Research report · September 2026

Evaluating autonomous
financial research

A benchmark for creating strategies, improving existing ones, and using the work of other research agents.

ABSTRACT

Financial research requires an agent to decide which ideas to test and what to do with inconclusive evidence. A strategy can improve repeatedly on the data used during research and still fail on a later period. Evaluating the final backtest alone does not reveal whether an agent discovered a useful signal, exploited the evaluator, or selected a favorable outcome from many attempts.

FAROS evaluates executable strategies on data withheld from the research process. It separates four tasks: originating a strategy, refining a supplied strategy, collaborating with other instances of the same system, and transferring lessons to a new research problem. Models run with their native coding harnesses. Research frameworks are compared under the same independent evaluator.

The primary output is an overall research score, accompanied by four capability scores under explicit resource budgets. The overall score weights the normalized capabilities equally; the breakdown shows strengths that a single ranking can hide.

01 / TASK

The research environment

An episode gives the agent a market dataset, a strategy interface and access to a development evaluator. The agent can inspect data, write code, run local analysis and submit candidate strategies. It decides which hypotheses to pursue and when to revise or abandon them.

The initial scope is daily decisions in liquid digital-asset markets. Strategies use structured, time-stamped data and execute without further LLM calls. Document analysis and discretionary execution are outside this version.

Example task specification

Construct a daily strategy for a fixed asset universe. Use only information available before each decision. Maximize net excess return under the supplied exposure, liquidity and turnover constraints.

Research environment responsibilities
ComponentAvailable to the researcher
DataDated price, volume and volatility observations; chronological development splits.
ToolsShell, code execution, local analysis and a budgeted development evaluator.
SubmissionDeterministic strategy code, dependencies, fixed parameters and a version history.
Portfolio constructionA fixed evaluator applies signal calibration, sizing, execution delays, trading costs and risk limits.
Hidden evaluationNeither the test data nor its results are accessible during the research episode.

Capability tracks

The task inputs and controls differ by track. Each comparison uses matched task instances and the same total allowance.

Capability tracks and controls
TrackStarting pointMeasured outcome
OriginationA strategy interface without a supplied trading strategy.Hidden net excess return of the submitted strategy.
RefinementThe same frozen v6_quality strategy for every model.Change in hidden score relative to the unchanged seed.
CollaborationFour continuing research islands per team, with sharing enabled or disabled.Difference between sharing and isolated teams.
TransferThe same target task and seed, with no notes, useful source notes, or unrelated notes.Difference between memory and no-memory conditions.

Refinement does not begin with each model’s own originated strategy. That would give weak originators more room to improve. Origination measures the ability to produce a useful strategy without a seed; it does not certify that the idea is historically novel.

02 / EVALUATION

Experimental protocol

The unit of comparison is a model and its native coding harness, with the model version, reasoning setting and harness version fixed. In the model comparison, all systems use CORAL to manage workspaces, strategy versions and research memory. In the framework comparison, the research framework changes while the model, native harness and task remain fixed.

An island is a continuing research workspace. Collaboration is allowed inside a team. Competing systems and repeated runs cannot exchange new discoveries. The evaluator operates outside these workspaces and controls access to data and scoring.

Reference experimental configuration
Reference configurationSpecification
Task instancesOne existing Binance spot-market task. The fixed dataset is reused across all systems.
RepetitionsThree runs per system and condition. These are registered targets, not a claim that all runs have completed.
Research allowanceOrigination, refinement and transfer: four evaluator calls and 45 elapsed minutes. Collaboration: eight calls across four islands and 60 elapsed minutes (240 worker-minutes). Waiting for grading counts toward time.
CheckpointsRefinement selections at charged calls 1, 2 and 4. Each checkpoint is selected from development feedback before any hidden evaluation.
Team conditionsFour islands in both collaboration conditions; two calls per researcher. Sharing enables native CORAL notes, attempts, skills and migration. Grading runs one evaluation at a time in both conditions.
Market horizon2023 public research; 2024 adaptive development; 2025 held out for one final operator batch. Source-note episodes fit on the first half of 2023 and receive feedback on the second half.

At each checkpoint, the system commits the strategy selected using development feedback. All checkpoints are frozen before hidden evaluation. A later hidden score cannot be used to choose an earlier checkpoint, revise a strategy or change the benchmark.

