Continuous intelligence.
Disciplined capital.

A research institution whose capacity grows through software.

Investment research has a recurring coordination problem. Relevant facts arrive at different times, in different formats, and at different levels of a business. A demand statement belongs to one document; a change in payment terms to another; a customer dependency to a third. The analytical work is to connect those facts, preserve their timing and determine whether the connection matters at the price available to an investor.

Base One’s investment thesis spans public and private markets, including startups. A shared research discipline connects these investments, while their horizons, evidence and liquidity requirements differ. The detailed quantitative protocols below address liquid securities; private-company diligence and allocation are addressed in Section 10. Our thesis begins with the work of connecting evidence. Our AI-native research and proprietary trading firm is organised around a continuous process: observe a change, assemble its evidence, formulate a hypothesis, test its incremental value and decide how much capital the evidence can support. The ambition extends beyond accelerating an analyst’s reading. It is to make investigations persistent, experiments reproducible and decisions traceable across a large research network.

Our central proposition is conditional: a reduction in the cost of reliable investigation can create investment value where interpretation remains expensive, the resulting information is not already fully reflected in prices, and implementation leaves an economic surplus. Each condition can fail. Faster research may merely reproduce consensus. A superior forecast may be impossible to trade economically. A profitable signal may add the same risk already held elsewhere.

We therefore treat research productivity, predictive usefulness and investment usefulness as separate claims. The firm becomes stronger when it can identify exactly where an idea fails. These three hypotheses organise the programme:

The AI decade: intelligence becomes an industrial input.

Our research architecture has an investment direction. We expect increasingly capable and economical AI to reshape the demand for infrastructure, the economics of software and the organisation of work over the coming decade. We intend to investigate the businesses enabling that transition and those able to turn it into durable cash flows. The agent network is the institution through which we examine this outlook, discover opportunities and revise our judgments.

The intellectual foundation

Leopold Aschenbrenner’s Situational Awareness: The Decade Ahead informs this outlook. His June 2024 essays connect greater computing scale, algorithmic efficiency and more capable agents with the possibility of transformative AI. They also examine how automated AI research could accelerate further progress. His forecast of AGI around 2027 is a scenario we take seriously, with substantial uncertainty; it is neither an established fact nor a deadline on which every investment must depend. [14] [15]

The public mandate of Situational Awareness LP similarly connects research on AI technology and economics with the search for beneficiaries and disrupted businesses. We adopt that connection as a research principle. Base One is independent and unaffiliated. Published essays and a public mandate do not disclose another firm’s complete portfolio, entry prices, financing or decision rules. Our strategy and risk policy must be specified and tested on their own terms. [22]

In his 2024 interview with Dwarkesh Patel, Aschenbrenner describes a research institution within an investment firm and emphasises the sequence of opportunities and survival through mistaken calls. Our coverage should similarly follow where AI changes a company’s earnings materially, and update as the constraints and beneficiaries change. [25]

Follow the physical requirements of intelligence

AI is also an industrial process. More useful models and wider deployment can require accelerators, memory, advanced packaging, networking, storage, data centres, cooling and dependable electricity. Aschenbrenner’s infrastructure essay makes the scale and constraints of that buildout central to the outlook. Our investment question is which constraints persist, who can resolve them and who retains economic value as supply responds. [16]

We organise investigation across three connected areas:

• Compute and its supply chain. Examine performance per unit of cost, capacity, customer concentration, substitution and pricing power. A scarce component today may become abundant before an investment earns its required return.
• Data centres and power. Distinguish announced capacity from financed, permitted, connected and operating capacity. Evaluate construction obligations, equipment replacement, contracted customers and cash flow after reinvestment.
• Deployment and disruption. Investigate platforms and applications with demonstrated customer value, defensible distribution and sustainable economics. Also identify businesses whose pricing or demand could weaken as intelligence becomes cheaper.

Crypto-related equities enter this framework through their actual businesses. A mining company’s proposed move into AI infrastructure requires evidence of suitable facilities, reliable power, customers and financing. A power contract alone does not establish an AI-ready data centre. Exchanges and other crypto businesses require a separate assessment of their revenues and risks; association with cryptocurrency is not evidence of an AI advantage.

