{
  "commitment": "We will continue to add verified studies, benchmark-level harness changes and revised syntheses as new evidence becomes available. Each evidence release will be dated and traceable; the paper’s frozen snapshot will remain identifiable.",
  "cadence": "AI-assisted literature searches are scheduled daily at 09:00 Pacific time. Verified additions and corrections update the bibliography, collection charts, repository overview and website after validation. Literature, evidence and presentation retain separate dates; a scheduled search does not guarantee new studies or complete coverage.",
  "steps": [
    {
      "title": "Find & screen",
      "text": "Search daily for new studies and revisions, with overlapping date windows; verify primary sources and distinguish reusable harness changes from model training and background work."
    },
    {
      "title": "Trace the change",
      "text": "Record the setting, seed, actual edit, result and source location. Keep missing artifacts, regressions and transfer conditions visible."
    },
    {
      "title": "Revise & release",
      "text": "Rebuild the collection and website, validate changes, and push a traceable commit. Revise benchmark comparisons only when new evidence supports them; preserve the paper snapshot."
    }
  ],
  "releases": [
    {
      "id": "project-video-2026-10-22",
      "date": "2026-10-22",
      "kind": "website",
      "title": "90-second project overview video in the README",
      "changes": [
        "Added the maintainer-selected 90-second silent Harness RSI overview video, preserving the original 1080 × 1350 MP4 without recompression.",
        "Added a video section and navigation link to the generated README, with a full-length 720 × 900 inline preview and original-file download links.",
        "Preserved literature membership, research dates, frozen evidence, dataset v0.1.0 and manuscript versions."
      ],
      "records": 95
    },
    {
      "id": "revised-author-preprint-2026-10-20",
      "date": "2026-10-20",
      "kind": "website",
      "title": "Revised GitHub reading PDF: ten studies added in a non-ACL layout",
      "changes": [
        "Published the 74-page October 20 author reading version with 128 references, adding Harness² and nine related studies to the main text and a dated source-traceable appendix. The PDF uses an independent single-column article layout, not the ACL template.",
        "Switched README and website reading links to the revised PDF; retained the October 19 dated PDF as an archive. This release does not update ARR, Overleaf or Preprints.org submission files and does not claim platform publication or peer review.",
        "Preserved the October 1 frozen evidence, all 95 category–source records, dataset v0.1.0, bibliography membership and search horizon. Targeted literature synthesis through October 20 is separate from frozen benchmark evidence."
      ],
      "records": 95
    },
    {
      "id": "literature-synthesis-2026-10-20",
      "date": "2026-10-20",
      "kind": "website",
      "title": "Dated synthesis: recursive reuse, regressions and memory-training boundaries",
      "changes": [
        "Connected ten already-collected studies in a source-linked comparison supplement checked through October 20, with inspected versions, method locations and attribution limits.",
        "Distinguished task-local harness search from retained improvement guidance, target repair from net gains and collateral breakage, and memory content from reusable memory procedures and trained policies.",
        "Kept the October 19 public PDF and submitted manuscripts unchanged. Frozen October 1 evidence, 95 records, dataset v0.1.0, collection membership and search horizon are unchanged; no new benchmark discovery or complete failure denominator is claimed."
      ],
      "records": 95
    },
    {
      "id": "public-author-preprint-2026-10-20",
      "date": "2026-10-20",
      "kind": "website",
      "title": "Read the author preprint: public PDF without the ACL template",
      "changes": [
        "Published the maintainer-authorized 19 October 2026 author preprint PDF in the public repository and website. This is the same 71-page independent single-column PDF submitted to Preprints.org, not an ACL review or preprint template.",
        "Added working paper-reading links to the repository overview, website hero and project resources. The author-hosted preprint is not peer-reviewed; no platform-issued DOI or journal/conference acceptance is claimed.",
        "Preserved all five authors, equal-contribution and Project Lead roles, corresponding-author email, and the paper’s October 1 evidence cutoff. Literature membership, search horizon, 95-record frozen evidence and dataset v0.1.0 are unchanged."
      ],
      "records": 95
    },
    {
      "id": "literature-update-2026-10-20",
      "date": "2026-10-20",
      "kind": "website",
      "title": "Literature backfill: memory skills, retrieval evolution and harness-training boundaries",
      "changes": [
        "Newly collected five previously released works: MemSkill, ERSkill, SESA, Harness Evolution Hits a Ceiling, and Self-Evolving Agent Harness. These are newly collected papers, not five papers published today.",
        "Separated reusable memory-construction and retrieval programs from trained controllers and solver policies; categorized the self-evolving harness definition paper as conceptual background rather than an empirical benchmark discovery.",
        "Checked current arXiv versions; ERSkill author Guangxu Zhu follows the primary full text rather than its misspelled abstract metadata. MemSkill NeurIPS 2026 acceptance remains explicitly author-reported.",
        "Living collection now contains 269 papers and 11 resources; targeted overlapping search through October 20. Frozen 95-record evidence, dataset v0.1.0, and submission manuscript remain unchanged."