Task selection must precede scored runs. Overlapping labels and holding periods are purged at split boundaries. Tasks that share a market period are treated as correlated observations, rather than as independent replications.

Registered model roster

Eight systems span two providers, several capability tiers and multiple generations. These choices test whether strengths differ across research tasks; they do not assume a ranking in advance. Each model runs in its native harness under CORAL.

ModelNative harnessReason for inclusion
GPT-6 AstraCodexFrontier model; tests the upper end of the research budget.
GPT-5.6 SolCodexWorkhorse tier; tests whether a general-purpose system stays competitive.
GPT-5.6 LunaCodexLightweight tier; tests the quality and speed trade-off.
GPT-5.5CodexEarlier generation; provides a concrete reference for progress.
Claude Fable 5.1Claude CodeFrontier model; a second provider at the high-capability end.
Claude Opus 5Claude CodeEarlier flagship; separates generation effects from tier effects.
Claude Sonnet 5Claude CodeMiddle tier; tests the quality and resource trade-off.
Claude Haiku 4.5Claude CodeSmall, earlier-generation model; an inexpensive capability baseline.

Registered native harnesses: Codex 0.153.4 with high reasoning; Claude Code 2.1.263 with native default effort; CORAL 0.7.19. Models use existing subscriptions without paid API fallback. Actual primary model and harness evidence is retained for each run. Research is in progress; completed development measurements appear below.

Source episodes receive four calls and 45 minutes to write at most 4,096 characters of notes. Notes are frozen without operator rewriting; unrelated notes match their length. Source cost is reported separately. Nine compatible earlier Astra, Luna and Sonnet refinement episodes occupy their original slots and are not rerun. Failed calls remain charged; infrastructure failures remain missing observations.

Scoring

S = Rstrategy, net − Rreference, netNet P&L is divided by fixed initial capital of $10,000. The market reference uses the strategy’s deployed exposure. Reported values are percentage points.

The reference is fixed for each task before evaluation and matches the permitted universe and exposure convention. Costs, execution lags and risk limits are evaluator-owned. Refinement, collaboration and transfer scores are paired differences in S against their respective controls. Raw percentage-point scores are normalized against fixed research references before they are combined. The overall score gives each capability a weight of 25%.

Overall research score

All four capabilities use the same economic unit. Nine frozen reference strategies were evaluated on public 2023 data. Their interquartile range of excess returns is 18.1474305489 percentage points, above the registered one-point floor. This scale is fixed before held-out evaluation and does not change when a model enters the leaderboard.

N = 100 × raw capability result / 18.1474305489Overall = (Norigination + Nrefinement + Ncollaboration + Ntransfer) / 4

Zero means zero raw excess return or zero paired gain, depending on the capability. A normalized score of 100 represents 18.1474305489 percentage points; it is not a percentile, a success probability or a ceiling. Negative scores and values above 100 are retained. The same economic gain receives the same weight across all four capabilities.

Equal weighting expresses equal importance for these four abilities, not equal expected investment value. An empirical report should include uncertainty for the complete composite and sensitivity to alternative weights. Final strategy performance and risk remain separate outcomes.

Aggregation, uncertainty and failures

This release uses one market task and three research repetitions per system and condition. Each capability mean requires all three measurements; the overall score requires all four capabilities. Sample standard deviation describes variation across research repetitions on the same market period, not uncertainty across markets. Seeded tracks retain the frozen seed if no completed child improves the development score. Origination without a completed candidate remains unavailable; it is not replaced by a zero-return cash strategy. Failed submissions consume their allowance. An invalid final replay remains missing, and incomplete results remain unranked. Interrupted runs retain their recorded calls and follow the registered recovery rules.

Reading the 95% intervals. Whiskers and shaded bands show pointwise Student t intervals across completed research repetitions on this same task. Hollow points mean an interval is not yet estimable; they do not mean zero uncertainty. With only three repetitions, normality and independent-run assumptions are strong and unverified. Zero observed variation can give a zero-width interval without establishing certainty. These exploratory intervals are not adjusted for repeated checking or all model comparisons, and do not measure uncertainty about future trading returns. Overall needs at least two observations on every axis for an interval; it uses conservative Bonferroni bounds to allow unequal counts and dependence between axes. Confidence interval method.

