👩‍🏫Proof of Alpha Network

Astros Research · Whitepaper · August 2026

Abstract

Public financial discourse produces millions of market opinions, yet the platforms that host this discourse measure only attention: followers, impressions, and engagement. None of these signals indicates whether an author has demonstrated repeatable insight. This paper introduces the Proof of Alpha Network, a framework that converts unstructured public market commentary into timestamped, testable records and evaluates each record against subsequent market outcomes within the context in which the opinion was made. The system maintains a continuously updated, context specific reputation for every source, organized as a graph that connects people, opinions, assets, events, and outcomes. We describe three technical foundations of the framework. First, the Alpha Oracle, an ensemble mechanism in which multiple independent artificial intelligence models act as redundant judges whose weighted agreement improves the accuracy and manipulation resistance of opinion extraction and evaluation. Second, an information theoretic account of how structured extraction compresses noisy discourse into high density evidence that fits within the narrow bandwidth of human decision making. Third, an analysis of how the same compressed records serve as an efficient interlingua for agent to agent networks, reducing token consumption by one to two orders of magnitude. The framework does not promise future returns; it transforms market reputation from a social claim into a continuously updated body of evidence.

The framework does not promise future returns; it transforms market reputation from a social claim into a continuously updated body of evidence.

Index Terms: Reputation systems, information entropy, ensemble learning, large language models, multi-agent systems, financial social networks, forecasting evaluation.

I. Introduction: The Conviction Gap

Trading infrastructure has spent decades optimizing execution. Exchanges compete on liquidity, spreads, latency, fees, and order types. These improvements matter, but they begin only after a user has already decided what to trade.

For many retail participants, the larger friction occurs earlier. A user may already be watching Bitcoin, a semiconductor equity, or gold. They have access to more information than they can process — a condition Simon identified half a century ago when he observed that a wealth of information creates a poverty of attention [1]. What such users often lack is not a buy button but sufficient conviction to press it.

Classical market theory assumes that prices rapidly incorporate available information [2], while Grossman and Stiglitz showed that some inefficiency must persist to compensate those who gather and interpret information [3]. Behavioral research adds that individual decisions under risk deviate systematically from rational benchmarks [4], and that retail investors disproportionately buy assets that capture their attention rather than assets supported by evidence [5]. Narratives, not verified track records, drive much of the resulting behavior [6].

The retail decision process is therefore better represented not as a single act of choice but as a pipeline: attention → interpretation → trust → conviction → action.

At a conceptual level, conviction can be expressed as:

$$C_{u,a,t} = f(P_{u,a,t}, E_{a,t}, R_{i,c,t}, S_{a,t}) \tag{1}$$

where $C_{u,a,t}$is user $u$'s conviction in asset $a$ at time $t$; $P_{u,a,t}$ is the user's prior belief; $E_{a,t}$ is the available market evidence; $R_{i,c,t}$ is the contextual reputation of source $i$ in context $c$; and $S_{a,t}$ is the level of social confirmation or disagreement. A trade occurs when conviction crosses a personal action threshold:

A trade occurs when conviction crosses a personal action threshold:

$$C_{u,a,t} > \theta_u \tag{2}$$

Equations (1) and (2) are not production formulas. They are a model of the product problem. Information alone rarely creates action; information becomes actionable when it is interpreted through trust. A bullish argument from an unknown account may have little effect. The same argument from a source with a strong historical record in semiconductors may materially change the user’s belief, and if several credible sources independently reach the same conclusion, conviction may rise further. Retail trading therefore resembles continuous belief updating in which trust weighted evidence revises a prior. This is the gap the Proof of Alpha Network is designed to address. The remainder of this paper is organized as follows. Section II reviews related work. Sections III through V describe how opinions become structured records, how reputation is computed and decays, and how the resulting opinion graph operates. Section VI describes the role of artificial intelligence as measurement infrastructure, and Section VII extends it with the Alpha Oracle, a multiple model consensus mechanism. Section VIII analyzes the framework through the lens of information entropy and human cognitive bandwidth. Section IX quantifies the benefits of structured records for agent to agent networks. Section X describes the product surfaces, Section XI discusses limitations and manipulation resistance, and Section XII concludes.