Separate the trend from the investment

We trace each candidate through an explicit chain: capability → adoption → demand → company economics → shareholder cash flow → value relative to price. Evidence supporting one link cannot establish the rest. Greater AI spending can benefit customers while suppliers earn poor returns. Revenues can grow while debt, dilution or replacement expenditure consume the shareholder’s gain. Even an exceptional business can be unattractive when its price already assumes an exceptional future.

Current evidence must replace old extrapolations. Microsoft’s July 2026 release reports Azure annual revenue above $100 billion and more than 30 million paid Microsoft 365 Copilot seats; total Azure revenue is not standalone AI revenue or profit. The Department of Energy’s current resource hub summarises a modelled 9.5–15.3% share of US electricity use for data centres in 2030. These observations and forecasts help define research questions. They do not verify every prediction made in 2024 or establish an investment return. [23] [24]

Security, state involvement and competing futures

Security and control are central to this worldview. Aschenbrenner’s security and alignment essays examine the protection of model capabilities and the difficulty of supervising more powerful systems. His geopolitical and government-project essays consider strategic competition and greater state involvement. For investors, these are questions about access, customers, intellectual property, supply chains and ownership economics, as well as sources of disruption. They are not guarantees of national dominance or favourable shareholder treatment. [17] [18] [19] [20]

Each company assessment must consider faster capability growth, slower commercial adoption and an overbuilding or funding contraction. We examine what happens if efficiency reduces demand for a particular component, if new supply removes scarcity, or if regulation changes who captures value. The skeptic can challenge the overarching outlook as well as the company thesis. The author’s own acknowledgment of large uncertainty reinforces this commitment to revision. [21]

The structural outlook spans five to ten years. Our current public-equity research focuses on US-listed beneficiaries across the AI value chain, including power, energy and infrastructure. Each company still needs its own investment horizon, review triggers and exit conditions. We assess substantial upside alongside the funding, dilution and downside that could prevent shareholders from capturing it. A compelling long-term view does not remove the need to survive the path to it.

Where a persistent information advantage could arise.

Grossman and Stiglitz show, in a model with costly information, why perfectly informative prices undermine the incentive to acquire information. The relevant lesson is that information processing has an economic cost and that price informativeness is an equilibrium outcome. It does not follow that any particular information-processing technology earns excess returns. [1]

Our inference is narrower. Where the cost of connecting evidence falls, some previously uneconomic investigations may become worth performing. We are interested in repeated comparisons across long histories, the propagation of developments through business relationships, and changes whose significance depends on several facts being considered together. A company’s description of stronger demand, for example, has a different interpretation when cash collection weakens and its largest customers reduce orders.

There are empirical precedents for studying these mechanisms. Cohen, Malloy and Nguyen document return predictability associated with changes in corporate reporting in a historical US sample. Cohen and Frazzini document delayed responses across customer–supplier links. These studies motivate research into document change and economic connections; their historical results are neither our expected returns nor proof that the effects persist after dissemination and contemporary costs. [2] [3]

The strongest objection is also central to the thesis: widely available AI may make markets faster at interpreting precisely these relationships. Access to a foundation model is therefore a weak basis for differentiation. Base One’s opportunity must come from the quality of its question selection, time-aware evidence, evaluation and execution. As competitors automate obvious signals, deeper contextual work may become relatively more valuable, or the entire opportunity may become less profitable. Both outcomes belong in the research agenda.

Our immediate public-market programme centres on company selection and the longer-term economics of AI adoption and infrastructure. Disclosures and market events prompt a review of the investment case; they do not impose a short holding period. Event-driven and systematic methods remain part of the broader research agenda. A market enters the programme only when its source history, trading access, costs and liquidity permit credible tests. Breadth of coverage is valuable only after this minimum standard is met.

AI changes the unit of research.

An AI-enabled analyst uses a model to summarise a report or write code. An AI-first workflow asks which parts of an investigation software can perform reliably. An AI-native institution makes the resulting evidence, tasks and experiments shareable, versioned objects from the start. This final step changes how research accumulates: the useful output of a task is a record another process can inspect and extend.