Memory and collaboration controls

The collaboration comparison changes only the sharing policy, keeping island count, concurrency and total resources fixed. The transfer comparison freezes memory before the target task begins and includes an unrelated, similarly sized memory control. Source-task research cost is recorded separately. A transfer score combines the ability to create useful notes and the ability to use them; swapping memory authors and readers would be a separate experiment.

Resource accounting

Researcher-hours sum active wall time across research workers; elapsed time is reported separately. Equal hours and evaluator calls do not mean equal FLOPs or equal inference cost. A task manifest must also pin local CPU, memory and process limits before execution. Token counts and measured provider costs are reported when available, without treating tokens from different providers as interchangeable compute. Subscription usage is not converted into invented dollar costs.

03 / RESULTS

Model comparisons

46 of 192 registered episodes are closed in this snapshot. The table and figures show real adaptive 2024 development results: three reused repetitions for Astra, Luna and Sonnet refinement; repetition 1 for Fable, Sol, GPT-5.5, Opus and Haiku refinement and the other available tracks. Missing observations remain missing. The 2025 held-out batch has not been evaluated, so these are progress measurements, not final rankings. Ongoing overall scores appear once all four capability axes have a measured result.

Overall research score and capability results
Development · 2024

Figure 1. Ongoing overall points and adaptive development capability measurements. Whiskers show 95% CIs across repetitions where estimable; hollow points have insufficient repetitions. Overall uses all four equally weighted axis means. These are not final held-out rankings.
Ongoing overall points and development capability scores with 95% confidence intervals
SystemOverall ongoingOriginationRefinementCollaborationTransfer
GPT-6 Astra / Codex21.44
4/4 axes
95% CI unavailable
15.32
n=1
95% CI unavailable
2.30
n=3
95% CI [-2.86, 7.47]
0.00
n=1
95% CI unavailable
-2.06
n=1
95% CI unavailable
GPT-5.6 Sol / Codex
2/4 axes

No result
1.30
n=1
95% CI unavailable

No result
0.23
n=1
95% CI unavailable
GPT-5.6 Luna / Codex47.87
4/4 axes
95% CI unavailable
33.58
n=1
95% CI unavailable
0.82
n=3
95% CI [-2.71, 4.35]
-0.33
n=1
95% CI unavailable
0.67
n=1
95% CI unavailable
GPT-5.5 / Codex
2/4 axes
11.01
n=1
95% CI unavailable
3.12
n=1
95% CI unavailable

No result

No result
Claude Fable 5.1 / Claude Code
2/4 axes
13.54
n=1
95% CI unavailable
0.00
n=1
95% CI unavailable

No result

No result
Claude Opus 5 / Claude Code
1/4 axes

No result
0.00
n=1
95% CI unavailable

No result

No result
Claude Sonnet 5 / Claude Code-25.78
4/4 axes
95% CI unavailable
-16.87
n=1
95% CI unavailable
0.17
n=3
95% CI [-0.55, 0.88]
-2.04
n=1
95% CI unavailable
0.03
n=1
95% CI unavailable
Claude Haiku 4.5 / Claude Code
1/4 axes

No result
0.00
n=1
95% CI unavailable

No result

No result

Table 1. Ongoing overall uses normalized points; capabilities use percentage points. Green and bold mark the largest observed value, including ties. n is the number of scored repetitions or matched pairs; unequal counts remain visible. A 95% CI needs at least two repetitions. Missing axes are not zero. Sol and Haiku lack origination scores after infrastructure failures; Opus lacks one after calibration failures.

How the overall score is calculated

The normalization scale is measured from nine frozen public reference strategies. Final model scores remain pending. Every raw capability result is divided by the same scale and multiplied by 100; the four resulting scores receive equal weight.

Measured public-reference normalization scale
CapabilityScale (pp)Weight
Origination18.1474325%
Refinement18.1474325%
Collaboration18.1474325%
Transfer18.1474325%

Ongoing overall is 100 × (origination + refinement + collaboration + transfer) / (4 × 18.147430548871807), using each axis’s current mean. All four axes must have a result. Unequal repetition counts are retained; an unavailable axis leaves overall unavailable. This is a development estimate, not the final held-out score.

