README.md

---
pretty_name: Dissei Financial Judgment — Sample
language:
- en
license: other
license_name: commercial-license-by-agreement
license_link: https://hub.harborframework.com/datasets/dissei/financial-judgment#access-policy
size_categories:
- n<1K
task_categories:
- question-answering
tags:
- rl-environment
- finance
- financial-reasoning
- financial-analysis
- question-previews
- evaluation-results
configs:
- config_name: previews
  default: true
  data_files:
  - split: test
    path: task-previews.jsonl
---

# Dissei Financial Judgment — Sample

[Website](https://dissei.ai) · [Harbor](https://hub.harborframework.com/datasets/dissei/financial-judgment) · [Hugging Face](https://huggingface.co/datasets/Dissei-Data/Dissei-Financial-Judgment) · [GitHub](https://github.com/Dissei-org/financial-judgment)

Dissei Financial Judgment examines how AI models turn financial evidence into well-supported conclusions.

Read [Financial judgment for LLMs](https://dissei.ai/guides/financial-judgment-for-llms) for more on how we think about financial judgment and its evaluation.

This sample contains seven tasks drawn from a single real-world deal, **Rung 3 anonymized while preserving the core financial reasoning**. The full evaluated sample includes a carefully designed environment, supporting exhibits, task-specific rubrics and reference answers. Across three dated evidence snapshots, models interpret evidence, assess explanations and alternatives, and reach a judgment at the stated decision date.

This release is an **illustrative sample of the types of financial-judgment tasks Dissei builds**, rather than the full task catalogue or evaluation environment. It presents the questions, briefs, methodology and recorded results for six models; supporting exhibits and execution tools are not included in these public previews.

Complete evaluation packages can be tailored to your requirements, including task-specific rubrics, an evaluation runner and supporting exhibits. To discuss a package and delivery, contact [tech@dissei.credit](mailto:tech@dissei.credit). Scope, access and permitted use are agreed before delivery.

[View the public Harbor sample](https://hub.harborframework.com/datasets/dissei/financial-judgment). The [Hugging Face viewer](https://huggingface.co/datasets/Dissei-Data/Dissei-Financial-Judgment/viewer/previews/test) shows the question previews; [`results.json`](https://huggingface.co/datasets/Dissei-Data/Dissei-Financial-Judgment/blob/main/results.json) contains the recorded scores and settings.

## Research question and scope

**Can a model identify what matters in a financial situation, connect the evidence to a conclusion, and distinguish what is supported from what remains uncertain?**

The seven task families cover diagnosis, prediction, explanation, quantitative interpretation, counterfactual reasoning, comparison and strategy. Together they exercise different analytical demands within the same transaction and at different decision dates.

The guide sets out our broader perspective; this pilot focuses on submitted analysis and recorded evidence use. It is a bounded assessment of financial judgment, rather than a comprehensive multi-step workflow evaluation.

## Tasks

Open a task to read its original question and brief. These previews illustrate the task types; complete evaluation packages are supplied separately under agreed terms.

In the viewer, `question` contains the question and `brief` contains the remaining context without repeating it. Linked Harbor instructions preserve the complete wording.

| Family | Information cutoff | Question |
|---|---|---|
| Diagnostic | 2021-12 | [What diagnosis should you make of Company CI recent margin pressure at the anchor?](https://hub.harborframework.com/tasks/dissei/finance-preview-dg04) |
| Predictive | 2021-12 | [Would you expect Company CI to be relatively resilient if demand conditions weakened after the anchor?](https://hub.harborframework.com/tasks/dissei/finance-preview-pr02) |
| Explanatory | 2022-06 | [Why did Company CI recent margin pressure matter for the investment case at the anchor?](https://hub.harborframework.com/tasks/dissei/finance-preview-ex01) |
| Quantitative | 2022-06 | [What does the segment-share exhibit imply about Company CI relative market position?](https://hub.harborframework.com/tasks/dissei/finance-preview-qn02) |
| Counterfactual | 2022-09 | [Had Company BR declined to roll equity and provide a seller note, how would Company ED have had to reshape the capital structure to close at the same price, and how would risk-sharing shift?](https://hub.harborframework.com/tasks/dissei/finance-preview-cf01) |
| Comparative | 2022-09 | [Company ED is negotiating to acquire Company CI at a $14 billion enterprise value. Measured against how public equity markets value comparable HVAC manufacturers, is that entry price attractive?](https://hub.harborframework.com/tasks/dissei/finance-preview-cp02) |
| Strategic | 2022-09 | [Which diligence workstream should Company ED treat as gating before signing the carve-out?](https://hub.harborframework.com/tasks/dissei/finance-preview-st06) |