II. Related Work

The idea that aggregated independent judgments can outperform individual experts has a long lineage, from popular accounts of collective intelligence [7] to the empirical literature on prediction markets, which shows that prices formed by many participants forecast events well when incentives are aligned [8].

Tetlock's research on forecasting demonstrates that individual predictive skill exists, is measurable, and is concentrated in people who make precise, falsifiable, and regularly updated forecasts [9]. Prelec's Bayesian truth serum shows that truthful reporting of subjective judgments can be incentivized even without ground truth at the moment of reporting [10].

A parallel literature applies computational text analysis to finance. Loughran and McDonald established that domain-specific dictionaries are required to interpret financial language correctly [11], and Bollen, Mao, and Zeng provided early evidence that aggregate social media sentiment carries information about subsequent market movements [12]. Modern transformer architectures [13] have made it feasible to extract far richer structure from free text than dictionary methods allow, including implied assets, direction, conviction, and time horizon.

Proof of Alpha differs from these strands in its object of measurement. Prediction markets price events; sentiment systems price crowds; Proof of Alpha prices sources. Its output is not a forecast but a continuously updated, context-conditioned record of who has demonstrated insight, in what domain, and over what horizon.

III. From Attention to Reputation

Social platforms are effective at measuring distribution. Follower count measures audience size, likes measure reaction, and reposts measure reach. None of them directly measures whether a person has historically demonstrated insight. A large audience does not prove alpha. A viral prediction does not establish consistency. A correct call remembered after the fact does not reveal how many incorrect calls were forgotten, deleted, or reframed.

A. The Structured Opinion Record

Proof of Alpha begins by converting each public market opinion into a structured record:

$$O_j = (i, a, d, t, h, q) \tag{3}$$

where $i$ is the source, $a$ the relevant asset or market, $d$ the direction (bullish, bearish, or neutral), $t$ the original timestamp, $h$ the expected time horizon, and q the clarity or confidence of the opinion. The distinction between vague and measurable statements matters. “BTC looks interesting” is not equivalent to “BTC will outperform over the next thirty days.” A measurable opinion requires enough specificity to establish what was claimed, when it was claimed, and over what period it should be evaluated. This mirrors the finding that forecasting skill can only be assessed when forecasts are precise and time bound [9]. Once an opinion is structured, market outcomes can be attached to it. The system can then evaluate not only whether the direction was correct but also entry timing, time horizon, maximum adverse movement, magnitude of the move, consistency across repeated calls, whether the original thesis was maintained or changed, and performance relative to an appropriate benchmark.

B. Reputation as a Contextual Vector

The result should not be a single permanent score, because alpha is contextual. Someone may have strong insight into Bitcoin but weak performance in equities. A long-horizon macro analyst may identify major trends while performing poorly on short-term timing. A trader who excels during momentum regimes may struggle in sideways conditions.

Reputation is therefore represented as a vector:

$$R_i = (r_{asset}, r_{horizon}, r_{regime}, r_{consistency}) \tag{4}$$

In other words, there is no universal alpha; alpha exists within a context. The system should answer not only "Is this person good?" but "What is this person good at, over what time horizon, and under which market conditions?"

IV. Proof is Continuous

Historical performance is useful, but it should not become permanent authority. Markets change, strategies decay, and expertise in one cycle may not transfer to the next. A reputation system must therefore update continuously as new evidence appears. A simplified representation is:

$$R_{i,c,t} = \frac{\sum_j k(c, c_j)\, q_j\, e^{-\lambda(t - t_j)}}{\sum_j k(c, c_j)\, e^{-\lambda(t - t_j)}} \tag{5}$$

where $q_j$ is the evaluated quality of historical opinion $j, k(c, c_j)$ measures how relevant that opinion is to the current context, and the exponential term reduces the influence of older evidence over time.

The principle is more important than the exact equation: recent, relevant, and clearly stated opinions should carry more weight than old, unrelated, or ambiguous ones. A person does not permanently own alpha because of one successful call. Alpha is not owned; it is continuously re-earned.