Inspect the frozen scale and nine reference results. The final report will include variation across three research repetitions and the registered weight-sensitivity comparisons. One market task does not support claims of generalization across market regimes.

Origination and refinement

Luna has the largest observed origination result in repetition 1: 33.58 pp. Sonnet’s selected strategy is negative at −16.87 pp. These are search-period outcomes, with one repetition per scored model; the remaining repetitions and hidden evaluation can change the comparison.

Capability relationships
Development · 2024
Figure 2. Higher values are better on both axes. Changing the axes reveals different capability relationships. The same system can improve more from a control while still having a lower final score. Horizontal and vertical whiskers show marginal 95% CIs, not a joint 95% region. Hollow points have an unestimable interval on at least one axis. These points do not establish statistical separation.

Research budget

The refinement curves combine nine reused episodes with the new completed runs. Each point is the mean development score of the strategy selected after 1, 2 or 4 charged calls, including the unchanged seed when no candidate improved it. These selection curves can rise by construction; they are not evidence of improvement on unseen data.

Refinement across research budgets
Development · 2024
Figure 3. Recorded 2024 development score at 1, 2 and 4 charged evaluator calls. Three-repetition means are shown for Astra, Luna and Sonnet; Fable, Sol, GPT-5.5, Opus and Haiku each have one completed repetition. The unchanged seed scores 32.13 pp. Final minus seed reconciles with refinement in Table 1. Whiskers and shaded bands show pointwise 95% CIs; hollow points have fewer than two repetitions. Calls are not matched inference compute.

Collaboration and transfer

Collaboration measures the benefit of sharing, rather than the performance of the sharing team alone. Transfer similarly measures the contribution of prior lessons. A high final score and a large gain from collaboration or memory are different outcomes. The paired comparisons show both the starting level and the contribution of the intervention.

Collaboration: sharing versus isolation
Development · 2024
04 / FRAMEWORKS

Comparing research frameworks

A framework comparison holds the model and native harness fixed while changing research orchestration, prompts and memory policies. Each framework receives the same tasks and total allowance. The evaluator remains independent of all frameworks, including CORAL.

The synthetic refinement results below show an interaction: no framework is best for every model. A framework’s mean across models would hide this difference.

Synthetic refinement gains across research frameworks
SystemFramework AFramework BCORAL
GPT-6 Astra / Codex1.52.13.0
GPT-5.6 Sol / Codex1.72.32.1
GPT-5.6 Luna / Codex0.91.61.2
GPT-5.5 / Codex0.60.30.4
Claude Fable 5.1 / Claude Code2.12.82.6
Claude Opus 5 / Claude Code2.01.51.8
Claude Sonnet 5 / Claude Code1.10.80.3
Claude Haiku 4.5 / Claude Code0.20.00.1

Table 2. Synthetic data. Refinement gain over the unchanged seed, in percentage points. Shading marks the largest value in each row. Frameworks A and B are placeholders. This comparison covers refinement only; it cannot establish a framework’s performance on the other tracks.

05 / DISCUSSION

Interpretation and limitations

Economic outcomes and research behavior

GPT-5.5’s no-notes transfer episode tested four candidates, scoring $3,788, $990, $3,195 and $2,232 in development net active P&L. It retained the first: a quiet-asset reversal without the seed’s stress filter. Softening the low-activity preference, rewarding volatility, and then strengthening the low-activity preference all scored worse. The first change also altered the deterministic tie-breaker, so its gain is not a clean filter ablation. The selected candidate still failed the paired lower-bound and quarter-concentration checks.

GPT-5.5’s source episode ended with four calibration failures and no source notes. It contributes to the completion count, but supplies neither an economic source result nor evidence of knowledge transfer. The frozen transfer payload is empty; failed evaluations are not converted into zero returns.

Sol’s unrelated-note condition retained the original seed at +32.13 points after four candidates scored −34.47, +24.04, +19.70 and +17.67 points. Its difference from the no-notes control is −3.33 points. Both note conditions received empty files: together, these outcomes illustrate run-to-run variation, not the value of useful versus irrelevant knowledge.

Sol’s first source-note condition also selected the stateless reversal formula: +35.69 points, a +0.23-point difference from its no-notes control. The source run wrote no notes, and the selected code differs in its tie-breaker rather than its economic formula. Its four attempts scored +24.39, +35.69, +30.95 and +17.84 points. This is one observed pair, with no estimable research-run confidence interval and no demonstrated knowledge-transfer benefit.

