{
 "batch": "final-01",
 "ranking": [
  {
   "rank": 1,
   "pr": 150556,
   "name": "hermes-ui-cards",
   "total": 28,
   "why": "Answerable cards in Desktop and Webapp with text fallback elsewhere: one tool, a real Desktop renderer, about 330 tests, no secrets or network, clean catalog check and a clear error on a bad card in my Hermes run; the idea itself is a known one and the README's demo-video slot is empty. (Useful 5, Reliable 5, Polish 5, Creative 3, Beyond 5.)"
  },
  {
   "rank": 2,
   "pr": 153714,
   "name": "tenure",
   "total": 28,
   "why": "Rebuilds Hermes's approvals as earned, revocable per-action trust with undo, a signed ledger, Desktop and dashboard pages and the field's biggest test suite; loses a Reliable point for three private core imports flagged by the catalog check, a heuristic shell classifier and an opt-in preset that can let high-stakes actions run unasked. (Useful 4, Reliable 4, Polish 5, Creative 5, Beyond 5.)"
  },
  {
   "rank": 3,
   "pr": 150644,
   "name": "hermes-canvas",
   "total": 27,
   "why": "The agent's plan drawn as a live Desktop canvas whose boxes turn green only from real writes and exit codes and which the user can edit mid-build; fully implemented with 206 tests and a clean catalog check, though 'cannot fake' overstates since the model picks its own check commands. (Useful 4, Reliable 4, Polish 4, Creative 5, Beyond 5.)"
  },
  {
   "rank": 4,
   "pr": 150840,
   "name": "kata",
   "total": 26,
   "why": "Mines the agent's own repeated tool sequences into approved, deterministic kata_ tools, a new idea with tight code, a dashboard and a clean catalog check; approval can be reached from an agent's terminal via the CLI and a new kata is only callable next session. (Useful 4, Reliable 4, Polish 5, Creative 5, Beyond 4.)"
  },
  {
   "rank": 5,
   "pr": 152458,
   "name": "housemate",
   "total": 26,
   "why": "Turns old phones and laptops into room speakers as a real Hermes gateway platform plus TTS provider with timers, intercom and spoken approvals, exactly Nous's own-hardware theme and deeply tested; setup is heavy, the audience is narrower, and anyone in earshot can talk to the agent. (Useful 3, Reliable 4, Polish 5, Creative 4, Beyond 5.)"
  },
  {
   "rank": 6,
   "pr": 152651,
   "name": "workroom",
   "total": 26,
   "why": "A finished Desktop workspace that ties figures in Word, PowerPoint and Excel files to workbook cells and stages owner-approved releases, with 31 tools and about 590 tests; an Office niche validated on Windows only, with guards its own README calls heuristic. (Useful 3, Reliable 4, Polish 5, Creative 4, Beyond 5.)"
  },
  {
   "rank": 7,
   "pr": 150596,
   "name": "ariadne",
   "total": 26,
   "why": "Records a session to a cassette and replays it offline through LLM and tool middleware to show what a Hermes upgrade or plugin changed, with CI against three Hermes versions and a polished Desktop page; mostly for plugin developers and cautious upgraders, with only 21 unit tests. (Useful 3, Reliable 4, Polish 4, Creative 5, Beyond 5.)"
  },
  {
   "rank": 8,
   "pr": 152691,
   "name": "grimoire",
   "total": 25,
   "why": "Locked-down DuckDB over the user's data files plus a check of every number in the reply; attach, query, refused DROP and missing-file error all behaved correctly in my Hermes run and many users would install it, but SQL over files is a familiar idea and the hooks add less than a platform change. (Useful 5, Reliable 4, Polish 5, Creative 3, Beyond 4.)"
  },
  {
   "rank": 9,
   "pr": 153844,
   "name": "fine-print",
   "total": 25,
   "why": "Detects when Hermes silently truncates, downscales or drops what the model sees and tells user and model, with 254 fixture tests, fail-soft hooks and a candid status table; a diagnostic niche pinned to Hermes 0.23 internals with 16 of 81 limits verified live. (Useful 3, Reliable 5, Polish 4, Creative 5, Beyond 4.)"
  },
  {
   "rank": 10,
   "pr": 150800,
   "name": "hypnos",
   "total": 25,
   "why": "Fine-tunes a small local LoRA model overnight from the user's approvals and denials and routes mastered turns to it, the boldest idea here with real SFT/DPO and routing code; habits come from about six hard-coded templates, evidence is a synthetic persona, the real-torch path has one test and its screenshots are mock-ups. (Useful 3, Reliable 3, Polish 4, Creative 5, Beyond 5.)"