For this reason, the system must also account for behaviors that weaken the quality of evidence:

• Deleting incorrect opinions

• Editing the meaning of a thesis after the market moves

• Publishing vague calls without a defined direction or horizon

• Repeatedly making contradictory predictions

• Highlighting winners while ignoring losses

• Taking excessive risk to generate a superficially high success rate

"Proof" does not mean certainty. It means that the same transparent measurement framework is applied before the outcome is known.

V. The Opinion Graph

Proof of Alpha is not simply a leaderboard; a leaderboard is one interface built on top of a larger network. The underlying system is a graph connecting people, opinions, assets, events, and outcomes.

A person publishes an opinion about an asset. The opinion is associated with a direction, timestamp, horizon, and thesis. Events either support or challenge that thesis. The eventual market outcome becomes historical evidence, the evidence updates the person's contextual reputation, and the updated reputation affects how future opinions are interpreted.

The loop can be summarized as:

opinion → outcome → reputation → trust → future conviction

This closed loop is what makes Proof of Alpha a network rather than a collection of profiles. Its purpose is to answer a continuously changing set of questions:

• Who believes what, and when did they begin believing it?

• What evidence supports their view?

• What happened after the opinion was published?

• Has this person demonstrated an edge in similar situations?

• Which credible sources agree or disagree?

• Is a thesis gaining support, losing support, or becoming invalidated?

The network creates bidirectional discovery. A user can begin with a person and traverse to their active opinions, the underlying market thesis, and finally a trade. Alternatively, a user can begin with an asset and traverse to the bull and bear theses, the credible supporters of each side, and finally a trade. Both paths operate on the same underlying graph.

VI. AI as Measurement Infrastructure

Artificial intelligence is essential to Proof of Alpha, but not because users should blindly follow machine generated trading decisions.

Its role is measurement. Public market discourse is unstructured: the same person may express a thesis across posts, threads, interviews, charts, and follow up comments; the asset may be implied rather than named; the horizon may be unclear; and the author may later revise, contradict, or abandon the original argument.

Large language models built on the transformer architecture [13] can convert this fragmented information into a structured system by extracting market opinions from public content, identifying the relevant asset, sector, or event, classifying direction and conviction, estimating the intended time horizon, detecting repeated, modified, or contradictory theses, connecting opinions with subsequent market outcomes, organizing bull and bear arguments, summarizing the evidence behind each side, and making the original source material inspectable.

The system must preserve the evidence behind its conclusions. Users should be able to understand why a reputation score changed and which historical opinions contributed to it. The framework should never ask users to trust an unexplained answer.

Artificial intelligence does not create trust; it makes trust measurable. Because measurement itself can err, the next section describes how the framework improves the reliability of machine judgment by treating no single model as a source of truth.

VII. The Alpha Oracle: Multiple Model Consensus

Every conclusion in the Proof of Alpha Network depends on machine judgment at two stages: extraction, in which free text is converted into the structured record of (3), and evaluation, in which a resolved outcome is scored against the original claim.

A single model performing these tasks is a single point of failure. It may misread sarcasm, miss an implied asset, hallucinate a horizon that was never stated, or drift as the underlying model is updated. Worse, a single known judge is a fixed target: an adversary who learns its blind spots can craft language that games extraction.

The Alpha Oracle addresses this by treating no individual model as a source of truth. Instead, multiple independent models act as redundant judges whose weighted agreement constitutes the system's verdict.

A. Theoretical Basis

The design rests on three well-established results:

1. Ensemble diversity reduces error — whenever individual members are better than random and make errors that are not perfectly correlated [14], a principle that also governs the combination of human forecasts [15].

2. The Condorcet jury theorem — if each independent judge is correct with probability greater than one-half, the probability that the majority is correct approaches one as the number of judges grows [16].

3. Sampling and refinement improve accuracy — sampling the same model several times and taking the modal answer measurably improves reasoning accuracy [17], and layering heterogeneous models so that later models refine the outputs of earlier ones improves quality further [18].

The Alpha Oracle applies all three: diversity across model families, redundancy within each model through repeated sampling, and a refinement layer for contested cases.