In Base One’s design, a research task has an objective, a permitted information set, a resource budget, an output schema and a termination condition. A result contains supported observations, calculations, unresolved conflicts and a decision about what to investigate next. The agent’s prose is an interface to this record; the record carries the institutional knowledge.

Language models are well suited to finding and interpreting heterogeneous text. Numerical transformations, accounting reconciliations, portfolio constraints and order checks belong in explicit code wherever practical. The distinction matters because fluent explanations do not guarantee correct arithmetic or a consistent application of a rule. Agents can help author an implementation, but its released version must be independently testable.

Text-as-data research supplies a useful methodological foundation: language becomes an input to measurement rather than a substitute for it. [4] Our extension is to connect those measurements to an ongoing investigative process. A model can flag that a disclosure changed; an agent can establish which clause changed, retrieve the relevant comparison, check an alternative explanation and produce a feature with a documented meaning.

Human responsibility concentrates at the points where errors have durable consequences: setting the research agenda, resolving consequential ambiguity, accepting an evaluation design, approving a strategy version and changing capital limits. A thousand parallel investigations should produce a smaller number of better-supported decisions, with clear ownership of each decision.

The evidence record is the foundation.

A research system needs to answer two different questions: when did an event happen, and when could this system have known about it? A quarter-end balance sheet may describe a period that ended weeks before publication. A revised dataset may contain facts that were unavailable at the original decision time. Treating either as contemporaneous creates an artificial information advantage.

Our evidence design retains the original document alongside structured fields. It records publication, observation and processing times separately; amendments become new versions rather than silent replacements. Entity identifiers distinguish issuers, securities and corporate relationships. Units, currencies, reporting periods and accounting definitions travel with the number. A source passage supports a claim only within that context.

Sources such as the SEC’s EDGAR APIs provide programmatic access to submissions and XBRL facts. They are useful building blocks, but a live interface alone does not reconstruct the dataset a researcher possessed years earlier. [5] Source-specific time conventions, amendments and ingestion delays still need to be reconciled. Licensed price histories, corporate actions, delistings and security mappings require the same discipline.

For a retrospective simulation of this architecture, the earliest usable time is bounded by the latest of public availability, simulated receipt and processing completion. Historical processing delays must be measured or conservatively assumed, not replaced by today’s retrieval speed. A feature relying on several records inherits the latest usable time of its inputs. If timing cannot be established, the observation is delayed conservatively or excluded under a rule fixed before evaluation.

Shared memory stores claims with provenance and validity periods. A company relationship observed today cannot be inserted into a historical network as though it were known then. A correction invalidates or flags the downstream features and experiments that depended on it. This dependency chain is what makes institutional memory useful: the system can revise its beliefs without erasing how it reached them.

A swarm needs a division of labour and a stopping rule.

Our ambition is a 24/7 cloud network spanning hundreds, and ultimately thousands, of specialised agents. Here, an agent is a bounded software worker that can choose tools and intermediate actions to complete a research task. A persistent coverage responsibility can survive many worker sessions; it does not require a model to generate text continuously. The swarm is the coordinated network of these responsibilities and investigations.

The operating desk gives these responsibilities concrete owners. A chief of staff coordinates evidence gathering, business research, valuation, quantitative work and an independent challenger. Portfolio, macro, operations and performance responsibilities connect the analysis to the wider investment process. Independent risk reports directly to the founder; a coordinator cannot waive its standing blocks. Agents can share findings directly while assignments and decisions follow the defined reporting structure.

Anthropic’s engineering report describes an orchestrator–worker pattern for parallel research and documents its increased resource demands. It supports the architectural possibility, not a financial-performance claim. [6] We take task decomposition seriously because not all work benefits from parallelism: accounting reconciliation with tightly shared state may be better handled by a deterministic pipeline than by agents exchanging interpretations.

Connecting the investment outlook to the swarm

The coordinator turns the AI outlook into bounded company questions. Business research tests the causal link to demand; valuation examines cash flow and market-implied expectations; quantitative research measures the strategy against simpler controls; an independent skeptic challenges both the company case and the macro assumptions. Portfolio construction examines shared exposures, and independent risk review can block an incomplete proposal. The founder brief preserves the strongest objection and the next observable test. More agents endorsing the same AI narrative do not constitute more independent evidence.

