Dissei-Data/Dissei-Financial-Judgment

Dissei Financial Judgment — Full Evaluation Package

Harbor · GitHub · Website · Hugging Face

Dissei explores financial reasoning and analysis behind institutional investment and credit decisions: interpreting evidence, weighing trade-offs and reaching well-supported conclusions.

Financial Judgment contains seven tasks drawn from one completed private-equity deal. Each asks for a written analytical judgment at a fixed date, using evidence from December 2021, June 2022 or September 2022. The package includes the task instructions, supporting evidence, 79 rubric criteria, judge prompts, scoring implementation, reference answers, dataroom tools and a separate verifier. Runs use the Harbor task files together with three Docker image archives containing the runtime and evidence.

Version 2.0.0 is available on public GitHub and Harbor. This Hugging Face mirror is public with manual gating: log in and receive approved access to download its files or use Data Studio.

The visualiser browses task specifications and reference/verifier files from the public GitHub v2.0.0 mirror. It cannot forward your Hub login to gated Hugging Face files or start a rollout.

Quickstart · Full run guide · GitHub release · Hugging Face files — login + approved access

Quickstart

Download the complete release

Use Git, the GitHub CLI (gh), Docker with Compose support and uv. Choose a host that supports linux/amd64 or linux/arm64. Start in a parent directory where financial-judgment-v2.0.0 does not already exist. These commands create a new checkout and download the three image archives into its images/ directory.

git clone --branch v2.0.0 --depth 1 \
  https://github.com/Dissei-org/financial-judgment.git financial-judgment-v2.0.0 &&
cd financial-judgment-v2.0.0 &&
gh release download v2.0.0 --repo Dissei-org/financial-judgment \
  --pattern 'financial-judgment-*.tar.gz' --dir images

Continue after the clone and all downloads succeed. A runnable package needs both parts: the repository supplies tasks/, images/SHA256SUMS, the run guide and historical attachments; the release assets supply runtime and evidence. Docker loads the compressed archives directly. Do not unpack them as task folders.

For direct downloads, save each archive below in the cloned repository's images/ directory. The matching SHA256SUMS is included at images/SHA256SUMS; image-inventory.json records image/platform digests. Keep files from different releases separate.

Snapshot	Docker image archive	SHA-256
2021-12	financial-judgment-2021-12.tar.gz	5f945820221ca7b3972b3044b2c5ac337eb6e39b0aff6e56bd1e7dfabb932edf
2022-06	financial-judgment-2022-06.tar.gz	f61750f2e4a1e84a3f5c80293c8758e7304d564808e92d16b7e62ba8b6180675
2022-09	financial-judgment-2022-09.tar.gz	b82a59b92f2bb1285cefff6923d01e9b5b3c7d3d1283173f991e0f4f66f39d6a

The same files are available through Hugging Face with login and approved access. GitHub and Harbor downloads are public. You supply the Docker host, provider account and judge endpoint for your run.

Run the commands below from the package directory containing both tasks/ and images/. If you downloaded the files separately, combine them into that layout first. Load the supplied images directly; no environment rebuild is needed.

1. Install Harbor and load the supplied images

uv tool install harbor==0.23.0
shasum -a 256 -c images/SHA256SUMS

Continue only if all three checksums pass:

docker load --input images/financial-judgment-2021-12.tar.gz
docker load --input images/financial-judgment-2022-06.tar.gz
docker load --input images/financial-judgment-2022-09.tar.gz

2. Configure your judge

Set DISSEI_JUDGE_BASE_URL to your OpenAI-compatible API base URL and DISSEI_JUDGE_MODEL to your chosen judge model. Supply DISSEI_JUDGE_API_KEY through your normal secret-management workflow; it may be empty only for an intentionally unauthenticated endpoint. The endpoint must be reachable from inside Docker.

These runs send evidence, answers and grading material to your chosen provider and may incur charges. Use an authorized endpoint. Harbor's saved configuration and process arguments can contain credentials: keep run directories private and do not enable shell tracing.

3. Run one included reference solution

harbor run --path ./tasks/fab01-2209-st06 \
  -a oracle -n 1 -k 1 \
  --ve DISSEI_JUDGE_BASE_URL="${DISSEI_JUDGE_BASE_URL:?Set your judge base URL}" \
  --ve DISSEI_JUDGE_MODEL="${DISSEI_JUDGE_MODEL:?Set your judge model}" \
  --ve DISSEI_JUDGE_API_KEY="${DISSEI_JUDGE_API_KEY:-}"

This runs and grades the supplied reference answer. The reward depends on your judge and can be below full credit. To run all seven reference solutions, replace the path with ./tasks. Use the next command to evaluate an agent.

