Add reviewed DRB-II development calibration

README.md

@@ -0,0 +1,37 @@
+---
+pretty_name: Sponge research evaluations
+license: cc-by-4.0
+language:
+- en
+task_categories:
+- text-generation
+tags:
+- evaluation
+- deep-research
+size_categories:
+- n<1K
+---
+
+# Sponge research evaluations
+
+This archive contains selected numeric research evaluations from
+[Sponge V2](https://sponge.computer). Each study includes its method, source
+identities, accounting scope and limits. Study IDs identify fixed results;
+corrections receive separate records.
+
+| Study | Comparison | Result | Scope |
+| --- | --- | --- | --- |
+| [DRB-II development calibration, 27 September 2026](studies/drb2-exposed-calibration-20260927-v1/method.md) | Revision versus revision with more retrieved evidence | Mean criterion satisfaction: 20.28% versus 32.33% | Two previously exposed tasks, four retained reports, one official-format evaluation per report |
+
+The calibration compares Sponge variants using a partial retained source corpus.
+It supports a descriptive result on those two tasks. Performance on unseen
+tasks and comparisons with external research systems require separate studies.
+
+Files contain authored methods, numeric results and source hashes. Research
+reports, retrieved documents, benchmark questions and rubrics, provider
+responses, and account records are excluded. The earlier incomplete adapted
+evaluation is recorded alongside the separate calibration.
+
+See each study's `metrics.json`, `provenance.json` and `rights.md` for its data,
+source identities and attribution. The authored export is licensed under
+[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

studies/drb2-exposed-calibration-20260927-v1/method.md

@@ -0,0 +1,72 @@
+# Reports with evidence refresh scored higher on two development tasks
+
+On two exposed DeepResearch-Bench-II tasks, Sponge V2 reports using evidence
+refresh averaged 32.33% criterion satisfaction, compared with 20.28% for revision
+alone. The difference was 12.06 percentage points. Both task-level differences
+were positive.
+
+| Task | Criteria | Revision satisfied | Refresh satisfied | Revision score | Refresh score |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| `drb2-10` | 67 | 13 | 24 | 19.40% | 35.82% |
+| `drb2-16` | 52 | 11 | 15 | 21.15% | 28.85% |
+
+## Comparison
+
+Both variants generated a draft, critique and final revision with
+`openai/gpt-5.4-mini`. They used the same partial retained corpus of six sources
+per task, with a 24,000-byte final-report limit. The refresh variant supplied
+16 additional passages before revision. Source acquisition was outside this
+comparison.
+
+The questions and reports had already guided development. All four reports
+were fixed before this calibration; no report was regenerated in response to
+the new grades. There was one report per variant and task and one evaluation
+per report.
+
+## Evaluation
+
+The evaluator used the prompt from
+[DeepResearch-Bench-II revision `087c1b8`](https://github.com/imlrz/DeepResearch-Bench-II/tree/087c1b8d4a0ed46fd3dd8615a0b5e93ce3acf6f8),
+with GPT-5.5, medium reasoning, a 32,768-token completion limit and chunks of
+50 criteria. Requests went through Vercel AI Gateway to OpenAI. The formatter
+matched the pinned upstream prompt byte for byte for all eight requests.
+
+A task's score is the count of criteria assigned `1`, divided by every
+criterion. Scores of `0` and blocked scores of `-1` remain in the denominator.
+The reported mean weights both tasks equally. Literal quotation checks were
+diagnostics and did not alter scores.
+
+The local runner accepted strict JSON or one whole-answer JSON code fence and
+made no repair or retry calls. Gateway routing and this parser/retry policy
+differ from the upstream runner. All eight grading calls completed, covering
+all four reports and 238 criterion evaluations. An invented two-criterion
+canary completed first with the expected scores.
+
+An independent Codex agent checked report, request, rubric and response hashes,
+upstream formatting, criterion coverage and score arithmetic. That check did
+not assess the substantive truth of every report claim.
+
+## Cost and time
+
+The eight grading calls cost $1.245172; the canary cost $0.012720. All nine
+charges were verified against provider generation metadata, for a total of
+$1.257892. These figures cover the new calibration only. Original report
+generation and source preparation are outside that cost total.
+
+`metrics.json` includes each original report's recorded generation duration.
+The calibration adapter did not record per-call wall-clock durations, so its
+elapsed-time field is null.
+
+## Limits and earlier evaluation
+
+This is a descriptive development result from two exposed tasks and a partial
+source corpus. It does not estimate unseen-task performance, end-to-end web
+research quality, repeated-run variability or performance against another
+research system. The official-format criterion scores also do not independently
+verify every citation and substantive claim.
+
+An earlier adapted mini-model evaluation of these reports completed 17 of
+24 grading batches; three failed and four were not run after earlier failures.
+Only one report received complete grading. That evaluation remains incomplete,
+with `analysisQualified: false` and no paired comparison. The present
+calibration is a separate protocol and does not replace its results.