B. Aggregation Algorithm

Let $M = {m_1, \ldots, m_N}$ be a set of N independent judge models drawn from distinct model families and providers. For a given piece of source content, each judge m_k produces a candidate structured record $O^{(k)} = (a^{(k)}, d^{(k)}, h^{(k)}, q^{(k)})$. Aggregation is performed field by field. For a categorical field such as direction, the oracle output is the weighted plurality

Aggregation is performed field by field. For a categorical field such as direction, the oracle output is the weighted plurality:

$$d^{*} = \arg\max_d \sum_k w_{k,\tau} \cdot \mathbb{1}[d^{(k)} = d] \tag{6}$$

where $w_{k,\tau}$ is the reliability weight of judge k on task type $\tau$ (for example, direction classification on cryptocurrency content), and $\mathbf{1}[\cdot]$ is the indicator function. Numeric fields such as horizon are aggregated by the weighted median, which is robust to a single extreme judgment. Weights are not static. Whenever ground truth becomes available, through resolved outcomes, human audit, or unanimous agreement of all other judges, each judge’s weight is updated multiplicatively in the style of the weighted majority algorithm [19]:

$$w_{k,\tau} \leftarrow w_{k,\tau} \cdot \beta^{\mathbb{1}[\text{judge } k \text{ was wrong}]}, \quad 0 < \beta < 1 \tag{7}$$

This guarantees that the aggregate error is bounded relative to the best judge in hindsight, even if some judges degrade over time [19].

The oracle also computes a disagreement score for every field — the normalized Shannon entropy of the weighted vote distribution:

• Low entropy → judges converge → record is committed automatically

• High entropy → escalation triggered → contested content, together with each judge's cited evidence, is passed to a refinement layer in the style of a mixture of agents [18]

• Persistent disagreement → record is flagged for human review or excluded from reputation scoring altogether

Excluding ambiguous cases is safer than guessing, because a reputation system inherits every error of its measurement layer.

Fault tolerance: if at most $F$ judges may be arbitrarily wrong or even adversarial on a given item, classical results on Byzantine agreement suggest provisioning $N \geq 3f + 1$independent judges so that the correct answer still commands a decisive majority [20]. In practice the oracle runs five to seven heterogeneous judges for routine extraction and enlarges the panel for high-stakes records, such as opinions from sources with large reputational weight.

Algorithm 1: Alpha Oracle Aggregation for One Content Item

Input:  content x, judges M = {m_1..m_N}, weights w, task tau
Output: committed record O* or escalation flag

1:  for each judge m_k in M do (in parallel)
2:      O_k <- extract(m_k, x)     // s samples, take mode
3:  for each field phi in {asset, direction, horizon, q} do
4:      v_phi <- weighted vote over {O_k.phi} using w[k,tau]
5:      H_phi <- normalized entropy of v_phi
6:  if max_phi H_phi <= theta_commit then
7:      O* <- field-wise winners; commit O* with provenance
8:  else if max_phi H_phi <= theta_escalate then
9:      O* <- refine({O_k}, x) via aggregator model; commit
10: else
11:     flag x for human audit; exclude from scoring
12: on later ground truth: update w[k,tau] by Eq. (7)

Beyond accuracy, the oracle has an incentive property. Because the panel is heterogeneous and its composition and weights evolve, no external party can optimize language against a fixed judge. And because every committed record stores the full panel's outputs as provenance, any user can inspect not just what the system concluded but how strongly the judges agreed — in the same spirit as mechanisms that make subjective judgments auditable without ground truth at reporting time [10].

VIII. Information Entropy and Human Bandwidth

The value of extraction can be made precise in information-theoretic terms. Shannon defined the entropy of a source as the expected number of bits required to describe its output [21]:

$$H(X) = -\sum_x p(x) \log_2 p(x) \tag{8}$$

Public market discourse is a high entropy channel in the raw sense but a low density one in the decision relevant sense. Shannon’s own experiments place the entropy of ordinary English near one bit per character [21], so a typical three hundred word thread carries on the order of 1,500 bits of text.

Yet the portion of that text that matters for a trading decision, the fields of the structured record in (3), is tiny. Direction over three classes carries at most $\log_2 3 \approx$ 1.6 bits; a horizon bucketed into eight ranges carries 3 bits; a confidence grade carries about 2 bits; and identifying one asset in a universe of ten thousand carries about 13.3 bits.

