PIÑATA SUPERINTELLIGENCE

​

Edge Superintelligence for Automated Equity Research and

Analysis

​

A Localized, Neuro-Symbolic Investment Intelligence System for Quantitative and Qualitative

Evaluation

​

Adam Thomas Contreras

​

Automated Equity Research and Analysis | Investment Intelligence

April 2026 · Revised October 2026

​

WHITE PAPER THESIS Piñata Superintelligence is a proprietary, locally deployed investment

intelligence system that combines quantitative analysis, qualitative intelligence, and deep

reasoning to systematically evaluate businesses and investments.

​

A note on terminology. Throughout this paper, "Superintelligence" and "SI" refer to the

constrained, locally deployed Domain-Specific Superintelligence described in Sections 2 through

5. The term is a descriptive label for the system's role as a high-capacity evidence processor

within a single domain: equity research. It does not claim general, fully autonomous, or

independent investment judgment. Every SI output in this architecture is bounded by explicit rules

and reviewed by a human decision-maker.

​

A note on ownership and use. Piñata SI is proprietary, internal infrastructure, built and used

exclusively by The Blacksmith Fund Limited Partnership. It is not a commercial product, is not

offered to third parties, and is described here to explain the fund's research approach.

 

EXECUTIVE ABSTRACT

Modern equity research operates under a fundamental contradiction. The volume of corporate

disclosure, macroeconomic data, and alternative information has grown beyond the practical

cognitive bandwidth (the finite mental energy and focus a human analyst can sustain before

quality degrades) of any individual analyst. Yet purely quantitative models routinely fail to capture

the qualitative characteristics of a business that can distinguish a compounding franchise (a

business that reinvests profits at high rates of return year after year) from a value trap (a stock that

looks cheap on paper but is cheap for good reason, such as a decaying business).

 

This paper presents the concept and architecture of a hybrid human-SI investment research

system. It is a Domain-Specific Superintelligence operating on localized edge infrastructure:

powerful, custom local compute that keeps proprietary research entirely within the operator's

Piñata Superintelligence · White Paper · Revised October 2026 1controlled environment. Its design is a neuro-symbolic hybrid architecture, pairing deterministic,

rules-based quantitative logic with constrained neural reasoning. The system integrates a

proprietary valuation engine, grounded in fundamental value investing frameworks (disciplined

methods for estimating what a business is worth and buying only at a meaningful discount), with

locally deployed SI constrained to function as a qualitative signal extractor (a tool that reads

documents and pulls out structured findings, rather than making decisions). These components

culminate in a structured human judgment layer for final investment evaluation and capital

allocation.

​

The central thesis is not that SI should replace the investor. Rather, when properly constrained, SI

can compress the qualitative research cycle without sacrificing analytical rigor.

This allows the human analyst to devote greater attention to strategic synthesis, judgment, and

investment decision-making, rather than the repetitive process of parsing and organizing

documents.

​

The result is a research architecture designed to combine the scale and speed of machine

intelligence with the contextual judgment of the human investor.

​

Architectural Development in Brief

The architecture was developed iteratively through working prototypes rather than as a purely

theoretical exercise.

​

In 2025, an initial quantitative layer was built to evaluate securities against fundamental

value-investing criteria. A qualitative research prototype was then developed to process company

disclosures and add contextual analysis to the quantitative output.

​

To test the architecture's range, two public companies of very different scale and complexity were

used as data ingestion and signal extraction stress-tests:

​

• Twin Disc Incorporated (TWIN), a small-cap industrial manufacturer with a market

capitalization (total market value of all outstanding shares) in the tens of millions of dollars. It

served as a test of whether the system could parse a smaller company's annual report and

extract competitive positioning, risk factors, and demand signals.

• Citigroup (C), a multi-billion-dollar global financial institution. It served as a test of whether the

same extraction framework could remain useful as the size, complexity, and information volume

of the underlying business increased.

​

The purpose of these tests was architectural. They measured processing throughput, extraction

quality, and cognitive compression (reducing days of manual reading to a short, reviewable

synthesis). They were not evaluations of investment outcomes.

​

This paper reports no investment performance. Nothing in it should be read as a representation of

returns, or as evidence that the described system will be profitable.