In Sol’s first no-notes transfer run, a liquidity-dependent momentum/reversal hypothesis scored −13.69 percentage points relative to the benchmark. Removing the seed’s lagged stress filter produced the selected candidate at +35.46 points, versus +32.13 for the seed. A half-strength filter scored +30.97; emphasizing extreme losers scored +29.63 and passed the quarter-concentration check, but remained below the selected candidate. All four are adaptive development results. The selected formula also appeared in Sol’s separate refinement run with a different deterministic tie-breaker and a +33.43-point score. This limits attribution of the gain to a new economic insight; neither candidate establishes held-out alpha.

Scores measure the value of executable strategies within the task’s economic assumptions. Research traces show how an agent changes a strategy and responds to another researcher’s evidence.

Observed development case: Astra, repetition 1. Four CORAL islands each received two evaluations. With sharing enabled, two researchers explicitly built on a completed peer strategy. One adapted another island’s local-stress filter; another combined a shared market-stress measure with its own recovery signal. Both follow-ups underperformed their peer reference, by 2.20 and 3.74 percentage points of development excess return.

The best strategy replaced an on/off market-stress switch with the fraction of assets that had fallen at the previous decision. The isolated team independently reached the same best development result. Sharing therefore changed the research path without improving the winning result in this pair. Both selected strategies failed the existing statistical lower-bound and quarterly-concentration checks. This is evidence of knowledge exchange, not evidence of reliable alpha or a held-out collaboration advantage. Inspect the strategy changes, lineage and evaluator references.

Observed development case: Luna, repetition 1. In isolation, all four researchers improved on their own first submission, but only one beat the supplied seed. The best result used a smooth volatility penalty instead of a sharper filter. Its 35.25 pp development result narrowly exceeded the sharing team’s 34.92 pp. Both winners still failed statistical lower-bound and quarterly-concentration checks. Inspect the revisions.

Observed development case: Sonnet, repetition 1. Three of four isolated researchers improved their first submission. The winner changed a volatility penalty into a reduction of the whole signal after severe prior market shocks, raising its development score from 32.74 to 36.09 pp. One island’s flow-filter reversal made its result worse. The winner still failed statistical lower-bound and quarterly-concentration checks. One charged submission used tune mode; an operator invoked native stop after the eight-call budget was exhausted, and two scoreless rejected attempts remain in the audit. Inspect the revisions and closure evidence.

Observed Sonnet sharing revision. In the sharing run, one researcher noticed that replacing a neutral multiplier with a volatility rank also reduced average signal strength. It doubled that rank multiplier to address the mismatch. Its next result rose from 21.84 to 24.84 pp, still below the unchanged seed at 32.13 pp. The code also changed its tie-break salt, so the gain cannot be attributed solely to the multiplier. This is a documented revision of the researcher’s own strategy, not evidence of peer benefit. Inspect the code change and evaluator evidence.

The sharing team selected a different idea: scale the volatility penalty by the fraction of assets that had fallen, instead of an on/off market switch. Its 34.05 pp result trailed isolation’s 36.09 pp by 2.04 pp. Both winners failed statistical lower-bound and quarterly-concentration checks. Sharing used eight charged attempts but only seven received scores: a third submission by one researcher failed before native admission, leaving a 3/2/2/1 allocation. No replacement slot was supplied. The comparison includes this allocation failure and the isolated run’s tune/stop caveat; it does not isolate the effect of sharing at a balanced allocation. Inspect both selections and the allocation audit.

Observed development case: Fable 5.1, refinement repetition 1. Its first change scored 3.27 pp versus the seed’s 32.13 pp. The researcher then revised its public-data fitting method to account for portfolio selection and turnover. Adding a preference for less-traded names scored 21.88 pp; reversing that preference scored −52.58 pp. A final change scored 21.54 pp. All four revisions stayed below the seed, which was retained at every checkpoint: refinement gain 0.00 pp. The trace shows adaptation to feedback, but no improvement in the selected strategy in this repetition. Inspect all submissions, frozen strategies and checkpoints.