4. Evaluate your agent on all seven tasks

Configure the agent provider's credentials separately and set DISSEI_AGENT_MODEL to its authorized provider/model identifier:

harbor run --path ./tasks \
  -a terminus-2 -m "${DISSEI_AGENT_MODEL:?Set your agent provider/model}" \
  -n 1 -k 1 --ak max_turns=40 --ak reasoning_effort=high \
  --ve DISSEI_JUDGE_BASE_URL="${DISSEI_JUDGE_BASE_URL:?Set your judge base URL}" \
  --ve DISSEI_JUDGE_MODEL="${DISSEI_JUDGE_MODEL:?Set your judge model}" \
  --ve DISSEI_JUDGE_API_KEY="${DISSEI_JUDGE_API_KEY:-}"

Here -n 1 limits concurrency to one trial and -k 1 selects one attempt per task. Inspect the job directory printed by Harbor: each trial's verifier/reward.json, verifier/score_breakdown.json, verifier/integrity.json and exception state. Missing required judge configuration and top-level graded/contradiction unavailability produce errors without a reward. Some per-criterion failures still allow a reward; inspect the unavailable flags and the judge-failure limitation.

These commands provide local setup settings. The historical comparison used the settings recorded below. After loading the supplied images, local runs require neither a Harbor Hub login nor a public registry upload. See the full run guide for optional judge settings, network configuration, dataroom commands and troubleshooting.

Package at a glance

Item	Scope
Tasks	Seven, one per reasoning family
Case	One completed private-equity transaction, with names suppressed
Anchors	December 2021, June 2022 and September 2022
Rubric criteria	79 total; 10–13 per task
Environment images	Three dated snapshots, each for linux/amd64 and linux/arm64
Harness	Harbor 0.23.0; local task directories and separate verifier containers
Reward	Continuous from 0 to 1; user-configured LLM judge
Historical records	Sanitized derivative of 16 jobs and 86 attempts
Current historical comparison	42 selected outcomes: six models × seven tasks
License category	Other / restricted; see LICENSE.md

Evaluation files and access

The package includes evaluator source, judge prompts, rubric criteria and reference answers. Runtime source and evidence can be extracted from the Docker images. Recipients can inspect the grading method and may infer aspects of task design from these files.

The runtime runs the dataroom and evaluator. Task-authoring machinery, private production workflows and authoring/build/orchestration modules are excluded. Private infrastructure addresses, local operator identities and revealing historical source excerpts have been removed or visibly redacted from the historical records. Required copyright and third-party notices are retained.

A separate verifier isolates grading inputs from the task agent during an ordinary Harbor run. Recipients who control the host can still read all downloaded files. When measuring performance, keep host-side tests/, solution/ and the historical answer archive out of the task agent's reach.

Tasks

Local task directory	Anchor	Family	Analytical demand	Criteria
fab01-2112-dg04	2021-12	Diagnostic	Diagnose a financial outcome from dated evidence.	12
fab01-2112-pr02	2021-12	Predictive	Judge how a situation may develop as of the anchor.	11
fab01-2206-ex01	2022-06	Explanatory	Explain why a development matters for an investment or financing decision.	13
fab01-2206-qn02	2022-06	Quantitative	Read an exhibit quantitatively and state its implications.	10
fab01-2209-cf01	2022-09	Counterfactual	Analyze a changed capital structure and allocation of risk.	13
fab01-2209-cp02	2022-09	Comparative	Compare an entry price with relevant public-market evidence.	10
fab01-2209-st06	2022-09	Strategic	Recommend and defend a diligence priority.	10

Each task asks one question at one anchor date. All seven draw on the same transaction. The three snapshots contain timeline narratives, exhibits and primary-source documents, with the original financial facts, dates, figures and source notices retained. Follow each task's instructions and the dataroom's point-in-time retrieval rules.

The metadata records identity_tier = "name_suppressed": case names are suppressed while the original financial facts remain. Exact dates, amounts, market data, transaction structure and combinations of details can permit re-identification or linkage to external sources. Name suppression does not guarantee anonymity or irreversible de-identification.