Piñata Superintelligence · White Paper · Revised October 2026 2SECTION 1 — THE ANALYST BOTTLENECK

 

PROBLEM

Traditional fundamental equity research is, at its core, a labor-intensive information compression

problem (the work of reducing enormous volumes of raw documents into a small number of

actionable conclusions). An analyst covering a universe of companies must process hundreds of

annual reports, quarterly filings, earnings call transcripts, proxy statements (documents describing

executive pay and shareholder votes), and industry data, while maintaining enough cognitive

bandwidth to form original, differentiated investment judgments.

The uncomfortable truth of this model is that it relies entirely on a human being, and human beings

are structurally imperfect instruments for this task.

​

• Human analysts get tired. Fatigue degrades the quality of judgment in ways that may not be

apparent to the analyst.

• Humans are inconsistent. The same analyst applying the same framework to two similar

companies in the same week may weigh qualitative factors differently depending on cognitive

load, recent experience, or other contextual conditions.

• Humans make errors under time pressure. Missed footnotes, misread figures, overlooked

risk disclosures, or incomplete consideration of relevant information can turn a well-intentioned

thesis into a capital-destroying mistake.

• Human bandwidth is finite. No matter how talented, a single analyst cannot maintain genuine

analytical depth across a large number of active research assignments at once.

This is not a criticism of analysts. It is a description of biological and cognitive constraints.

The constrained SI system at the core of this platform does not experience fatigue. It does not lose

consistency simply because it is evaluating the fifteenth company in a single research session. Its

attention does not degrade with workload, and its capacity to process large volumes of information

does not approach a ceiling in the way human attention does. This structural asymmetry is the

foundation of the bottleneck thesis.

​

When the cost of thorough qualitative research becomes prohibitively high for human analysts,

stocks that are quantitatively attractive but qualitatively complex can become systematically

under-researched. This creates conditions under which SI-assisted fundamental analysis may

surface opportunities that receive insufficient analytical attention, not by predicting the future, but

by covering more ground, more consistently, than a human-only research operation can

reasonably achieve.

​

SECTION 2 — THEORETICAL FOUNDATION

2.1 · The Bottleneck as Structural Market Inefficiency

The analyst bottleneck is not merely an operational inconvenience. It can represent a structural

source of market inefficiency (a persistent gap between a company's price and its underlying

Piñata Superintelligence · White Paper · Revised October 2026 3value, caused by a lack of careful attention).

When the cost of thorough qualitative research is prohibitively high, companies that are

quantitatively cheap but qualitatively complex may remain systematically under-researched. This

creates conditions under which disciplined, SI-assisted fundamental analysis may identify

situations that receive insufficient analytical attention.

Value investing has long rested on the principle of buying securities at a meaningful discount to a

conservative estimate of what the underlying business is worth. Its foundation is identifying

securities whose fundamentals support a valuation that the market price does not fully reflect.

The innovation proposed here is to apply a similar principle to the research process itself. Rather

than limiting the search to companies that are easy for a human analyst to understand, the system

seeks to identify quantitatively attractive companies and then deploy constrained SI to perform the

qualitative evaluation necessary to investigate them more deeply.

In this framework, the objective is not simply to find companies that appear cheap. It is to

systematically investigate whether the apparent discount survives qualitative examination.

Quantitative screening identifies where further investigation may be warranted. Qualitative

analysis determines whether the underlying investment case remains intact.

2.2 · Constrained Superintelligence (SI) in Financial Analysis

Standard large language models (software trained on vast amounts of text to generate human-like

writing) present two critical challenges in financial analysis:

• Hallucination (the generation of plausible-sounding but false claims, such as inventing a

number that appears in no filing).

• Staleness (the model's training data may not contain current company filings or other recent

information).

The architecture described in this paper is designed to address both limitations. It constrains SI to

verified source evidence, so that outputs can be traced back to source material and subjected to

human review.

​

Crucially, this paper advocates for local deployment of the SI system. Operating as a

Domain-Specific Superintelligence on localized edge infrastructure, on hardware the operator

physically controls rather than through an outside provider, reduces data-exfiltration risk (the

chance that proprietary information leaks to a third party), removes dependence on external

vendors, and allows proprietary research processes and insights to remain within the operator's

controlled environment.

​

For a serious investment research operation, maintaining control over proprietary analytical

information is a fundamental architectural requirement.

Piñata Superintelligence · White Paper · Revised October 2026 4SECTION 3 — GENESIS: THE ARCHITECTURAL

PROOF-OF-CONCEPT

​

The system's development began in April 2025, not in a laboratory, but through hands-on

construction and iterative testing of a working research pipeline.