## Recorded results

Reward / 100 is the equal-weight mean of the seven saved continuous task rewards, multiplied by 100. It is not accuracy. Values use unrounded rewards before display. Scores come from the complete evaluation tasks, not the read-only public previews. No task data, rubric or recorded reward changed for this publication.

| Model | Reward / 100 | Submitted | Run notes |
|---|---|---|---|
| Claude Fable 5.1 | 60.73 | 7/7 | 7 submitted; no infrastructure recovery |
| GPT-6 Astra | 52.45 | 7/7 | 7 submitted; no infrastructure recovery |
| Claude Opus 5.5 | 52.17 | 7/7 | 7 submitted; no infrastructure recovery |
| GLM-5.3 | 44.91 | 7/7 | 7 submitted; 1 task recovered; earlier recovery interrupted by the VM limit |
| Kimi K3 | 35.79 | 7/7 | 7 submitted; no infrastructure recovery |
| Gemini 3.8 Flash | 9.01 | 1/7 | Protocol-limited: 1 submitted, 6 turn limits; 5 tasks recovered after infrastructure failures |

Gemini 3.8 Flash's result is protocol-limited: six selected attempts reached the turn limit, and the JSON command parser rejected 50 responses. A non-submission receives zero without an LLM grade. The number therefore measures this harness interaction as well as the model. GLM-5.3's recovery completed after a previous recovery was interrupted by the VM's runtime limit; the other six successful original trials were retained. No successful trial was rerun to seek a higher score.

### Why the score is continuous

A financial answer can be partly sound: it may reach a defensible conclusion while missing a material qualification or supporting evidence. Continuous reward captures these distinctions instead of reducing every response to right or wrong.

Required-check pass, submission and answer quality are separate measures. Passing required checks does not earn full credit. The displayed reward is a quality score, not accuracy or a probability of correctness; additional decimal places do not establish measurement precision.

## Diversity and reward distribution

The broader [Dissei evaluation record](https://dissei.ai/bench) spans multiple cases, decision dates and the seven reasoning categories. Its pooled model results have unequal case coverage and should be read descriptively. The six-model comparison above uses the same seven task versions from one transaction.

A separate [historical two-family study](https://dissei.ai/research/where-equal-models-come-apart) evaluated 47 tasks with three rollouts per task per model—282 rollouts in total. The July 2026 methods brief reports per-task mean rewards spanning near zero to about 0.9, with task-level standard deviations of about 0.23 for both families on a 0–1 scale. This shows variation across task-average rewards, beyond the headline mean.

These historical distribution figures belong to that separate study. They describe observed score spread, not uncertainty intervals, judge reliability or demonstrated training gains.

## An inspectable assessment

An [already-public worked assessment](https://dissei.ai/bench#assessment-first-lien-dispersion) illustrates the distinction: its recorded gate result is **Passed**, while its score is **46.4/100**. The answer reads the observed price dispersion but omits yield and recovery-scenario analysis and overstates the arbitrage conclusion. Equal ranking alone does not establish identical cash flows, exposure, recovery or a free arbitrage.

This historical assessment is separate from the seven-task pilot. The source presents selected answer excerpts and commentary, not the full private evidence or executable grader; it does not validate the pilot's scores.