      ],
      "records": 95
    },
    {
      "id": "literature-update-2026-10-19",
      "date": "2026-10-19",
      "kind": "website",
      "title": "Monday literature backfill: reusable memories, tool documentation and failure-aware evaluation",
      "changes": [
        "Newly collected six existing studies after overlapping discovery and Monday survey/core-paper reference follow-up: MUSE, ContDa, AgentFactory executable subagents, EvoHarness-RL, Does Failure Feedback Direct Agent Self-Modification?, and Two Perspectives on Recursive Self-Improvement. These are newly collected works, not six papers published today.",
        "Merged MUSE's differently titled 2025 preprint and final ACL Findings paper, and AgentFactory's preprint and ACL demonstration version, as one study each. EvoHarness-RL uses its current v2 title and authors and is categorized as policy training plus external-experience evolution; its workshop acceptance is author-reported.",
        "Kept ContDa's release at verified proceedings-month precision because an earlier posting was not verified. Mem2Evolve remains pending while its earlier OpenReview posting date is unresolved.",
        "Living collection now contains 264 papers and 11 resources; search horizon advances to October 19. The frozen 95-record evidence set, dataset v0.1.0 annotations and submission manuscript remain unchanged."
      ],
      "records": 95
    },
    {
      "id": "wenxin-github-confirmed-2026-10-18",
      "date": "2026-10-18",
      "kind": "website",
      "title": "Verified Wenxin Long's GitHub profile through PSU identity sources",
      "changes": [
        "Confirmed Wenxin Long's GitHub account as VivLon: the PSU Statistics directory email matches the named author email in spVelo's setup.py and public commits attributed to VivLon.",
        "Added the verified profile link to the repository overview and homepage. All five authors now have GitHub links; author order, roles, literature and frozen evidence are unchanged."
      ],
      "records": 95
    },
    {
      "id": "author-profiles-2026-10-18",
      "date": "2026-10-18",
      "kind": "website",
      "title": "Survey author roles and verified GitHub profiles",
      "changes": [
        "Listed all five survey authors in manuscript order on the repository overview and homepage, preserving co-first authorship, Project Lead, and corresponding-author roles and contact email.",
        "Linked verified GitHub profiles for Xin Yu, Chengyuan Liu, Haochen Zhang and Lingzhou Xue. Wenxin Long remains listed without a GitHub link pending identity confirmation.",
        "Kept research authorship separate from project and AI assistance; did not fabricate commits or alter GitHub's automatic contribution graph. Literature and frozen evidence are unchanged."
      ],
      "records": 95
    },
    {
      "id": "dataset-pilot-2026-10-18",
      "date": "2026-10-18",
      "kind": "website",
      "title": "Evidence dataset v0.1.0 and supplementary literature screening",
      "changes": [
        "Published a versioned additive dataset view of the unchanged 95 category-source records, with JSON, JSONL, CSV, schema, checksums, citation and a dataset card. Dataset units are not independent interventions.",
        "Added a selected structured-annotation pilot: four records, six scoped signals, with explicit outcome, retention, attribution and source links. The other 91 records remain unreviewed for these labels; no failure-rate or complete-annotation claim is made.",
        "Added homepage dataset downloads and outcome comparisons, distinguishing regressions, no-gain proposals, trade-offs and usage observations. No dataset-specific reuse license is granted.",
        "Newly collected five previously omitted papers: HarnessSR, HarnessTTT, HarnessNet, Harness^2 and WorkflowOps, first submitted October 6-12. Total living collection: 258 papers and 11 resources; ten papers newly collected across today’s two updates.",
        "Bibliography additions do not extend the frozen evidence cutoff, its 95 records, or the submission manuscript. Unresolved candidates remain in local screening notes."
      ],
      "records": 95
    },
    {
      "id": "literature-update-2026-10-18",
      "date": "2026-10-18",
      "kind": "website",
      "title": "Workflow reuse and skill-generalization literature update",
      "changes": [
        "Newly collected five existing studies after primary-source screening: CB-Orchestrator, LAWA, DAGO, InFlowOp and Generalizable Skill Optimization. These are newly collected older works, not five newly published papers.",
        "Separated reusable workflow search and skill-generalization evidence from InFlowOp, which is categorized as adjacent task-time workflow repair rather than a retained cross-task harness discovery.",
        "Deduplicated workshop and arXiv versions of CB-Orchestrator and DAGO as one study each. Their exact earliest public posting dates remain unresolved and are recorded with year precision instead of treating October arXiv dates as first publication.",
        "Completed overlapping targeted discovery for October 10-18 and retained unresolved candidates locally. Refreshed collection statistics and downloads; the 95-record evidence snapshot, manuscript cutoff and benchmark findings are unchanged."