The catalyst was a direct question:

Could a single operator, using locally deployed constrained SI in conjunction with automated

financial screening, compress weeks of fundamental research into hours, while preserving the

rigor and human oversight that serious analysis requires?

​

3.1 · The First Pipeline

The first deployment was deliberately primitive. A classical, traditional computer script was built to

ingest financial statement data across the U.S. equity universe. For each company, the pipeline

calculated a baseline valuation and compared the result with the prevailing market capitalization to

produce a margin-of-safety assessment (the discount between what an asset is truly worth and

what the market is currently charging, which protects against downside error).

The initial output consisted of thousands of securities. Successive filtering passes then narrowed

the universe to a shortlist numbering in the dozens, removing financially distressed companies,

illiquid securities, and securities whose valuations appeared disconnected from underlying

fundamentals.

​

This represented the first stage of the research funnel: broad quantitative identification followed by

progressively deeper evaluation.

The SI was introduced separately, as a rough prototype built specifically to streamline qualitative

research. Initially it assisted the operator with the manual queries that qualitative review demands.

It then evolved, step by step, into a multi-agent system that replaced those manual human queries.

​

3.2 · Stress-Test One: Small-Cap Industrial Filing (Twin Disc Incorporated)

One company that emerged from the screening process was Twin Disc Incorporated (NASDAQ:

TWIN), a manufacturer of power transmission equipment with a market capitalization in the tens of

millions of dollars. It was selected as a data ingestion and signal extraction stress-test on a

smaller-capitalization industrial company.

Rather than requiring the human operator to manually read the full annual report (the 10-K, a

company's comprehensive yearly filing with the SEC), the constrained SI system was queried

against the document. The test measured whether the system could:

• Extract competitive positioning language from the business description.

• Identify potential risk factors from the risk disclosures.

• Surface management commentary bearing on industrial demand conditions.

Piñata Superintelligence · White Paper · Revised October 2026 5• Compress this synthesis into a form a human reviewer could evaluate quickly.

In this test, the system processed the filing and produced a structured, source-traceable synthesis

(every finding linked to the passage it came from) that the human operator could review in under

two hours. A review of this kind can otherwise consume several days of manual analysis.

The measured outcomes were throughput (how fast a document is processed), extraction fidelity

(how accurately findings reflect the source text), and cognitive compression (how much human

reading time is saved).

​

3.3 · Stress-Test Two: Mega-Cap Financial Institution (Citigroup)

A second stress-test applied the same framework to Citigroup (NYSE: C), a substantially larger

and more complex company. The question was architectural:

Can the same qualitative signal extraction framework extend from a small-cap company with a

market capitalization in the tens of millions of dollars to a large financial institution with a market

capitalization in the multi-billion-dollar range?

A large financial institution presents a different class of challenge: complex balance-sheet

structure (the layered assets, liabilities, and regulatory capital of a global bank), extensive

regulatory disclosure, and far greater information volume. The test examined whether the system

could parse these materials and organize them into the same structured, reviewable format,

without losing source traceability.

Together, the two stress-tests demonstrated that the extraction architecture could be applied

across companies of materially different scale and complexity. They were evaluations of system

capability, not of investment results.

​

SECTION 4 — SYSTEM ARCHITECTURE: THE THREE-STAGE

FUNNEL

The operational architecture is a three-stage funnel that progressively concentrates an investment

universe of thousands of securities into a small set of thoroughly vetted candidates for human

evaluation.

Each stage applies a distinct mode of intelligence: deterministic quantitative computation,

constrained qualitative extraction, and structured human judgment. The stages operate in strict

sequential dependency, with each layer narrowing and refining the output of the one before it.

UNIVERSE (thousands)

|

v

STAGE 1 · Deterministic Engine -- classical code

|

v

STAGE 2 · SI Qualitative Layer -- constrained evidence extraction

|

v

Piñata Superintelligence · White Paper · Revised October 2026 6STAGE 3 · Human Judgment Node -- final evaluation

​

STAGE 1 · QUANTITATIVE VALUATION ENGINE

Proprietary models · Margin-of-safety determination · Universe compression

The first stage is classical code: a high-throughput, strictly deterministic filter. It ingests financial

statement data across the U.S. equity market and applies proprietary models blending qualitative

and quantitative analysis to derive a conservative margin of safety based on fundamental

economic moats. The result is a margin-of-safety score (a number expressing the gap between

estimated intrinsic value and market price) for each company in the universe.