A task ends with a checkpoint recording evidence, findings, uncertainty and the next action. Shared memory and versioned source packets preserve the handoff across sessions and model platforms. Corrections are new linked records rather than silent edits. The broader architecture calls for bounded retries, task reservations and explicit model-version logging as orchestration develops. Reproduction means preserving what actually happened, including model outputs; replaying the same prompt cannot guarantee identical generation.

Disagreement is a routing signal. It may identify an ambiguous definition, a missing document or a genuine alternative explanation. Majority voting is insufficient when agents share training data, sources or prompts. Checks should use independently located evidence, deterministic calculations and blinded review where feasible. External documents remain untrusted data: their embedded instructions must not change permissions, access secrets or authorise orders.

The swarm must beat simpler alternatives.

Agent count is an input, not an achievement. Cemri and colleagues identify failures in system design, inter-agent alignment and task verification across multi-agent frameworks. Their findings reinforce the need to test the complete workflow, including handovers and termination, rather than infer system reliability from the capabilities of an individual model. [7]

Weak-to-strong generalisation research, coauthored by Aschenbrenner, examines how stronger models can learn from weaker supervision while also learning its errors. Its task-dependent results do not establish reliable financial oversight. For our council, the implication is to verify sources and calculations and test reviewers on known errors, rather than assume that a more capable model or another layer of review guarantees correctness. [26]

Our evaluation design compares three systems on the same frozen source packets and questions: a scripted extraction-and-retrieval pipeline, a single capable agent with tools, and the coordinated network. One comparison fixes the compute budget; another fixes the deadline and reports the additional cost. Without both views, a swarm may appear superior simply because it receives more resources.

The test set includes routine disclosures, amended statements, tables with changing units, contradictory passages, sparse evidence and documents containing malicious instructions. It is stratified by source type and difficulty. Independent reviewers establish reference answers and adjudicate disagreements without seeing which system produced an answer. Evaluation documents are held apart from prompt development; repeated failures may inform a new version, which needs fresh evaluation.

For extraction, we measure precision and recall at the field level, numerical and unit accuracy, and whether the cited passage supports the claim. For investigation, we measure coverage of relevant evidence, unsupported conclusions, correct abstentions and the human minutes needed to reach an accepted record. Selective accuracy must be paired with coverage: a system that answers only easy questions can otherwise look deceptively reliable.

We also remove individual components and repeat the comparison. Does a challenge agent catch material errors? Does shared memory reduce repeated investigation without spreading stale claims? Does parallel search increase useful source coverage after deduplication? These ablation tests attribute improvement to a mechanism. A component that adds cost without a measurable benefit should be simplified or removed.

Model confidence is not automatically a probability. If the system reports confidence, we compare it with observed correctness on held-out cases and inspect calibration across task types. Repeated runs expose stochastic instability. A model upgrade is evaluated as a change to the measurement system, with shadow comparisons and a rollback path. Passing research-quality tests qualifies the output for quantitative investigation; it does not establish a trading edge.

From structured observations to measurable forecasts.

For liquid markets, our quantitative agenda connects three research families: changes in company information, propagation across economic relationships, and systematic price-and-risk behaviour. Their value is assessed through a common modelling discipline. A candidate feature must have a defined meaning, an availability time, a plausible mechanism and a benchmark it is expected to improve.

Document changes and event information

A disclosure feature might measure the direction and magnitude of a guidance revision, a change in payment terms, or a divergence between management language and operating data. The baseline includes ordinary financial variables and transparent text measures such as term counts or document similarity. An LLM-derived interpretation earns its place only if it improves on these cheaper measurements, with the same source history and decision time.

Features should describe an economic observation rather than an unconstrained instruction to “predict the stock”. For instance, inventory growth relative to sales growth is a reproducible quantity; whether management attributes the difference to a temporary supply buffer is an evidence-linked classification. The numerical measure and textual explanation can then be tested separately and jointly.