Package layout

Run commands from the directory containing this README:

README.md
USAGE.md
CHANGELOG.md
LICENSE.md
publication-plan.md
cards/
dataset.toml
local-task-index.json
.financial-judgment-owned.json
tasks/
  fab01-2112-dg04/
    instruction.md
    task.toml
    environment/
      Dockerfile
    tests/
      Dockerfile
      verify.py
      test.sh
      task.yaml
    solution/
      solve.sh
  fab01-2112-pr02/
  fab01-2206-ex01/
  fab01-2206-qn02/
  fab01-2209-cf01/
  fab01-2209-cp02/
  fab01-2209-st06/
images/
  financial-judgment-2021-12.tar.gz
  financial-judgment-2022-06.tar.gz
  financial-judgment-2022-09.tar.gz
  SHA256SUMS
image-inventory.json
run-index.json
run-records.tar.gz
evaluation-protocol.json
redaction-summary.json

Each task uses the eight-file layout shown for fab01-2112-dg04. instruction.md is the agent prompt; task.toml selects the supplied image and defines execution and metadata; tests/ contains the task YAML, rubric source and separate-verifier entry points; solution/solve.sh supplies the reference answer and oracle entry point. Together, the seven tasks contain 79 criteria. Harbor 0.23.0 requires an environment/ directory; its one-line Dockerfile references the corresponding supplied image with FROM. Runtime, evidence, wrappers and requirements are supplied in the images.

The three image archives contain the dataroom and evaluator runtime at /opt/dissei/runtime and the dated evidence at /corpus. Task-specific rubrics and reference solutions live in the task packages, outside the task image. Harbor builds the separate verifier from tests/Dockerfile on top of the corresponding loaded image. This small verifier build uses the supplied environment as its base.

dataset.toml describes the public registry dataset and pins its seven task packages by content digest. For the local workflow, run harbor run --path ./tasks from the package directory after loading the images. The Harbor dataset and its task packages are public. Hosted New Job support and remote --dataset execution have not been verified for this release; USAGE.md documents the local workflow.

.financial-judgment-owned.json records file hashes used to recognize generated package files. Private authoring source and identity mappings are excluded.

Data Studio overview

The tasks / test viewer indexes seven questions from one shared case across three dated environment snapshots. Each row provides instructions, exhibit catalogue entries and links to the task and image files. The complete evaluation payload is distributed in those files. Hugging Face login and approved access are required for the viewer.

Field	What it contains
task_id, family, decision_date	Existing task identifier, reasoning family and decision anchor in YYYY-MM precision.
task, question	Full original agent-facing instruction, plus the unchanged short question separately.
world, environment_snapshot	Case slug company-ci-carveout, shared by all seven tasks, with a dated snapshot: 2021-12, 2022-06 or 2022-09.
exhibits	Complete exhibit catalogue available at the task's anchor, with original IDs and titles in catalogue order.
task_url, instruction_url	Direct links to the full task folder and exact instruction file; Hugging Face access rules apply.
evidence_location, image_url	Docker archive download and the in-container evidence path /corpus/company-ci-carveout.

Each December and June task has four visible catalogue entries; each September task has eleven. Exhibit IDs are case-scoped and can refer to different content in different snapshots. Evidence bodies live in the three supplied images. The viewer lists every catalogue entry available at the task's anchor, with no reference answers, rubric fields or grading-based evidence selections. Run the Harbor task files with the supplied images using the Quickstart.

Reward and verifier

The evaluator uses the original task-specific rubrics, judge prompts, scoring math and reference answers:

• Critical criteria gate or scale the score according to which required claims are satisfied.
• Graded assessment returns a holistic 0–10 judgment and per-criterion pass, partial or fail diagnostics; criterion weights guide the assessment.
• Negative-weight pitfall criteria penalize specified reasoning errors.
• A contradiction finding halves the score.
• A retrieval modulator between 0.7 and 1.0 can reduce, but never increase, the answer score according to credited evidence retrieval.

The verifier runs in a fresh container and receives only declared episode artifacts from the agent. It checks submission provenance, replays retrievals against its own corpus and enforces the dataroom call budget. A missing or invalid submission receives zero. Missing required judge configuration produces exit 3 without a reward. A scoring exception that reaches the entry point, or a breakdown with top-level graded_unavailable or contradiction_unavailable, produces exit 4 without a reward. See USAGE.md for error states and output files.