The output is a ranked shortlist, sorted according to the difference between estimated intrinsic

value and current market capitalization.

Critically, no SI inference touches this layer.

The valuation logic is a deterministic valuation (rules-based math that always outputs the exact

same answer from the same inputs, leaving no room for guesswork). It is reproducible and

auditable (every result can be re-run and checked step by step), keeping the quantitative

foundation independent of probabilistic language-model behavior.

​

STAGE 2 · SI-ASSISTED QUALITATIVE LAYER

Constrained SI · Document signal extraction · Economic adjustment engine

Stage 2 is where the system's core innovation operates.

The quantitative shortlist contains companies that appear attractive based on the numbers, but

numbers primarily describe historical financial performance. Stage 2 asks a different question:

Does the economic reality behind those numbers support the investment thesis?

The constrained SI system operates against company reports, including annual filings, quarterly

updates, earnings transcripts (written records of management's quarterly calls with investors),

investor presentations, and other relevant disclosures.

SI is prompted to extract structured qualitative signals across several proprietary evaluation

frameworks. Each signal is accompanied by measures of how well-supported and how current the

underlying evidence is.

​

SI does not assign a final valuation number. Qualitative evidence is permitted to inform the

analytical process only through structured, rule-bound channels, never by letting unconstrained

generative output override the underlying deterministic framework. The speculative generative

capacity of the language model is therefore bounded by explicit rules. Qualitative outputs are

treated as evidence and structured signals, not as autonomous valuation conclusions.

The system maintains disciplined controls and a complete audit trail, so that outcomes can be

traced and reviewed over time.

Piñata Superintelligence · White Paper · Revised October 2026 7STAGE 3 · HUMAN JUDGMENT NODE

Fundamental value investing frameworks · Narrative evaluation · Final conviction scoring

The final stage is where machine-generated research meets trained human judgment.

The operator reviews the SI-generated synthesis not as an oracle, but as a research assistant's

first draft. The human:

• Validates narrative coherence (whether the story the numbers and documents tell actually

hangs together).

• Identifies potential value traps.

• Evaluates competitive dynamics and timing.

• Challenges unsupported conclusions.

• Makes the final decision.

The Twin Disc stress-test illustrates how this stage functions as a workflow. The quantitative

screen identified the company as attractively valued. The SI layer investigated the underlying

disclosures and organized the relevant qualitative signals. The human operator then reviewed the

synthesis and applied independent judgment.

The workflow compressed a research process that could have consumed several days of manual

analysis into approximately two hours of review, while preserving the human oversight required for

final conviction.

​

SECTION 5 — SOFT JUDGMENT CODIFICATION: THE ECONOMIC

ADJUSTMENT ENGINE

The next architectural challenge is ambitious: how can the qualitative reasoning that separates a

competent analyst from an exceptional one be systematically represented?

This is the domain of Soft Judgment Codification: the structured extraction of tacit investment

knowledge (know-how that experts apply by instinct and rarely write down) into reusable and

auditable reasoning components.

The underlying insight is that much of the qualitative reasoning performed by skilled investors,

while appearing intuitive, is also patterned.

Management quality inference, narrative credibility evaluation, competitive moat assessment

(judging how well a business is protected from competitors), capital allocation analysis (how well

management uses the company's cash), and other qualitative judgments are not necessarily

random acts of intuition. They often reflect learned heuristics (mental shortcuts refined by

experience) that can be partially formalized, encoded, tested, and applied consistently.

Piñata Superintelligence · White Paper · Revised October 2026 8The objective is not to reduce investment judgment to a single formula. It is to identify the recurring

analytical patterns within experienced fundamental investing and convert them into structured

components that can be evaluated systematically.

​

5.1 · Mitigating Human Cognitive Bias at Scale

One of the most important properties of this codification framework is not simply efficiency. It is

consistency.

Human analysts, regardless of skill, bring unavoidable cognitive biases to qualitative evaluation:

• Recency bias (overweighting the latest information) can cause recent management

commentary to receive disproportionate weight.

• Confirmation bias (favoring evidence that supports what you already believe) can influence

how ambiguous disclosures are interpreted.