The decision relevant content of a long thread is therefore roughly 20 bits embedded in 1,500, a signal density below two percent. The rest is framing, hedging, repetition, and rhetoric, which is informative as narrative [6] but nearly worthless as evidence.

A. Extraction as Rate-Distortion

Extraction is best understood as lossy compression that minimizes description length while preserving the mutual information between the record and the quantities a decision depends on [22], [23]. Formally, the system seeks the shortest representation $O$ of source content $X$ such that $I(O; Y)$ is approximately $I(X; Y)$, where $Y$ denotes the decision relevant claim (what was predicted, about what, by when, and how confidently).

Under this criterion, deleting the rhetoric is not a loss at all: the record achieves a compression ratio near 75 to 1 in bits while retaining essentially all of the evidentiary content. The effective information rate delivered to the reader improves by the same factor:

$$\rho = \frac{I(O; Y)}{L(O)} \gg \frac{I(X; Y)}{L(X)} \tag{9}$$

where $L(\cdot)$ is description length. The Alpha Oracle of Section VII ensures the compression is trustworthy; this section explains why it is necessary.

B. The Human Bottleneck

The necessity comes from a hard biological constraint. Although human sensory systems take in on the order of a billion bits per second, careful measurements across reading, memory, and decision tasks converge on a deliberate cognitive throughput of only about ten bits per second [25]. Working memory compounds the constraint: people can hold and compare only about seven chunks of information at once [24].

At ten bits per second, a fully attentive human processes at most roughly 36,000 deliberate bits per hour. Reading raw discourse at a two-percent signal density, that hour yields only about 700 bits of usable evidence — the equivalent of perhaps thirty-five structured opinions, and in practice far fewer, because attention is diverted to whatever is most salient rather than most reliable [5].

Compression changes what a fixed cognitive budget can buy. If each structured record delivers its roughly 20 decision bits in a two-second glance instead of a several-minute read, the same hour of attention can cover on the order of a thousand distinct pieces of evidence rather than a few dozen — an improvement of one to two orders of magnitude.

Just as important, the reputation vector of (4) tells the user which of those records deserves any budget at all. Entropy reduction and trust weighting are therefore two halves of one mechanism: compression fits the evidence into the channel, and reputation prioritizes what flows through it. This is how the framework converts Simon's poverty of attention [1] from a structural handicap into a solvable allocation problem, and it is why conviction in the sense of (1) can form quickly enough to be actionable.

IX. Structured Records in Agent Networks

The same compression that serves human bandwidth serves machine economics. Autonomous research and trading agents built on large language models increasingly consult one another and shared knowledge sources [27], [28]. When such agents exchange raw social content, every consumer pays the full token cost of every thread it reads — and pays it again on every subsequent read, because model context is not persistent.

Prompt compression research shows that large reductions are possible with little task loss [26]; Proof of Alpha goes further by compressing once, at ingestion, into a canonical schema that every agent can consume without reparsing prose.

A. Token Cost Model

Let $T_{\mathrm{raw}}$ be the token length of a raw thread with context, $T_{\mathrm{rec}}$ the length of its structured record, $T_{\mathrm{ext}}$ the one-time cost of oracle extraction, and $R$ the number of downstream reads across the network. Total consumption under the two regimes is

$$T_{network} = R \cdot T_{raw} \quad \text{versus} \quad T_{ext} + R \cdot T_{rec} \tag{10}$$

With representative values — $T_{raw}$ $\approx 1{,}500$ tokens for a thread plus the replies needed to disambiguate it, $T_{rec} \approx 60$ tokens for the record with provenance pointers, and $T_{ext} \approx 12{,}000$ tokens for a seven-judge oracle pass — the structured regime breaks even after only nine reads and thereafter approaches the asymptotic ratio $T_{raw}/T_{rec} = 25$

A network of one thousand agents each consulting two hundred opinions per day would consume roughly 300 million tokens daily on raw content, versus about 12 million on records plus an extraction budget shared across all consumers — a saving of more than an order of magnitude that grows linearly with the number of readers per record.

B. Qualitative Benefits

1. Fewer parsing errors — a fixed schema eliminates the per-agent parsing errors that arise when each agent interprets prose independently; the network disagrees less because it reads the same fields.