Judge-failure limitation

The evaluator retains an inherited limitation: unavailable critical verdicts are flagged in critical_results but excluded from the gate denominator. If no critical verdicts are available, gate_score defaults to 1.0. Unavailable pitfall verdicts are flagged in pitfall_results and contribute zero penalty. Neither per-row condition triggers the top-level unavailable check, so a critical-only or pitfall-only failure can still produce a reward.

Missing or malformed graded criterion diagnostics can also coexist with a usable holistic score. With multiple judge votes, some unusable votes can be discarded while an aggregate verdict remains usable under the existing voting rules. Inspect these diagnostics alongside the reward and exception state. The release leaves these scoring rules unchanged.

The score combines evidence retrieval, financial reasoning, judged writing and harness control. Weak analysis, failed retrieval or protocol errors can each lower it. The historical agent protocol allowed 20 dataroom calls, 40 turns and 30 minutes per task. A model that never submits receives zero without an LLM grade.

Failure-case checks

Release checks compared the original and cleaned runtime on fab01-2209-cf01 in 20 controlled failure and edge-case runs: ten deliberately injected scenarios on each of linux/amd64 and linux/arm64. Separate reference-answer comparisons covered all seven tasks.

Three components handle a run:

• The dataroom runtime tracks the retrieval budget and records evidence access and final submission.
• The Dissei verifier checks submission integrity, replays retrievals against the corpus, rejects tampering and handles the judge outcomes described above.
• Harbor runs the agent and verifier phases, transfers declared artifacts and records rewards or trial errors.

Injected condition	Verified behavior
No submission	Reward 0; no judge call.
Retrieval budget exhausted, then submission	Submission remains accepted; terminal submission is not counted as an extra retrieval.
Forged evidence-retrieval record	Replay mismatch detected; integrity marked failed.
Answer edited after submission	Invalid submission; reward 0 and no judge call.
Missing judge configuration	Verifier exits 3; no reward file.
Whole judge endpoint unreachable in the injected scenario	Verifier exits 4; no fabricated zero score.
Wholly unusable decisive judge verdicts in the injected scenario	Verifier exits 4; no reward file.
Forged calls beyond the budget	Excess records dropped; integrity marked failed.
Answer file planted without submission	Reward 0; no judge call.
Old reward file present, followed by configuration failure	Stale reward is not reused; verifier exits 3 with no reward file.

All 20 comparisons matched the original behavior. Two additional cutoff checks confirmed that later-dated evidence remained inaccessible on both architectures.

These checks cover the conditions in the table: absent or invalid submissions received zero, while the injected whole-endpoint outage and wholly unusable decisive verdicts produced errors without scores. Per-role or per-criterion unavailability can still yield a reward. The comparisons show that sanitization preserved the tested execution rules. To attribute a historical model failure to Harbor or another component, inspect the particular recorded attempt.

Historical results

Preliminary comparison recorded on 2026-09-24–25 (UTC)

These outcomes come from the earlier task versions identified in the records. The records retain their provenance and selection rules; version 2.0.0 has new package and image digests and has not received a new model evaluation.

The selected GLM-5.3 recovery completed at 2026-09-25T00:46:07Z; the other 41 selected outcomes completed on 2026-09-24 UTC.

The comparison used Harbor 0.23.0, Terminus-2's JSON harness, high agent reasoning, 40 turns, a 16,384-token output limit per request and one selected outcome per task. The common judge was gemini-3.1-pro-preview, with low reasoning and one vote. Provider/model identities and parameters are retained; private endpoint details are redacted. Different provider clients can frame requests differently.

Reward / 100 is the equal-weight mean of seven saved continuous task rewards, multiplied by 100. Means use unrounded rewards before display. The score reflects the documented rubric, retrieval, harness and judge settings. Accuracy and probability of correctness require separate measurements.

