Restore original Dissei organization profile

README.md

@@ -8,53 +8,30 @@ pinned: false
 ---
 
 # Dissei Data
-
-**We explore financial reasoning and analysis for institutional investment and credit decisions.**
-
-Dissei develops task environments in which an agent reads transaction evidence, weighs trade-offs and produces an analytical judgment. Our work connects financial reasoning with evidence retrieval, tool use and rubric-based evaluation.
-
-## EBTIDA — explore the sample
-
-[**Open the seven task previews and recorded results on Harbor**](https://hub.harborframework.com/datasets/dissei/ebitda).
-
-EBTIDA presents seven questions from one completed private-equity transaction, one per reasoning family. It is a sample of a 70-task environment, not the full set or a claim about general financial ability. Company and person names are suppressed in the task material.
-
-### Task questions
-
-| Reasoning 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) |
-
-Each public task page contains its question, brief and metadata. The previews do not include case evidence or answer keys and are not runnable evaluations.
-
-### Recorded model results
-
-| Model | Reward / 100 | Submitted |
-|---|---|---|
-| Claude Fable 5.1 | 60.73 | 7/7 |
-| GPT-6 Astra | 52.45 | 7/7 |
-| Claude Opus 5.5 | 52.17 | 7/7 |
-| GLM-5.3 | 44.91 | 7/7 |
-| Kimi K3 | 35.79 | 7/7 |
-| Gemini 3.8 Flash | 9.01 | 1/7 |
-
-Scores are the equal-weight mean of seven saved continuous task rewards, multiplied by 100—not accuracy. They come from the complete controlled-access tasks, not the public previews. The comparison uses Harbor 0.23.0 and the Terminus-2 JSON harness, with 20 dataroom calls, 40 agent turns per task, a 30-minute task limit and a 16,384-token output limit per request. The fixed judge is `gemini-3.1-pro-preview`, with low reasoning and one vote; agent reasoning is high.
-
-**Interpretation matters.** Gemini 3.8 Flash submitted one of seven selected answers; six tasks reached the **40-turn-per-task cap**. Five tasks were rerun after a gateway authentication fix. GLM-5.3 includes one recovered task, with its earlier VM-interrupted recovery retained in the private record. These are single-case results with one selected outcome per task, no repeated-run error bars, and dependence on the judge, harness and provider framing. They are not an external certification or a ranking of general financial ability.
-
-## Research access
-
-We publish the sample questions, results and methodology so researchers can understand the work before requesting access. Full case evidence, execution environments, rubrics, reference answers, evaluator source and run records remain controlled-access materials.
-
-After a mutual NDA and review terms are agreed, approved researchers can receive the exact evaluated task versions, submitted answers, trajectories, grading breakdowns and original failure/recovery records. The private **EBITDA PRIVATE** package contains the full sample write-up and saved evaluation records. Dissei grants access explicitly; signing an NDA does not automatically unlock a repository.
-
-Evaluation access does not grant training or redistribution rights. Those rights, retention and permitted use are defined by agreement. We do not publish private benchmark files or raw trajectories on this organization card.
-
-**Methods & findings:** [dissei.ai/research](https://dissei.ai/research)  
-**Contact:** [info@dissei.credit](mailto:info@dissei.credit) · [Request research access](https://dissei.ai/contact)
+**Outcome-graded RL environments for institutional financial judgment.**
+
+Dissei builds reinforcement-learning environments from real institutional credit
+decisions — each one a documented transaction rebuilt as an agentic task, worked with
+analyst tools and graded against what actually happened.
+
+- **Point-in-time by construction.** A model sees only what was knowable at the decision
+  date; every exhibit is anchor-gated, so hindsight is excluded structurally — not by an
+  instruction it can ignore.
+- **Graded, not checked.** Reward decomposes into fail-closed critical gates, an
+  expert-authored graded rubric, a contradiction penalty, and a retrieval modulator that
+  rewards grounding in the record.
+- **Grading automation, disclosed.** Rubrics are authored and anchored by credit
+  practitioners who carried real risk; execution runs on LLM judges at temperature 0 that
+  apply the sealed rubric and never invent criteria. Answers are graded twice — against
+  best practice, and against the realized outcome.
+- **Sealed by architecture.** Answer material is time-locked and anchor-gated; every asset
+  ships with a per-asset contamination statement and right-to-train lineage.
+- **Operators, not an annotation farm.** The corpus is the by-product of an operating
+  lending business — incentive-bearing decisions with real consequences, not opinions
+  commissioned for a dataset.
+
+Environments (RL post-training) and sealed evaluations are available to frontier labs
+**under agreement**. We do not publish tasks, rubrics, graders, or leaderboards.
+
+**Methods & findings:** https://dissei.ai/research
+**Contact:** info@dissei.credit · Mutual NDA before anything sensitive.