  },
  {
   "rank": 11,
   "pr": 153851,
   "name": "foreman",
   "total": 24,
   "why": "A git-diff gate that blocks an agent from skipping, deleting or weakening tests before commit, push or PR, plus second-model review and human-confirmed merge, with about 436 tests and honest limits; one private core import flagged by the catalog check and it only sees commands it can parse. (Useful 4, Reliable 4, Polish 4, Creative 4, Beyond 4.)"
  },
  {
   "rank": 12,
   "pr": 150251,
   "name": "gemini-live",
   "total": 24,
   "why": "Full-duplex voice calls in Desktop and dashboard with the key read correctly from Hermes's secret store and only an ephemeral token in the browser; the Python side registers just a CLI, picker integration needs an unmerged upstream patch, and phone writes skip Hermes approvals by default. (Useful 4, Reliable 3, Polish 4, Creative 3, Beyond 5.)"
  },
  {
   "rank": 13,
   "pr": 153809,
   "name": "living-spaces",
   "total": 24,
   "why": "Agent-built persistent mini-apps in a Desktop Spaces tab with a sandbox, SSRF-safe fetcher and about 600 tests, a genuinely new surface; the catalog check flags rules 8 and 9 (raw window.hermesDesktop bridge, private cron modules), its own README says it is not catalog-approved, and it schedules model-spending jobs without an approval step. (Useful 4, Reliable 2, Polish 3, Creative 5, Beyond 5.)"
  },
  {
   "rank": 14,
   "pr": 153955,
   "name": "shadow",
   "total": 24,
   "why": "Blind A/B of a candidate model on the user's own past turns with a statistical verdict and gradual switch, original and rigorous on paper; the catalog check flags over 40 private core imports, the end-to-end acceptance test is a skip stub, and it needs several providers and real spend. (Useful 3, Reliable 2, Polish 4, Creative 5, Beyond 5.)"
  },
  {
   "rank": 15,
   "pr": 149738,
   "name": "trust-ladder",
   "total": 23,
   "why": "Compiles a user's correction into the strongest enforceable rule (block, check, reminder or approval), compact and defensive with a clean catalog check and clear argument errors in my run; guards are best-effort regex, auto-learn is on by default, there is no UI and the install line still has an <owner> placeholder. (Useful 4, Reliable 4, Polish 3, Creative 4, Beyond 4.)"
  },
  {
   "rank": 16,
   "pr": 149640,
   "name": "hermes-ark",
   "total": 23,
   "why": "Encrypted, deduplicated backups of the Hermes home with item-level restore, secrets read from Hermes's secret store and a Desktop page; backup is a common category, its 'agent cannot switch it off' claim rests on regex matching, and ark_checkpoint was not offered to the model in my fresh-home run. (Useful 4, Reliable 4, Polish 4, Creative 3, Beyond 4.)"
  },
  {
   "rank": 17,
   "pr": 151887,
   "name": "cmi",
   "total": 23,
   "why": "Replaces Hermes's context engine with deterministic digests and byte-exact recall, a deep change with many tests; but it is a Claude Code/Codex proxy project thinly wrapped for Hermes: five-line manifest, six private core imports flagged by the catalog check, and savings that are modelled rather than measured. (Useful 4, Reliable 2, Polish 3, Creative 4, Beyond 5.)"
  },
  {
   "rank": 18,
   "pr": 153020,
   "name": "breadcrumbs",
   "total": 23,
   "why": "A verbatim content-addressed memory provider plus context engine with zoom levels and about 680 tests; the catalog check flags five private core imports, the manifest under-declares what it registers, the benchmark table has no harness in the repo, and the scope and private vocabulary sprawl. (Useful 3, Reliable 2, Polish 4, Creative 4, Beyond 5.)"
  },
  {
   "rank": 19,
   "pr": 150353,
   "name": "fluent",
   "total": 22,
   "why": "Checks the agent's shell commands against the installed programs' own help text and blocks invented flags, a fresh and conservative guard with a clean catalog check; the '542 tests' claim is not visible in the repo (69 test functions plus corpus cases), hook glue is untested and scope is long flags on local backends. (Useful 4, Reliable 3, Polish 3, Creative 4, Beyond 4.)"
  },
  {
   "rank": 20,
   "pr": 152908,
   "name": "contract-necromancer",
   "total": 22,
   "why": "Freezes a project's tests by hash and holds 'done' until they pass, rigorous with about 390 tests and clean, clear tool answers in my run; a coding-only niche with no UI, a headline figure borrowed from another product, and a key file renamed to avoid looking like a credential under catalog rule 11. (Useful 3, Reliable 4, Polish 3, Creative 4, Beyond 4.)"
  }
 ]
}