• Anchoring (over-relying on the first number or idea encountered) can distort how new

information changes an existing thesis.

These are not character flaws. They are features of human cognition under uncertainty.

A constrained SI system operating within a structured, rule-based evaluation framework can help

mitigate the influence of these biases within the qualitative translation process. When the system

evaluates management quality, for example, it applies consistent criteria across companies rather

than shifting its methodology because of fatigue, recency, or emotional context. This supports

consistency enforcement and reduces heuristic drift (the gradual, unnoticed change in how an

analyst applies their own standards over time).

This does not eliminate bias. Language models can carry biases of their own, and rule-based

frameworks reflect the assumptions of their designers. Mitigation depends on testing, human

review, and continuous recalibration.

Properly constrained, SI does not merely accelerate the work. It can make the analytical process

more consistent and help expose systematic sources of human error.

At scale, greater consistency across thousands of evaluations could become a structural

advantage over a purely human research operation.

​

5.2 · Evolution of the Factor Framework

The system's qualitative layer has evolved significantly from its April 2025 origins.

The platform now operates a proprietary factor framework (a structured set of measurable

qualities, such as management track record or capital discipline) spanning multiple analytical

domains. It is the product of sustained development: translating the tacit heuristics of experienced

fundamental investors into rigorous, measurable, and continuously validated components. Factors

are admitted only where they demonstrate both statistical relevance and a clear, defensible

economic rationale.

​

Piñata Superintelligence · White Paper · Revised October 2026 95.3 · Multi-Agent Orchestration Architecture

The system has evolved from a single-model interface into a multi-agent orchestration architecture

(breaking a massive research task into specialized digital workers, where each agent masters one

specific job). This architecture replaced the manual human queries of the earliest prototype.

Rather than relying on one SI system to perform every qualitative function, the platform routes

analytical tasks to specialized agents, each operating against targeted data sources and

domain-specific frameworks.

The data sources ingested by the platform have expanded well beyond annual reports to a broad

range of corporate disclosures, regulatory filings, insider transaction data (records of company

executives buying or selling their own stock), earnings transcripts, and other structured and

unstructured sources. Each source is treated as a distinct signal channel.

An independent review layer challenges weak, inconsistent, or contradictory evidence before

findings are presented to the human operator. This additional layer is intended to further reduce

the risk of misinterpretation before qualitative findings influence the evaluation.

​

SECTION 6 — SYSTEM RISK ANALYSIS & FAILURE MODES

6.1 · Regime Dependence

Deep-value strategies (approaches that favor out-of-favor, inexpensive companies) are historically

cyclical. In high-liquidity, growth-driven market regimes, quantitative screens can systematically

surface companies that continue to underperform in the near term.

The system therefore uses several proprietary analytical frameworks designed to evaluate

companies across different market regimes and valuation environments. The objective is to reduce

dependence on any single methodology or market regime.

6.2 · Data Quality and Historical Bias

The quantitative screening parameters were initially calibrated against historical financial data

(tuned to how past data looked).

Structural shifts in reporting standards, accounting rules, disclosure practices, or the way

companies classify assets can cause screening logic to misclassify companies.

• Regular recalibration against new data is a standing operational requirement.

• The system must be evaluated not only on whether its historical models produced attractive

results, but on whether those models remain economically meaningful when applied to new

data and changing market conditions.

6.3 · Compute and Latency Constraints

Operating a powerful local SI creates real constraints on throughput. Processing a full company

shortlist through the qualitative layer can require substantial time.

Piñata Superintelligence · White Paper · Revised October 2026 10This may be acceptable for a single-operator research system, but it becomes a potential

bottleneck at institutional scale. Expanded local compute, improved efficiency, and optimized

orchestration therefore represent known infrastructure requirements on the development roadmap.

 

SECTION 7 — EXPANSION PROTOCOLS & INFRASTRUCTURE

ROADMAP

7.1 · Extended Valuation Model Suite

The expansion roadmap extends current valuation coverage to a broader suite of proprietary

models, each calibrated to evaluate companies across different business types and economic

structures.

The objective is a multi-perspective analytical framework rather than reliance on a single valuation

lens.

7.2 · Expanded Data Ingestion Protocols

The data ingestion layer is being systematically expanded to incorporate a wider range of

regulatory, governance, insider, communication, and alternative data (non-traditional information

sources).