      ],
      "records": 95
    },
    {
      "id": "literature-update-2026-10-17",
      "date": "2026-10-17",
      "kind": "website",
      "title": "First daily literature update",
      "changes": [
        "Added 13 primary-source-checked preprints to the living collection: 11 harness, skill, memory or improver methods, plus two separately categorized infrastructure and reliability studies.",
        "Recorded first-submission dates and v1 method locators. Skill-V and EvoAlloc were first submitted on October 8 and are newly collected older papers, not new publications today. Other additions were first submitted October 10–15.",
        "Extended the completed search horizon from October 10 to October 17, with overlapping discovery queries and unresolved candidates retained for later screening. Updated collection statistics and downloads; the 95-record frozen evidence snapshot and its cutoff are unchanged."
      ],
      "records": 95
    },
    {
      "id": "daily-maintenance-2026-10-16",
      "date": "2026-10-16",
      "kind": "website",
      "title": "Daily literature discovery is scheduled",
      "changes": [
        "Scheduled daily AI-assisted discovery with primary-source checks, version deduplication and weekly searches for overlooked studies.",
        "Documented incremental bibliography updates, validation, automatic commits and pushes, and verification of the deployed website.",
        "Added explicit non-interactive publication checks and recovery guidance, separating search completion, push, deployment and scheduled-run verification.",
        "Maintenance setup only: this release adds no papers or findings and does not advance the existing search horizon or frozen evidence cutoff."
      ],
      "records": 95
    },
    {
      "id": "public-launch-2026-10-15",
      "date": "2026-10-15",
      "kind": "website",
      "title": "The research atlas is public",
      "changes": [
        "Opened the project repository and interactive website to all readers, with public paper-suggestion and evidence-correction forms.",
        "Enabled validated website deployment from the main branch and added the public homepage link to the repository.",
        "Public release only: the evidence snapshot, literature-search horizon and research findings are unchanged."
      ],
      "records": 95
    },
    {
      "id": "comparison-studio-2026-10-11",
      "date": "2026-10-11",
      "kind": "website",
      "title": "Compare mechanisms, not just domain labels",
      "changes": [
        "Added a two-domain comparison studio for interface repairs, decision evidence and recovery, with concrete processing diagrams and linked evidence.",
        "Added separate paper, collection and website dates, plus explicit curation and correction policies.",
        "Reworked the repository overview around the survey’s questions, deliverables and maintenance commitments. This presentation update adds no studies or experimental findings."
      ],
      "records": 95
    },
    {
      "id": "project-authors-2026-10-10",
      "date": "2026-10-10",
      "kind": "website",
      "title": "Meet the survey authors",
      "changes": [
        "Added the five confirmed manuscript authors, in order, and their shared Penn State Statistics affiliation to the website and repository homepage.",
        "Separated survey authorship from project development and AI assistance. The preprint link remains forthcoming.",
        "Project information update only; no new literature search or evidence revision is claimed."
      ],
      "records": 95
    },
    {
      "id": "website-spatial-2026-10-10",
      "date": "2026-10-10",
      "kind": "website",
      "title": "Explore the editable layer",
      "changes": [
        "Added an interactive spatial cover with four selectable harness surfaces and direct links to their evidence.",
        "Opened four domain gateways and refined the result spotlights; animation supports pause and reduced-motion preferences.",
        "Presentation update only: evidence counts, original findings and the paper snapshot are unchanged."
      ],
      "records": 95
    },
    {
      "id": "website-observatory-2026-10-10",
      "date": "2026-10-10",
      "kind": "website",
      "title": "A closer look at the evolving harness",
      "changes": [
        "Redesigned the homepage with an interactive harness diagram, three source-linked result spotlights, and an explorable improvement workflow.",
        "Added benchmark shortcuts and a reading guide for attribution, evidence strength and reuse conditions.",
        "This is a presentation update: the 95 domain records, 235 collected papers and original evidence cutoff are unchanged."
      ],
      "records": 95
    },
    {
      "id": "evidence-2026-10-10",
      "date": "2026-10-10",
      "title": "Benchmark evidence, opened up",
      "changes": [
        "Added all 95 domain records with setting/seed, reported changes and evidence limits.",
        "Added 27 comparison lenses and benchmark/setting filters across four domains.",
        "Separated transfer settings, single-source observations and reporting gaps; clarified that ModularRSI’s detailed Go result is external to Terminal-Bench."
      ],
      "records": 95,
      "kind": "evidence"
    }
  ],
  "policy": [
    {
      "title": "Trace every substantive change",
      "text": "New evidence identifies a task setting, starting harness, reusable edit, observed result and primary-source location. Reporting gaps remain explicit."
    },
    {
      "title": "Keep versions and units distinct",
      "text": "A revised paper updates its source record. One artifact evaluated on several settings is not counted as several independent discoveries."
    },
    {
      "title": "Publish corrections with context",
      "text": "A correction records what changed and why. Conflicting code, unclear retention and negative results remain attached to the original setting."
    },
    {
      "title": "Preserve the paper snapshot",
      "text": "The frozen survey cutoff stays identifiable. A bibliography import or website redesign does not silently extend the evidence collection."
    },
    {
      "title": "Verify publication and recover failures",
      "text": "Updates are complete only after validation, a matching remote commit and a successful website deployment. Failed searches or publications retain separate status and are retried; neither a schedule nor an authentication test counts as a completed update."
    }
  ]
}