Economic relationships and transmission

A company network can represent disclosed customer, supplier or other material dependencies. Edges need a source, an observed date, a relationship type and, where available, a measure of economic importance. The experiment asks whether a change at a connected company adds information beyond the focal company’s own disclosures, sector returns and common market shocks.

A weighted aggregation over verified links provides a transparent baseline. More elaborate graph models are candidates only when enough reliable historical relationships exist to estimate them. Comparisons with shuffled or deliberately irrelevant links help test whether the economic relationship matters. Such checks diagnose a predictive mechanism; they do not, by themselves, establish causality.

Market signals and nonlinear models

Trend, relative strength, volatility and liquidity provide a baseline against which richer information must compete. Regularised linear models make coefficients and stability inspectable. Boosted trees can test threshold effects and interactions. Neural networks can learn richer representations where the effective sample and task justify them. Gu, Kelly and Xiu find gains from nonlinear predictor interactions in their historical asset-pricing comparison; this motivates model comparison rather than an assumption that complexity will win. [8]

Here, x contains conventional company and market features; z contains agent-derived observations; and m describes contemporaneous market conditions. Parameters θ are estimated only on the training window. We compare this model with an otherwise equivalent model that excludes z. This isolates the contribution of the research layer more convincingly than comparing a complex new system with a deliberately weak benchmark.

The outcome may be a future fundamental change, a return forecast or a risk estimate. These are different tasks with different labels. A correct earnings forecast can still produce a poor trade if it is already anticipated in price. Likewise, low prediction error need not imply profitable ranking or useful portfolio weights. Model selection therefore reports statistical accuracy, ranking quality, forecast calibration where relevant, turnover and net portfolio contribution.

“Deep” describes a model architecture, not a standard of evidence. Sequential models require enough independent history; many correlated observations do not create many independent market regimes. Embeddings must retain their model version. Normalisation, missing-value treatment and dimension reduction are fitted within the training fold. If a small model performs comparably with lower instability and cost, it is the stronger implementation.

A demand claim, a cash-conversion question, a falsifiable test.

Consider a fictional manufacturer reporting stronger demand. Its current and prior comparable periods contain the figures below. The example is deliberately small enough to audit; it illustrates a research protocol, not an investment recommendation or a Base One result.

The deterministic calculations produce an inventory-growth gap of 15 percentage points relative to revenue growth, and a receivables-growth gap of 22 percentage points. Neither is a trading signal by itself. Inventory is a stock and revenue is a flow; seasonality, period comparability, acquisitions, currency translation and accounting changes can invalidate a naive comparison.

The extraction worker records the values, periods and source cells. The investigation worker checks comparable seasonal periods and management’s prior explanations. A relationship worker looks for contemporaneous evidence from material customers. The challenge worker examines whether a product launch, acquisition or deliberate stocking policy explains the pattern. Unsupported causal language is removed from the final record. If the periods cannot be reconciled, the case remains unresolved.

A concrete experimental question follows: does the conjunction of working-capital divergence and a disclosed extension of payment terms improve a forecast of subsequent cash-conversion deterioration, and does it contain additional information about post-publication returns? The fundamental outcome and the return outcome are evaluated separately. Discovering one does not establish the other.

An illustrative preregistered experiment

Use non-financial listed companies in one market with archived, timestamped quarterly disclosures and a security history that includes later delistings. Fix eligibility, liquidity requirements and all exclusions before examining evaluation outcomes. Compute the two growth gaps using like-for-like periods and preserve the payment-terms classification with its supporting passage. Missing disclosures remain a documented source of selection bias, not a reason to quietly remove inconvenient companies.

Choose one primary return horizon (for illustration, 20 trading sessions) and define the label as the security’s total return less its predefined sector benchmark over that interval. Entry follows the later of document availability and processing completion, under a stated venue and execution assumption. Define the separate fundamental outcome as the change in operating cash flow divided by revenue at the next report, relative to its prior-year comparable period. Where cash-flow statements are cumulative, derive comparable discrete periods before forming the ratio; invalid denominators and unreconcilable statements follow predefined exclusion rules.