Each new source category is admitted to the platform only after validation that it adds genuinely

new information.

7.3 · Automated Macro Regime Filtering

A macro overlay module will allow the system to adapt to the prevailing market and credit

environment. The objective is to prevent the quantitative screen from mechanically surfacing

deep-value candidates in environments where the structural conditions for value realization may

be impaired.

7.4 · Anomaly Detection Layer

An automated anomaly detection layer will examine company financials for signs of accounting

irregularity, including:

• Aggressive revenue recognition (booking sales earlier than is prudent).

• Unusual accrual patterns (gaps between reported profit and actual cash).

• Abnormal working-capital movements (odd swings in receivables, inventory, or payables).

• Related-party transaction activity (dealings between the company and insiders or affiliates).

This layer is designed to scale readily to comprehensive forensic testing. Its purpose extends

beyond calculating a margin of safety: it enriches the overall qualitative picture of every evaluation,

giving the human operator a fuller view of financial reporting quality.

Piñata Superintelligence · White Paper · Revised October 2026 117.5 · Adversarial Committee Architecture

For each investment thesis, a structured adversarial debate layer will deploy competing analytical

perspectives, constructing the strongest independent cases both for and against a potential

position, before human review.

Their outputs will be reconciled by a synthesis layer that presents the human operator with an

explicit map of agreement, disagreement, and the evidence underlying each position.

This architecture is intended to approximate the challenge function of an institutional investment

committee and to systematically reduce the confirmation bias that can arise when a single

analytical model evaluates its own conclusions.

 

SECTION 8 — SCALABILITY & FUTURE MACHINE LEARNING

OPERATIONS

A skilled fundamental analyst operating at full capacity, with genuine depth and rigor, can

realistically maintain active coverage of perhaps a dozen companies simultaneously. Adding more

companies does not necessarily add proportional insight. Beyond a certain point, it introduces

dilution of attention and analytical depth.

This is not readily resolvable within a human-only research paradigm. It is fundamentally

constrained by human time, attention, and cognitive capacity.

The architecture described in this paper operates under a different constraint set:

• The quantitative engine can screen an investable universe of thousands of companies with

consistent application of the same computational rules.

• The qualitative SI layer does not experience fatigue when processing its thousandth company.

• The human operator is engaged primarily at the final judgment node, reviewing a

pre-synthesized, evidence-grounded research packet rather than manually processing every

underlying document.

The marginal cost of covering an additional company within the quantitative screening layer

approaches zero once the infrastructure is established. The qualitative layer substantially reduces

the human time required for deeper investigation.

Deep fundamental analysis, which might require several days of senior analyst time in a traditional

operation, can potentially be compressed to hours.

This changes the economics of fundamental research by making a broader level of analytical

coverage possible within a much smaller operating structure.

8.1 · Future Machine Learning Operations

Piñata Superintelligence · White Paper · Revised October 2026 12The current architecture represents the first generation of a research platform designed to

compound its capabilities over time.

As the system accumulates a structured history of analysis and decisions, that record becomes a

potential training substrate (a body of examples the system can later learn from) for future

machine learning operations.

The roadmap includes meta-cognitive capabilities (systems that evaluate the performance of the

analytical process itself), allowing the platform to monitor its own signals, assess how well its

stated confidence matches actual accuracy, and propose refinements for human review.

Beyond self-improvement, the platform will develop deeper signal-research capabilities, studying

which qualitative patterns have historically been informative about business outcomes and capital

allocation behavior.

At the frontier of this roadmap are experimental regime-detection functions, designed to identify

structural transitions in market behavior, competitive dynamics, or macroeconomic conditions

before those transitions become fully reflected in consensus data (the widely shared view of

analysts and the market).

The long-term vision is a research platform that compounds structured knowledge, helps mitigate

systematic sources of human bias, operates without cognitive fatigue across a broad universe of

companies, and continuously improves through feedback from its own accumulated analytical

record.

SECTION 9 — CONCLUSION: THE INFRASTRUCTURAL ELEVATION

OF THE ANALYST

The central claim of this paper is architectural: Superintelligence, when properly constrained within

a deterministic valuation framework and supervised by trained human judgment, can address the

analyst bottleneck without compromising analytical rigor.

The April 2025 proof-of-concept was not presented as evidence that SI can select investments. It

was an early demonstration that constrained SI can compress the research cycle to a speed at

which a disciplined investor can investigate more companies, maintain analytical depth, and make

informed decisions from a broader evidence base.

