast has a realized outcome, and every realized outcome updates the next allocation. The differentiable corporation is the limit of that process.

The Economic World Model is a network trained on the joint distribution of priced economic data—a foundation model for allocation in the same sense that a large language model is a foundation model for text. Text describes economic activity; prices settle it; the same underlying forecast can therefore be executed at more than one depth in the real economy rather than only as a paper position.

The controller can be instantiated in three computational regimes, each removing one piece of hand engineering. The first is the one this blueprint develops in detail: a hand-factored chain rule whose per-channel scaling laws compose into a one-step gradient and equilibrate against the marginal ROI ${\lambda }_t^{\ast }$. The second replaces the hand factorization with a single neural Economic World Model trained end-to-end on the firm’s operating history; $J_t$ becomes the discounted sum of predicted log-equity returns under a finite MPC roll-forward and ${\nabla }_{{\mathbf{a}}_t}{J}_t$ is recovered by autodifferentiation through the rollout [16, 17]. The third parameterizes the allocation policy and lets gradients of cumulative log-equity flow back through the world model and into the policy itself: the differentiable-simulator paradigm of PILCO [62], the Dreamer family [63, 64, 23], and the broader differentiable-world-model line [21, 22]. The certificate of monotone improvement (Certificate of monotone improvement) extends across all three regimes: each candidate update is admissible only when the held-out t-RSI clears the Sharpe-margin threshold $\delta$ and the Fisher-information readiness floor ${\epsilon }_c$ on every active channel.

These regimes describe how the controller reasons; the question of what it acts on is orthogonal and develops along its own axis. If we take intelligence to be the capacity to acquire, preserve, and compound command over future resources, those resources need not arrive through a brokerage API. A forecast over future supply and demand has value because it implies future prices, constraints, and margins. The shallowest channel that realizes a predicted edge is a futures position; a deeper channel realizes the same edge as a sequence of actuators in ${\mathbf{\Xi }}_t$ (procurement, transport, refining, wholesale distribution) and captures the entire margin between input and output rather than the basis alone. Each version is the same economic prediction executed at a different depth, scored by the same per-dollar log-growth functional that the portfolio optimizer compares across heterogeneous channels in cycle $t$. The depth at which the firm operates is not a strategic preference but a Coasean comparison the framework already makes [65]: the firm verticalizes the next step of a production chain when two conditions hold together—the executing channel clears the certificate of Certificate of monotone improvement, and the certificate-corrected risk-adjusted cost of running that step in-house is lower than the market price of buying the same output. New actuators—humanoid platforms, AI agents, autonomous logistics services—enter the firm’s action set as soon as their success distributions are reliable enough for the EWM to learn and condition on. Because deeper execution demands a more capable $\hat{\mathcal{W}}$, the regimes and the depth axis are coupled: better reasoning unlocks deeper action, and deeper action enriches the operating record that the next regime trains on. At sufficient depth and scale the small-firm approximation no longer holds: a firm whose actions move corn prices, contract refining capacity, or absorb a meaningful slice of the financing pool perturbs ${\mathbf{E}}_{t+1}$ by its own choices. This is not a contradiction of the framework; it is absorbed by it. The firm’s filtration (Firm filtration) accumulates the joint record of its own actions and the world’s responses to them, so the EWM trained on ${\mathcal{F}}_t$ can in principle learn the reaction term that the small-firm limit zeroed out. Operationally the change is local: the same channel-row laws are refit on a richer history, and the certificate continues to gate every commit.

The standardized create-vs-decay distance, t-RSI (Trsi Net), is computable at AlphaFund. We believe quant trading is among the first domains where such a statistic is practically computable. A clothing retailer asking “does this material blend increase next-quarter equity” faces a numerator whose sign is contested between marketing-campaign quality, retail traffic, supply-chain cost, and the broader economic cycle; the standard error on the marginal effect is so large relative to the marginal effect itself that t-RSI is, in practice, undefined. A platform company selling tools—advertising slots, API tokens, software seats—inserts a product–market-fit layer between the loss its model minimizes and the revenue its balance sheet collects; the bridge is opaque, lagged, and customer-dependent, and the implied $\partial \mathrm{Equity}/\partial (\text{model improvement})$ is dominated by exogenous customer behavior rather than by the model. In both cases the underlying problem is structural: the standard errors on $\partial \mathrm{Equity}/\partial \mathrm{Company}$ are large relative to the effects they would measure at any reasonable sample size. The moat is therefore not the value of AlphaFund’s t-RSI. It is that t-RSI is practically computable in this firm and this industry. The single scalar that summarizes the legibility of a corporation to itself is, at the time of writing, much harder to construct elsewhere.