Model	Reward / 100	Submitted / 7
Claude Fable 5.1	60.73	7
GPT-6 Astra	52.45	7
Claude Opus 5.5	52.17	7
GLM-5.3	44.91	7
Kimi K3	35.79	7
Gemini 3.8 Flash	9.01	1

Task	Claude Fable 5.1	GPT-6 Astra	Claude Opus 5.5	GLM-5.3	Kimi K3	Gemini 3.8 Flash
fab01-2112-dg04	0.7882	0.7338	0.7094	0.4892	0.0000	0.6306
fab01-2112-pr02	0.7094	0.7338	0.7094	0.5708	0.5518	0.0000
fab01-2206-ex01	0.4570	0.2285	0.2285	0.2285	0.2340	0.0000
fab01-2206-qn02	0.4333	0.4000	0.4333	0.5000	0.3467	0.0000
fab01-2209-cf01	0.6688	0.5320	0.5964	0.6080	0.5250	0.0000
fab01-2209-cp02	0.7455	0.7455	0.7455	0.5218	0.6171	0.0000
fab01-2209-st06	0.4490	0.2982	0.2294	0.2251	0.2307	0.0000
Equal-weight mean	0.6073	0.5245	0.5217	0.4491	0.3579	0.0901

The comparison selects 42 outcomes from the 86 retained attempts. It includes genuine non-submission zeros; unresolved trial exceptions remain errors. Infrastructure recovery attempts are linked in the records. No successful trial was rerun to seek a higher score.

• Gemini 3.8 Flash submitted once; six selected attempts reached the 40-turn cap, and the JSON parser rejected 50 responses. Five infrastructure-failed tasks were retried, while the original valid submission and genuine no-submission zero were retained. Its score is protocol-limited.
• GLM-5.3 retained six original successful outcomes. The remaining task's original timeout and interrupted recovery are preserved alongside the completed recovery, which scored 0.608 on fab01-2209-cf01.
• Kimi K3 submitted on fab01-2112-dg04 and received zero after a judged critical-criterion failure.
• A separately labeled historical Gemini 3.1 Pro baseline has seven selected outcomes, seven submissions and mean reward 0.3124. The six-model comparison above excludes this baseline.

Record scope and redaction

run-records.tar.gz contains sanitized records of all 16 jobs and 86 attempts: the comparison above, the historical baseline, failures, recoveries and separately labeled execution QA/no-op/stub runs. Numerical rewards, score breakdowns, model identities and selection relationships are retained. Redactions make this a derivative of the original logs. QA/no-op/stub rewards are execution checks and must not be interpreted as model performance.

run-index.json links stable pseudonymous jobs and trials to archive paths and selected outcomes. evaluation-protocol.json records historical settings and provenance; its historical task digests identify the versions evaluated at the time. redaction-summary.json describes replaced identities, infrastructure and source excerpts, including copies in trajectories and terminal recordings. Original private records and the private identity mapping remain private. Native Harbor job pages retain separate access controls.

Limits

Deterministic local judge smoke runs exercised dataroom, submission, verifier and scoring mechanics. Their scope excludes finance-grading quality and replication of the historical rankings. The release has no repeated-run variance estimate, contamination-free claim or independent certification.

This is a small, single-case pilot with one selected outcome per model/task. Its scope excludes general financial knowledge, spreadsheet/model-building skill and long-horizon workflows. Small score differences are not established as statistically significant. Reference answers are public; disclose prior exposure when reporting future evaluations and keep independent holdouts separate.

Running and rights

Use your own Docker host, agent/provider access and judge credentials. Runs require no private proxy, Dissei-operated grading service or bundled commercial agent CLI. Model and judge calls can incur charges and transmit prompts, evidence or answers to your chosen provider; select endpoints and spending limits deliberately. Local runs do not automatically upload results.

The owner has confirmed authority to redistribute the included source material. Downstream training, adaptation and redistribution remain subject to LICENSE.md, copyright and source-specific notices. Obtain any rights not already granted by those terms. Public access does not grant unrestricted reuse.

Published destinations and access

• GitHub repository and version 2.0.0 release: public; task files and historical attachments are in the repository, and the three Docker archives are release assets.
• Harbor dataset: public dissei/financial-judgment-full@2.0.0, with seven public task packages.
• Hugging Face files: public mirror with manual gating; login and approved access are required for file downloads and the gated viewer. GitHub and Harbor copies remain publicly accessible.

Existing private package/image histories and original records remain private and unchanged. publication-plan.md records the plan preceding publication of version 2.0.0. Native Harbor job pages and interactive trajectories retain separate access controls.

Questions, licensing and private security reports: tech@dissei.credit. When reporting a run, identify task and image digests, model identities, harness/judge settings, attempts, selection and failures. Do not publish credentials or unreviewed logs.