Use an elastic-net regression (a linear model with penalties that shrink and select coefficients) as the initial estimator for each outcome. Compare the enriched model with a baseline containing the same accounting quantities, sector, size, prior returns, volatility and the price reaction already observed before entry. Tune both models under the same validation budget. A simple text classifier supplies an additional benchmark. Develop and tune on chronological windows, freeze the specification, and evaluate on later observations. Assess whether the text feature adds anything after the numerical gaps, and whether linked-company evidence adds anything after both. Report the full distribution of results and uncertainty. Any secondary horizons or revised feature definitions count as additional tests; a disappointing primary result cannot be replaced silently by the most attractive variant.

Finally, apply the identical sizing, liquidity and cost rules to baseline and enriched portfolios. If the improvement vanishes after a realistic delay, it is not executable at that speed. If it survives only in securities too illiquid to trade at the assumed size, it does not support that capital allocation. If no incremental effect remains, retain the negative result and stop the strategy branch unless genuinely new evidence changes the question.

A larger search demands stronger statistical discipline.

A network that generates hypotheses rapidly increases the opportunity for discovery and the number of ways to fool itself. The experiment registry must therefore include failed trials, changes in labels, prompts, features, universes, costs and model settings. A strategy’s best result is interpretable only alongside the search that produced it.

Chronology and selection

The primary design is a rolling or expanding walk-forward evaluation: train on earlier data, select settings on a subsequent validation window, and test on a later window untouched by that selection. Repeat the process using only information that would have been available at each retraining date. All preprocessing and feature selection occur inside the appropriate training window.

Outcome windows must also respect the boundary. A training observation whose future-return label extends into the test interval is removed. Where a cross-validation design uses observations on both sides of an evaluation block, remove overlapping labels and apply an appropriate post-block embargo. The exclusion period follows label and information dependencies; an arbitrary fixed gap is not a universal cure. The final holdout has a finite role: once used to revise the method, it becomes development data.

Bailey and colleagues’ probability-of-backtest-overfitting framework examines how selecting a winner among configurations can degrade performance outside the selection sample. Their combinatorial diagnostic complements, rather than replaces, the chronological deployment simulation. The Deflated Sharpe Ratio addresses selection and non-normality in reported Sharpe ratios. Both require care with assumptions and trial dependence; neither converts an incomplete experiment history into credible evidence. [9] [10]

Uncertainty and economic relevance

Daily portfolio observations are serially dependent, and company events cluster around shared market shocks. Uncertainty estimates should respect those dependencies: for example, resample blocks of portfolio dates and use issuer/date-aware inference for event panels. Block length and clustering choices need sensitivity checks. Report interval estimates, variation across time windows, concentration of profits, tail losses, turnover and the effect of removing the strongest period or names.

There is no universal minimum number of months or trades that proves a strategy. Required evidence depends on effect size, dependence, tail behaviour and the number of alternatives considered. A thousand highly correlated trades can contain less independent information than a much smaller number of distinct events. Evaluation length and the minimum economically useful improvement are chosen before inspecting the final comparison, with uncertainty carried into the allocation decision.

The language model’s historical knowledge

A historical document is not necessarily a historical information set when a modern model reads it. Glasserman and Lin distinguish look-ahead bias from distraction caused by company knowledge; in their experiment, anonymised headlines outperform the originals in-sample, illustrating that these effects are not a simple one-directional inflation story. [11]

Our protocol records model versions and known training boundaries, preserves prompts and outputs, restricts retrieval by time and tests anonymisation where useful. None of these proves the absence of memorised future information. Historical LLM-derived results therefore receive less evidential weight when training overlap is unresolved. Frozen forward collection on newly arriving information is essential, with any model change beginning a separately identified evaluation cohort.

Execution realism

A backtest includes spreads, fees, slippage, market impact, financing and borrow costs where applicable. It accounts for corporate actions, halted trading, rejected orders and unavailable borrow. Capacity analysis varies capital and participation rather than scaling returns linearly. Stress cases widen spreads, delay execution and remove expected liquidity. A closing-price fill is not acceptable merely because the database offers that price.

Forward simulation then tests the complete path from arrival to intended order under frozen rules. A limited live pilot, where authorised, tests actual fills and operational behaviour. Continuous monitoring does not justify continuous strategy tuning: promotion and stopping rules must be documented, and formal repeated statistical looks require a sequential-testing design or a predefined review schedule.