studies/drb2-exposed-calibration-20260927-v1/metrics.json

@@ -0,0 +1,101 @@
+{
+  "contract": "sponge.public-study-metrics.v1",
+  "studyId": "drb2-exposed-calibration-20260927-v1",
+  "date": "2026-09-27",
+  "protocolId": "drb2-official-format-gpt55-calibration-v1",
+  "comparisonScope": "descriptive paired calibration of retained reports on exposed development tasks",
+  "analysisQualified": true,
+  "assignedCases": [
+    "drb2-10",
+    "drb2-16"
+  ],
+  "exposure": "development",
+  "assignedReports": 4,
+  "completeReports": 4,
+  "failedReports": 0,
+  "missingJudgments": 0,
+  "gradingCalls": 8,
+  "failedGradingCalls": 0,
+  "unrunGradingCalls": 0,
+  "rows": [
+    {
+      "caseId": "drb2-10",
+      "armId": "coverage-v1",
+      "status": "complete",
+      "criteria": 67,
+      "satisfied": 13,
+      "blocked": 0,
+      "score": 0.19402985074626866,
+      "reportSha256": "03b8b0a75bcf49b4491121105b9b7a79d180aa23bcf80c70270641f55084bbea",
+      "gradingCalls": 2,
+      "gradingCostMicros": 356815,
+      "originalGenerationElapsedMs": 44092
+    },
+    {
+      "caseId": "drb2-10",
+      "armId": "coverage-refresh-v1",
+      "status": "complete",
+      "criteria": 67,
+      "satisfied": 24,
+      "blocked": 0,
+      "score": 0.3582089552238806,
+      "reportSha256": "7ce7ee233daafcbf8bf792c19a41922b8c86c32c18083289a167ef98e63684e7",
+      "gradingCalls": 2,
+      "gradingCostMicros": 433255,
+      "originalGenerationElapsedMs": 49308
+    },
+    {
+      "caseId": "drb2-16",
+      "armId": "coverage-v1",
+      "status": "complete",
+      "criteria": 52,
+      "satisfied": 11,
+      "blocked": 0,
+      "score": 0.21153846153846154,
+      "reportSha256": "a00e9fde17e293bbab0d14775f806eaf203e8bc86b470afe4bc3ca90b51ff234",
+      "gradingCalls": 2,
+      "gradingCostMicros": 219296,
+      "originalGenerationElapsedMs": 37985
+    },
+    {
+      "caseId": "drb2-16",
+      "armId": "coverage-refresh-v1",
+      "status": "complete",
+      "criteria": 52,
+      "satisfied": 15,
+      "blocked": 0,
+      "score": 0.28846153846153844,
+      "reportSha256": "44f726daec3ecdab6d77cd88c42217aaaeb619d36c1aebbb149a584815be03c4",
+      "gradingCalls": 2,
+      "gradingCostMicros": 235806,
+      "originalGenerationElapsedMs": 42941
+    }
+  ],
+  "aggregation": "unweighted mean of task-level satisfied / all criteria",
+  "means": {
+    "coverage-v1": 0.2027841561423651,
+    "coverage-refresh-v1": 0.32333524684270953
+  },
+  "meanDifference": 0.12055109070034442,
+  "cost": {
+    "currency": "USD",
+    "gradingMicros": 1245172,
+    "canaryMicros": 12720,
+    "totalNewMicros": 1257892,
+    "verifiedCalls": 9,
+    "unknownCosts": 0,
+    "scope": "new official grading and invented canary only; original report-generation costs excluded"
+  },
+  "calibrationElapsedMs": null,
+  "calibrationElapsedReason": "The private calibration adapter did not record per-call wall-clock durations.",
+  "priorAdaptedEvaluation": {
+    "protocolId": "adapted-v7",
+    "analysisQualified": false,
+    "plannedBatches": 24,
+    "scoredBatches": 17,
+    "failedBatches": 3,
+    "unrunBatches": 4,
+    "completeReports": 1,
+    "pairedComparison": null
+  }
+}