2. Relieves context window pressure — short records allow an agent to hold hundreds of trust-weighted opinions in working context at once, mirroring the human working-memory argument of Section VIII.

3. Precomputed trust prior — the reputation vector of (4) is itself a precomputed trust prior (roughly thirty tokens) that would otherwise require each agent to re-derive a source's history from thousands of tokens of archives.

In effect, Proof of Alpha amortizes the epistemic labor of the entire network: the oracle pays the interpretation cost once, and every human and machine consumer inherits the result at marginal cost. The opinion graph thus functions as shared memory for an agent economy, in which credibility metadata travels with every claim.

X. From Proof to Trade

The network is surfaced through two connected product experiences.

A. Reputation Arena

The Reputation Arena begins with the question: whom should I trust?

Users can discover who has recently performed well, who has remained consistent over time, and who has demonstrated an edge in a particular asset, sector, horizon, or market environment. Leaderboards create distribution, competition, and social discovery, while the historical record provides the evidence beneath the ranking.

B. Market Thesis

Market Thesis begins with a different question: what should I believe about this market?

For a user already watching Bitcoin, a semiconductor equity, gold, artificial intelligence, or interest rates, the product structures:

• The strongest bullish arguments

• The strongest bearish arguments

• The people supporting each side

• Key catalysts and invalidation conditions

• Relevant historical situations

• The credibility of the sources involved

The two interfaces are connected. A user viewing an equity thesis can see which historically credible sources are bullish or bearish, and a user viewing a trader's profile can see their current active theses across markets. This creates a unified flow from discovery → conviction → trade.

Traditional social platforms end with content; research platforms end with information; copy-trading platforms begin with an existing position. Astros connects the reputation and opinion layer directly with execution, so that once a user has evaluated the evidence and formed a view, they can act without leaving the system.

XI. What Proof Does Not Mean

Proof of Alpha is not a promise that historical winners will continue to win. Public opinions do not always represent a person's complete portfolio, true entry price, position size, or private risk management, and market outcomes are influenced by randomness, leverage, liquidity, and changing conditions. Grossman and Stiglitz's argument implies a further caution: any edge that is fully published and fully believed will tend to erode [3].

Measurement systems also invite gaming. When a measure becomes a target, it ceases to be a good measure [29], and a reputation score is a target by construction. The framework's defenses are the ones described throughout this paper:

• Transparent scoring

• Inspectable source material

• Context-specific evaluation rather than a single number

• Penalties for deletion, revision, vagueness, and selective disclosure

• Oracle-level redundancy so that no fixed judge can be reverse-engineered

• Appropriate treatment of risk and horizon

• Explicit handling of uncertainty

• Continuous iteration as manipulation patterns emerge

Proof of Alpha should be treated as evidence, not authority. The system is designed to improve the quality of judgment, not to eliminate the need for judgment.

XII. Conclusion

The internet has built sophisticated systems for measuring attention and weak systems for measuring credibility. Financial markets contain millions of public opinions, yet those opinions are still evaluated through follower counts, screenshots, selective memory, and reputation by association.

Proof of Alpha proposes a different model:

Market opinions become structured records; structured records become historical evidence; historical evidence becomes contextual reputation; and reputation helps both humans and machine agents decide which information deserves weight.

A multiple-model oracle keeps the measurement honest, information theory explains why the compression is what makes the evidence usable within human and machine bandwidth, and connection to execution lets conviction become action.

This is not an attempt to replace human judgment with an algorithm. It is an attempt to make the history behind human judgment visible. Markets have always priced assets. Proof of Alpha begins pricing credibility.

References

1. H. A. Simon, "Designing organizations for an information-rich world," in Computers, Communications, and the Public Interest, M. Greenberger, Ed. Baltimore, MD, USA: Johns Hopkins Press, 1971, pp. 37–72.

2. E. F. Fama, "Efficient capital markets: A review of theory and empirical work," J. Finance, vol. 25, no. 2, pp. 383–417, 1970.

3. S. J. Grossman and J. E. Stiglitz, "On the impossibility of informationally efficient markets," Amer. Econ. Rev., vol. 70, no. 3, pp. 393–408, 1980.