This paper makes no claim about investment outcomes, and the architecture's usefulness as a

research tool should not be taken as evidence of future profitability.

The broader significance lies in the architecture: a localized, neuro-symbolic, Domain-Specific

Superintelligence that is difficult to replicate because it is the product of sustained, iterative,

hands-on development rather than any single technique.

The future of investment research is not necessarily algorithmic replacement of the analyst. It is

the infrastructural elevation of the analyst.

Piñata Superintelligence · White Paper · Revised October 2026 13By automating the repetitive and cognitively expensive work of document ingestion, evidence

extraction, organization, and signal identification, the platform seeks to free human judgment for

the work that remains most difficult to automate: asking better questions about competitive

advantage, capital allocation, business quality, valuation, and the durability of economic moats

(the lasting advantages that protect a business from competition).

THESIS IN ONE SENTENCE

Constrained Superintelligence (SI), bounded by deterministic valuation guardrails, grounded in

verified source evidence rather than unrestricted generative inference, and supervised by human

judgment grounded in fundamental value-investing frameworks, addresses the analyst bottleneck

by reducing the cost and time required for institutional-grade equity research and creating a

scalable, SI-augmented research infrastructure.

LEGAL DISCLAIMER & IMPORTANT DISCLOSURES

Please read carefully before relying on any information in this document.

FOR EDUCATIONAL AND INFORMATIONAL PURPOSES ONLY. This document is an

academic and technical systems white paper, produced solely for educational and informational

purposes. Nothing contained herein constitutes investment advice, financial advice, legal advice,

tax advice, or any other form of professional advice. The information presented describes an

experimental research architecture and workflow concept. It is not intended to serve as the basis

for any investment decision.

NOT AN OFFER OR SOLICITATION OF SECURITIES. This document does not constitute an

offer to sell or a solicitation of an offer to buy any security, investment product, or financial

instrument of any kind in any jurisdiction. Any such offer or solicitation, if made, would be made

only pursuant to formal offering documents that comply with applicable federal and state securities

laws.

NO PERFORMANCE REPRESENTATION. This document does not report, and should not be

read as implying, any investment performance, returns, or profitability of any account, portfolio,

fund, or strategy. References to specific public companies, including Twin Disc Incorporated and

Citigroup, are made solely to illustrate the technical operation of a document-processing and

signal-extraction system. They are not recommendations and do not represent holdings, trades, or

results. There is no representation, warranty, or guarantee that any methodology, research

architecture, or process described herein will achieve profitability or avoid losses. All investments

involve risk, including the possible loss of principal.

DESCRIPTION OF TECHNOLOGY. References to "Superintelligence" and "SI" are descriptive

labels for the constrained, locally deployed, domain-specific systems described in this paper. They

do not represent that the system possesses general intelligence, acts autonomously, or can

independently make investment decisions. Descriptions of system capabilities, including speed,

scale, and consistency, reflect design objectives and limited testing. They are not guarantees of

Piñata Superintelligence · White Paper · Revised October 2026 14performance, accuracy, or reliability. SI outputs may contain errors and require human review.

NO INVESTMENT ADVICE; CONSULT A QUALIFIED PROFESSIONAL. Nothing in this

document should be construed as a recommendation to buy, hold, or sell any specific security or

to adopt any particular investment strategy. Readers should consult a qualified financial

professional before making investment decisions.

PROPRIETARY INFORMATION & INTELLECTUAL PROPERTY. The methodologies,

frameworks, system architectures, factor libraries, analytical processes, and related intellectual

property described herein are proprietary to the author and to The Blacksmith Fund Limited

Partnership and are protected under applicable intellectual property laws. No portion of this

document may be reproduced, distributed, transmitted, or published without the express prior

written consent of the author, except as permitted by applicable law.

FORWARD-LOOKING STATEMENTS. Certain statements in this document, including statements

about planned capabilities and the development roadmap, may constitute forward-looking

statements. These statements involve known and unknown risks and uncertainties that may cause

actual results to differ materially from those expressed or implied. The author undertakes no

obligation to update or revise any forward-looking statements.

© 2026 Adam Thomas Contreras. All Rights Reserved. Piñata SI · SI-Native Equity Research

Piñata Superintelligence · White Paper · Revised October 2026 15