studies/drb2-exposed-calibration-20260927-v1/provenance.json

@@ -0,0 +1,37 @@
+{
+  "contract": "sponge.public-study-provenance.v1",
+  "studyId": "drb2-exposed-calibration-20260927-v1",
+  "projectUrl": "https://sponge.computer",
+  "benchmark": {
+    "name": "DeepResearch-Bench-II",
+    "url": "https://github.com/imlrz/DeepResearch-Bench-II",
+    "revision": "087c1b8d4a0ed46fd3dd8615a0b5e93ce3acf6f8",
+    "datasetSha256": "263aaabb8c279fb16cbe7c9499afe82d657a8ab3ccfb07ace084387e367d921a",
+    "selectedRubricSha256": "63e12b6f1c8f9a7a98cfa7041dc5ac4682ae26dfc0df8b8ceb03fac47fba48f6"
+  },
+  "evaluator": {
+    "model": "openai/gpt-5.5",
+    "route": "Vercel AI Gateway / OpenAI",
+    "reasoning": "medium",
+    "maximumCompletionTokens": 32768,
+    "criterionBatchSize": 50,
+    "configurationSha256": "300a2d1f2312a2f6232aa470cd7971369ff401a73f643f8da8fb6d43a4a776ea",
+    "spongeTransportSourceRevision": "a887ebd1edbf675fbf8d94aeff5eac53f56f071a",
+    "spongeRepository": "https://github.com/hraness/spongev2",
+    "privateCalibrationRunnerSha256": "7eeeb44154660eebe6d3ddd1e5ef0a657d6f1c45fcd00c53106d91ec122a5b1d",
+    "promptTemplateTextSha256": "2b40577e219d87294e49f9e4ce8e1038556e56d120b871bd1bfd3c327c609a50",
+    "promptTemplateFileSha256": "ff726e03c3841c60e0d85f5f808201a18fd8985eda54114b2ea131117d429f1a"
+  },
+  "generation": {
+    "model": "openai/gpt-5.4-mini",
+    "harnessSha256": "af8fedd6ce5a47d8dbc3bfe88ab842762da08efd0fa09db12f4005e77fb39037",
+    "manifestSha256": "b212bfa5faee7eecb640d16c8e216d72aa7bf04e659829f2a1f634d3d1782ad9",
+    "generationFileSha256": "7e508e21e591899d166c7e0f8b6983721d0cd59c2cc0fbe66f07cd99f74b8961"
+  },
+  "calibrationPlanSha256": "e6dcc20c6f51f76c5157913a5a4c4a38342a8c8d3dd21b2f2fdffb9932946057",
+  "independentAuditSha256": "5435f9db8e2a152f15bd7cfa57f4a3bc127cf671a5759734c8ac225dfcae42f9",
+  "methodDraftedBy": "OpenAI Codex root agent",
+  "methodReviewStatus": "independently reviewed",
+  "generationStyleVersion": "hraness-generation-style/v1",
+  "methodReviewedBy": "OpenAI Codex independent agent (repo_recon)"
+}

studies/drb2-exposed-calibration-20260927-v1/rights.md

@@ -0,0 +1,19 @@
+# Rights and attribution
+
+The numeric observations, method description and archive card were prepared
+for Sponge V2 by OpenAI Codex agents. The export owner is Hraness. These authored
+export files are offered under
+[Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/).
+Attribute the archive as “Sponge research evaluations, Hraness” and retain the
+study ID and revision when citing a result.
+
+DeepResearch-Bench-II is the benchmark used for evaluation. Credit its authors
+through the [upstream project](https://github.com/imlrz/DeepResearch-Bench-II)
+and its requested scholarly citation. Its source revision and dataset hashes
+identify the evaluated material. This export does not redistribute benchmark
+questions, rubrics, reference reports or the upstream evaluator prompt.
+
+Generated research reports, third-party source bodies, raw provider responses,
+private paths, account records and credentials are excluded. Their absence
+limits public reproduction to the numeric aggregation and described method.
+Licensing this authored export grants no rights to those excluded materials.