4. D. Kahneman and A. Tversky, "Prospect theory: An analysis of decision under risk," Econometrica, vol. 47, no. 2, pp. 263–291, 1979.

5. B. M. Barber and T. Odean, "All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors," Rev. Financial Stud., vol. 21, no. 2, pp. 785–818, 2008.

6. R. J. Shiller, "Narrative economics," Amer. Econ. Rev., vol. 107, no. 4, pp. 967–1004, 2017.

7. J. Surowiecki, The Wisdom of Crowds. New York, NY, USA: Doubleday, 2004.

8. J. Wolfers and E. Zitzewitz, "Prediction markets," J. Econ. Perspect., vol. 18, no. 2, pp. 107–126, 2004.

9. P. E. Tetlock and D. Gardner, Superforecasting: The Art and Science of Prediction. New York, NY, USA: Crown, 2015.

10. D. Prelec, "A Bayesian truth serum for subjective data," Science, vol. 306, no. 5695, pp. 462–466, 2004.

11. T. Loughran and B. McDonald, "When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks," J. Finance, vol. 66, no. 1, pp. 35–65, 2011.

12. J. Bollen, H. Mao, and X. Zeng, "Twitter mood predicts the stock market," J. Comput. Sci., vol. 2, no. 1, pp. 1–8, 2011.

13. A. Vaswani et al., "Attention is all you need," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), Long Beach, CA, USA, 2017, pp. 5998–6008.

14. T. G. Dietterich, "Ensemble methods in machine learning," in Proc. 1st Int. Workshop Multiple Classifier Syst. (MCS), LNCS 1857, Cagliari, Italy, 2000, pp. 1–15.

15. R. T. Clemen, "Combining forecasts: A review and annotated bibliography," Int. J. Forecasting, vol. 5, no. 4, pp. 559–583, 1989.

16. M. de Condorcet, Essai sur l'Application de l'Analyse à la Probabilité des Décisions Rendues à la Pluralité des Voix. Paris, France: Imprimerie Royale, 1785.

17. X. Wang et al., "Self-consistency improves chain of thought reasoning in language models," in Proc. Int. Conf. Learn. Represent. (ICLR), Kigali, Rwanda, 2023.

18. J. Wang, J. Wang, B. Athiwaratkun, C. Zhang, and J. Zou, "Mixture-of-Agents enhances large language model capabilities," arXiv:2406.04692, 2024.

19. N. Littlestone and M. K. Warmuth, "The weighted majority algorithm," Inf. Comput., vol. 108, no. 2, pp. 212–261, 1994.

20. L. Lamport, R. Shostak, and M. Pease, "The Byzantine generals problem," ACM Trans. Program. Lang. Syst., vol. 4, no. 3, pp. 382–401, 1982.

21. C. E. Shannon, "A mathematical theory of communication," Bell Syst. Tech. J., vol. 27, no. 3, pp. 379–423, 1948.

22. T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. Hoboken, NJ, USA: Wiley, 2006.

23. J. Rissanen, "Modeling by shortest data description," Automatica, vol. 14, no. 5, pp. 465–471, 1978.

24. G. A. Miller, "The magical number seven, plus or minus two: Some limits on our capacity for processing information," Psychol. Rev., vol. 63, no. 2, pp. 81–97, 1956.

25. J. Zheng and M. Meister, "The unbearable slowness of being: Why do we live at 10 bits/s?," Neuron, vol. 113, no. 2, pp. 192–204, 2025.

26. H. Jiang, Q. Wu, C.-Y. Lin, Y. Yang, and L. Qiu, "LLMLingua: Compressing prompts for accelerated inference of large language models," in Proc. Conf. Empirical Methods Natural Lang. Process. (EMNLP), Singapore, 2023, pp. 13358–13376.

27. Q. Wu et al., "AutoGen: Enabling next-gen LLM applications via multi-agent conversation," arXiv:2308.08155, 2023.

28. G. Li, H. Hammoud, H. Itani, D. Khizbullin, and B. Ghanem, "CAMEL: Communicative agents for 'mind' exploration of large language model society," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), New Orleans, LA, USA, 2023.

29. M. Strathern, "'Improving ratings': Audit in the British university system," Eur. Rev., vol. 5, no. 3, pp. 305–321, 1997.