Observed development case: Sol, refinement repetition 1. Changing the volatility preference scored 30.18 pp, below the seed’s 32.13 pp. Removing the market-stress volatility filter entirely then scored 33.43 pp, a gain of 1.30 pp. Stronger concentration in less-traded names scored 26.98 pp; combining reversal with momentum scored −13.71 pp. Sol retained the simpler second submission at the final checkpoint. That selected strategy still failed statistical lower-bound and quarterly-concentration checks. Inspect the four submissions and frozen selections.

Observed development case: GPT-5.5, refinement repetition 1. Replacing the on/off market-stress rule with a gradual penalty scored 35.25 pp, up 3.12 pp from the seed. Favoring less-traded names more strongly reduced costs from $1,744 to $1,327 and passed the quarterly-concentration check, but lowered the score to 30.09 pp. Delaying the stress penalty scored 33.33 pp; emphasizing larger prior price falls scored 32.18 pp. The first revision remained selected at every checkpoint. It still failed statistical lower-bound and quarterly-concentration checks. Inspect the four revisions, risk measures and frozen checkpoints.

Observed development case: Opus 5, refinement repetition 1. Favoring fresh price drops scored 19.50 pp; reversing that preference to favor continuing declines scored 20.87 pp. Both stayed below the seed’s 32.13 pp. Opus then abandoned that filter. Avoiding the least-traded names scored 15.01 pp; increasing the preference for less-traded names scored 26.94 pp. The researcher changed its hypothesis in response to feedback, but none of its four revisions improved the selected strategy. The unchanged seed was retained at every checkpoint: refinement gain 0.00 pp. Inspect the revisions, evaluator results and frozen checkpoints.

Observed development case: Haiku 4.5, refinement repetition 1. Replacing lagged market stress with a continuous volatility adjustment scored 20.46 pp, below the seed’s 32.13 pp. A purported threshold change produced the same score: the formula rescales the volatility multiplier instead of changing an activation boundary. The third submission restored the seed’s scoring logic while describing it as a new hybrid, returning to 32.13 pp. Halving the volatility penalty then scored 29.29 pp. The seed remained selected at every checkpoint, giving 0.00 pp refinement gain. These traces distinguish a change in the research narrative from a change in executable behavior. Inspect the threshold formula, the restored seed logic and all frozen results.

Observed source research: Sol, repetition 1. The first submission failed calibration. A revised implementation received a score of −$961.80 in net active P&L on the public July–December 2023 source window; two later ideas scored −$1,982.48 and −$2,138.99. These source scores are separate from the 2024 capability table. No transfer-notes file was produced, so the frozen source-note and unrelated-note conditions both contain empty files. Their later outcomes cannot demonstrate transfer of written knowledge. Inspect the closed source run.

Historical exposure

Withholding test data prevents feedback from that evaluator. It does not exclude market knowledge acquired during pretraining. Dated data audits, controlled tool access and prospective evaluations are needed for stronger temporal claims. Masking names and dates is a partial mitigation, not proof of an uncontaminated task.

Task coverage and uncertainty

Repeated runs measure variability in the research system. Different market periods measure a different source of uncertainty. Many assets or research briefs from the same period cannot substitute for independent market conditions. The reference task count must be revisited after an outcome-free data audit; narrow confidence intervals from only a few periods would be misleading.

Risk and deployment

Net excess return is reported with portfolio return, benchmark return, maximum drawdown, turnover, exposure and concentration. These risk results are not fabricated in this preview and would be required in an empirical release. Historical paper performance does not establish that a strategy can be traded at the assumed costs or capacity.

Scope of the comparison

Native harnesses are part of the evaluated systems. A difference between Sonnet in Claude Code and Astra in Codex cannot be attributed solely to model weights. Collaboration results depend on the sharing policy; transfer results depend on the source tasks and memory construction. These are conditional capability measurements, not universal rankings.

APPENDIX

Data and reproducibility

An empirical report must identify exact model, harness and framework versions; task and data hashes; resource usage; every submitted strategy; and evaluator receipts. The charts and downloadable data use recorded development results with source hashes. Framework examples remain synthetic and are excluded from the development download. The observed development case links to its frozen strategies, individual evaluator references and cross-island lineage. Final held-out scores and measured benchmark uncertainty are not yet available. Inspect this development snapshot and its source hashes.