The portfolio is a constrained decision, not a collection of opinions.

The object of proprietary trading is the portfolio outcome. A forecast becomes useful only in relation to existing positions, uncertainty, liquidity and the cost of changing exposure. Two apparently distinct signals may both be long the same industry or sensitive to the same funding shock. Capital cannot be assigned by counting how many agents endorse an idea.

Boyd and colleagues provide a framework that combines return forecasts, risk, transaction costs and holding costs in portfolio optimisation. Their work addresses how to use forecasts rather than how to create them. [12] A simplified formulation makes our decision logic explicit:

The forecast vector μ̂ estimates returns; Σ̂ estimates covariance; λ expresses risk aversion; C estimates the cost of trading from the current weights; and H covers holding costs such as borrow. The feasible set W encodes capital accounting, instrument permissions, gross and net exposure, concentration, liquidity and other hard limits. This is a public description of an allocation problem, not a disclosure of our production optimiser.

Estimation uncertainty matters as much as the optimum. Forecasts should be shrunk toward conservative baselines, covariance estimates regularised and sensitivity to plausible input changes examined. Historical covariance does not describe every crisis; stress scenarios test concentration and liquidity under conditions the fitted model may underestimate. A small forecast change should not cause an unexplained large position reversal.

Admission to capital requires a versioned evidence package: mechanism, source dependencies, experiment history, forward observations, cost model, capacity assumptions, failure modes and a named owner. A staged progression from historical research to shadow operation, a bounded live pilot and possible scaling allows each stage to answer a different question. Capital increases follow evidence from the actual operating conditions rather than a calendar or a target agent count.

The founder reviews recommendations and retains order authority; research agents have no execution authority. A future execution service, if separately authorised, would consume approved strategy outputs through a constrained interface. It checks price and data freshness, permissions, position limits and order identifiers before submission; reconciles broker acknowledgements and fills; and compares intended with realised positions. On a timeout, it checks order state before retrying. A research worker has no route to change these controls or access trading credentials.

Suspension rules distinguish a loss within the model’s expected behaviour from broken assumptions or an operational fault. Stale inputs may block new orders. A position mismatch may require reconciliation before further trading. A breached exposure limit may require controlled reduction. The response must account for outstanding orders and liquidity: an indiscriminate liquidation can increase harm during disruption. Responsibility for suspension, recovery and any restart is explicit.

Private companies within the same capital discipline

A private investment begins with a company-level hypothesis: which customer problem is valuable, what evidence supports demand, and what milestones could change the business’s value. Agents can compare markets, organise diligence materials and track evidence over time. Customer references, product inspection and reconciliation of reported revenue, cash and financing needs remain substantive checks. AI-generated confidence does not replace verification.

Assessment considers entry valuation and ownership terms alongside business quality. Downside, base and upside scenarios examine cash runway, future financing, dilution and potential exit proceeds. A strong company can still be an unattractive investment at the wrong price or terms. These scenarios expose assumptions; they do not turn sparse private-company histories into statistically validated return forecasts.

The SEC’s investor bulletin highlights illiquidity, limited disclosure and the risk of total loss in private placements. These characteristics reinforce the need for investment-specific diligence and patient capital. [13]

Private holdings require explicit limits on concentration, illiquid exposure and any capital reserved for follow-on investment. Capital planning considers a prolonged holding period, further funding needs and the possibility of no exit. Periodic valuation marks are not realised returns, and infrequent marks can understate economic volatility. The liquid-portfolio optimisation above is therefore not applied mechanically to private holdings; their cash commitments and downside scenarios must still enter the firm-wide capital budget.

Company information received in diligence is permissioned and kept within its authorised use. Potentially material non-public information is escalated for appropriate review before any related public-market trading. A shared research institution does not imply unrestricted movement of confidential information between investment decisions.

Realised performance in liquid strategies is decomposed into forecast contribution, common exposures, sizing, trading costs and execution deviations. This decomposition routes failure back to the right part of the institution. A good signal with poor fills needs a different response from an accurate extraction that never contained predictive information.