## Evaluation method

- Seven fixed, content-addressed task versions, with December 2021, June 2022 and September 2022 information cutoffs.
- Harbor 0.23.0, Terminus-2 JSON agent harness, 20 dataroom calls, 40 agent turns and a 30-minute task limit.
- High agent reasoning, with at most 16,384 output tokens per request.
- A separate `gemini-3.1-pro-preview` judge, low reasoning and one vote, applies the authored rubric.
- One selected outcome per task. Infrastructure-only recovery attempts are retained in the private audit record with the original failures. An unresolved trial exception excludes an aggregate score; a genuine non-submission remains a zero.
- Model identities, task digests and original results are retained for controlled verification. Provider access paths and client framing can differ; this is not a claim of identical provider implementations.

## Evaluation integrity

- **Separate holdouts.** Dissei maintains separate holdout material for evaluation, distinct from this public sample.
- **Point-in-time analysis.** Each task fixes a decision date and asks for a judgment supported by the evidence available then. Later outcomes can inform review without becoming the sole answer key.
- **Versioned comparisons.** Recorded results refer to fixed task versions and documented settings. Changes to evidence or grading require separately identified evaluations.
- **Diagnostic review.** Saved outcomes and execution records support investigation of evidence use, analytical errors and harness failures, rather than relying on an aggregate score alone.

**Temporal judgment.** Some context deliberately includes information outside a task's decision date. Models must distinguish relevant, contemporaneous evidence from distractors rather than treating every visible date as usable. The tasks are designed to permit full credit without relying on those distractors.

Rung 3 replaces company and person names while retaining dates and financial figures for analysis. Those details can permit re-identification. Separate holdouts and anonymization do not establish absence of prior model exposure to underlying sources; this public sample is not claimed to be an undisclosed, contamination-free holdout.

## Interpretation and review

This seven-task pilot provides a focused comparison within one transaction. It has one selected outcome per model and task, with no repeated-run variance or judge-agreement estimate. Judge, harness and provider differences remain relevant; small score differences should not be treated as established model superiority.

## Access policy

Public questions, briefs, metadata, results and linked assessment excerpts are available for inspection. This public sample does not include the full case evidence, runner or grading materials.

Complete evaluation packages—including task-specific rubrics, an evaluation runner and supporting exhibits—are supplied separately under an agreed scope and license. Package contents, model/provider use, retention, redistribution and training rights are agreed before delivery. Task-authoring machinery and private production workflows are not part of the standard package.

For Dissei-operated studies, model selection, evidence disclosure, spending limits, run count and report contents are agreed before execution. Reports can include version identifiers, settings, selected outcome summaries and disclosed failures; answer or evidence excerpts require disclosure review. The recorded Dissei-operated results on this page are not independently reproduced results, and Harbor does not certify the scores.

An inquiry does not itself grant package access or subscribe anyone to a mailing list.

## Contributions

Maintainers, researchers and practitioners are welcome to inspect the sample, challenge the methodology and propose evaluation runs under agreed terms. Discuss research questions in the [Hugging Face Community tab](https://huggingface.co/datasets/Dissei-Data/Dissei-Financial-Judgment/discussions), or open a [GitHub issue](https://github.com/Dissei-org/financial-judgment/issues) for a correction or proposal. The GitHub repository contains the research README, not the private evaluation environment.

Research directions we welcome:

- **Temporal relevance:** how models select evidence at the decision date, including controlled distractor ablations.
- **Failure analysis:** distinguishing evidence-selection errors, financial-judgment errors and harness failures.
- **Grading reliability:** judge agreement and how reward changes with answer quality.
- **Repeatability and efficiency:** variation across attempts and quality versus token, tool and inference cost.

These are proposed analyses, not additional results from this pilot. Focused documentation pull requests are welcome. Run reports should identify task and model versions, harness and judge settings, who operated the run, and any failures or recovery attempts. Discuss changes to tasks or scoring before proposing them; existing results must remain tied to their evaluated versions.

For evaluation access, research proposals or private security reports, contact [tech@dissei.credit](mailto:tech@dissei.credit). Please keep credentials, protected case material, raw run records and exploit details out of public posts. Contributions do not expand the access or licensing terms above.