index.html

@@ -1,108 +1,4 @@
 <!doctype html><html lang="en"><head><meta charset="utf-8"><title>Dissei Data</title>
 <meta http-equiv="refresh" content="0; url=https://dissei.ai/research"></head>
-<body>
-<h1>Dissei Data</h1>
-<p><strong>We explore financial reasoning and analysis for institutional investment and credit decisions.</strong></p>
-<p>Dissei develops task environments in which an agent reads transaction evidence, weighs trade-offs and produces an analytical judgment. Our work connects financial reasoning with evidence retrieval, tool use and rubric-based evaluation.</p>
-<h2>EBTIDA — explore the sample</h2>
-<p><a href="https://hub.harborframework.com/datasets/dissei/ebitda"><strong>Open the seven task previews and recorded results on Harbor</strong></a>.</p>
-<p>EBTIDA presents seven questions from one completed private-equity transaction, one per reasoning family. It is a sample of a 70-task environment, not the full set or a claim about general financial ability. Company and person names are suppressed in the task material.</p>
-<h3>Task questions</h3>
-<table>
-<thead>
-<tr>
-<th>Reasoning family</th>
-<th>Information cutoff</th>
-<th>Question</th>
-</tr>
-</thead>
-<tbody>
-<tr>
-<td>Diagnostic</td>
-<td>2021-12</td>
-<td><a href="https://hub.harborframework.com/tasks/dissei/finance-preview-dg04">What diagnosis should you make of Company CI recent margin pressure at the anchor?</a></td>
-</tr>
-<tr>
-<td>Predictive</td>
-<td>2021-12</td>
-<td><a href="https://hub.harborframework.com/tasks/dissei/finance-preview-pr02">Would you expect Company CI to be relatively resilient if demand conditions weakened after the anchor?</a></td>
-</tr>
-<tr>
-<td>Explanatory</td>
-<td>2022-06</td>
-<td><a href="https://hub.harborframework.com/tasks/dissei/finance-preview-ex01">Why did Company CI recent margin pressure matter for the investment case at the anchor?</a></td>
-</tr>
-<tr>
-<td>Quantitative</td>
-<td>2022-06</td>
-<td><a href="https://hub.harborframework.com/tasks/dissei/finance-preview-qn02">What does the segment-share exhibit imply about Company CI relative market position?</a></td>
-</tr>
-<tr>
-<td>Counterfactual</td>
-<td>2022-09</td>
-<td><a href="https://hub.harborframework.com/tasks/dissei/finance-preview-cf01">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?</a></td>
-</tr>
-<tr>
-<td>Comparative</td>
-<td>2022-09</td>
-<td><a href="https://hub.harborframework.com/tasks/dissei/finance-preview-cp02">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?</a></td>
-</tr>
-<tr>
-<td>Strategic</td>
-<td>2022-09</td>
-<td><a href="https://hub.harborframework.com/tasks/dissei/finance-preview-st06">Which diligence workstream should Company ED treat as gating before signing the carve-out?</a></td>
-</tr>
-</tbody>
-</table>
-<p>Each public task page contains its question, brief and metadata. The previews do not include case evidence or answer keys and are not runnable evaluations.</p>
-<h3>Recorded model results</h3>
-<table>
-<thead>
-<tr>
-<th>Model</th>
-<th>Reward / 100</th>
-<th>Submitted</th>
-</tr>
-</thead>
-<tbody>
-<tr>
-<td>Claude Fable 5.1</td>
-<td>60.73</td>
-<td>7/7</td>
-</tr>
-<tr>
-<td>GPT-6 Astra</td>
-<td>52.45</td>
-<td>7/7</td>
-</tr>
-<tr>
-<td>Claude Opus 5.5</td>
-<td>52.17</td>
-<td>7/7</td>
-</tr>
-<tr>
-<td>GLM-5.3</td>
-<td>44.91</td>
-<td>7/7</td>
-</tr>
-<tr>
-<td>Kimi K3</td>
-<td>35.79</td>
-<td>7/7</td>
-</tr>
-<tr>
-<td>Gemini 3.8 Flash</td>
-<td>9.01</td>
-<td>1/7</td>
-</tr>
-</tbody>
-</table>
-<p>Scores are the equal-weight mean of seven saved continuous task rewards, multiplied by 100—not accuracy. They come from the complete controlled-access tasks, not the public previews. The comparison uses Harbor 0.23.0 and the Terminus-2 JSON harness, with 20 dataroom calls, 40 agent turns per task, a 30-minute task limit and a 16,384-token output limit per request. The fixed judge is <code>gemini-3.1-pro-preview</code>, with low reasoning and one vote; agent reasoning is high.</p>
-<p><strong>Interpretation matters.</strong> Gemini 3.8 Flash submitted one of seven selected answers; six tasks reached the <strong>40-turn-per-task cap</strong>. Five tasks were rerun after a gateway authentication fix. GLM-5.3 includes one recovered task, with its earlier VM-interrupted recovery retained in the private record. These are single-case results with one selected outcome per task, no repeated-run error bars, and dependence on the judge, harness and provider framing. They are not an external certification or a ranking of general financial ability.</p>
-<h2>Research access</h2>
-<p>We publish the sample questions, results and methodology so researchers can understand the work before requesting access. Full case evidence, execution environments, rubrics, reference answers, evaluator source and run records remain controlled-access materials.</p>
-<p>After a mutual NDA and review terms are agreed, approved researchers can receive the exact evaluated task versions, submitted answers, trajectories, grading breakdowns and original failure/recovery records. The private <strong>EBITDA PRIVATE</strong> package contains the full sample write-up and saved evaluation records. Dissei grants access explicitly; signing an NDA does not automatically unlock a repository.</p>
-<p>Evaluation access does not grant training or redistribution rights. Those rights, retention and permitted use are defined by agreement. We do not publish private benchmark files or raw trajectories on this organization card.</p>
-<p><strong>Methods &amp; findings:</strong> <a href="https://dissei.ai/research">dissei.ai/research</a><br />
-<strong>Contact:</strong> <a href="mailto:info@dissei.credit">info@dissei.credit</a> · <a href="https://dissei.ai/contact">Request research access</a></p>
-</body></html>
+<body><p>Dissei Data — outcome-graded RL environments for institutional financial judgment.
+Methods &amp; findings: <a href="https://dissei.ai/research">dissei.ai/research</a></p></body></html>