Research capacity and trading capacity are different constraints.

The cloud network should expand where additional investigation creates useful evidence at an acceptable cost. Persistent monitoring can be inexpensive; deep investigations can be selective. Cached documents, shared extractions and reusable calculations prevent thousands of workers from paying repeatedly for the same fact. Compute is allocated to unresolved questions, not to keeping every agent busy.

Research cost includes model inference, retrieval, data licences, storage, engineering, human review and correction. A useful operating metric is total cost per accepted research record, accompanied by coverage, latency and error severity. Measuring only tokens or completed tasks rewards cheap, superficial work. Measuring only successful hypotheses hides the cost of rejected experiments.

A second ledger evaluates strategy economics. The relevant surplus is net trading contribution less the attributable research and operating burden. Shared infrastructure should be visible at firm level even when it cannot be assigned neatly to an individual signal. Under uncertainty, research budgets can be released in stages: a low-cost feasibility check, a bounded historical study, then forward evaluation if the prior stage resolves enough uncertainty.

Trading capacity places a separate ceiling on the business. A strategy may work at modest size and decay as participation increases, execution worsens or competitors crowd the same signal. It is therefore possible for research capacity to grow while economically deployable capital does not. Base One’s scale ambition requires a portfolio of sufficiently distinct opportunities, not merely more agents discovering variants of one trade.

Task allocation also needs protection from self-reinforcing bias. If all compute flows to historically successful research themes, the system can become blind to new conditions. A bounded exploration budget preserves room for unfamiliar questions, while repeat investigations need a reason grounded in new evidence. The value of an experiment includes learning that a promising direction should be abandoned.

What would make us change our mind.

A credible investment thesis specifies the observations that would weaken it. The following objections are not peripheral risks; they challenge the causal chain of the business itself.

No model makes this uncertainty disappear. Market participants adapt, data-generating processes change and rare events remain difficult to estimate. Our response is to make assumptions inspectable and revisions controlled. Conviction becomes more valuable when the institution can distinguish a temporary disappointment from evidence that its explanation was wrong.

The institution should become more informed with every experiment.

The execution sequence follows the dependencies of the thesis. First, establish a reliable evidence substrate and a narrow research task with a credible baseline. Next, show that agent collaboration improves that task after its full cost. Then test whether the resulting measurements improve forecasts. Only then does the capital process ask whether those forecasts justify executable positions. This sequence makes progress observable without disclosing the particular signals or parameters that may become proprietary.

Each stage produces a durable artifact: an adjudicated evaluation set, a versioned source history, a registered experiment, a reproducible result or a strategy admission record. Expansion into another market or research family should reuse the infrastructure while repeating the relevant evidence tests. A broader universe is a new data and execution problem, not merely another prompt.

The desk’s learning protocol connects an observed error or insight to its root cause, a concrete process change and a named owner. A lesson is adopted only when that change is visible in a rule, template, script or gate. Daily manifests and weekly reconciliation define how research records are brought back into a coherent institutional history.

Our intended institutional advantage accumulates in the relationships between these artifacts. Source reliability informs future retrieval. Failed experiments narrow the next search. Live execution discrepancies refine future cost assumptions. A correction to an accounting definition can identify every dependent signal. This is how learning becomes operational rather than remaining in a researcher’s memory or an unstructured archive.

Foundation models will change, and their general capabilities will become available to competitors. The more durable work is the integration of evidence histories, domain-specific evaluations, research judgment and tested implementation. Even that advantage must be maintained: stale memory, unchecked model upgrades and unreviewed strategy drift can turn accumulated knowledge into accumulated error.

Base One applies this vision to its own capital across public and private markets, including selective investments in startups. Research, diligence and allocation form one institution, with methods adapted to the liquidity, information and ownership characteristics of each investment. Our ambition is the reach of a large, continuously operating research network with the coherence of a single accountable decision process.

The thesis ultimately rests on three achievements: reliable research at lower cost, information that improves decisions, and disciplined conversion of that information into portfolio value. We aim to earn all three through a system that can explain what it knows, reproduce how it learned it and recognise when capital should remain uncommitted.

Research first.
Conviction earned.