{
 "batch": "final-09",
 "ranking": [
  {
   "rank": 1,
   "pr": 152458,
   "name": "housemate",
   "total": 29,
   "why": "A real platform plugin that turns any old phone or laptop into a wake-word room speaker for your own Hermes, with TLS pairing, spoken approvals, a Desktop page, 453 tests run in CI on real Hermes and claims tiered by what is reproducible: the entry Nous's own hardware push describes.",
   "scores": {
    "useful": 5,
    "reliable": 4,
    "polish": 5,
    "creative": 5,
    "beyond": 5
   }
  },
  {
   "rank": 2,
   "pr": 153714,
   "name": "tenure",
   "total": 27,
   "why": "Revocable, evidence-based trust that gates every tool call through pre_tool_call directives with HMAC undo and a signed passport; 1,453 of its tests passed here, it ships Desktop and dashboard UIs in five locales, and two test failures plus a 5,200-line classifier are the only reliability doubts.",
   "scores": {
    "useful": 4,
    "reliable": 4,
    "polish": 4,
    "creative": 5,
    "beyond": 5
   }
  },
  {
   "rank": 3,
   "pr": 152908,
   "name": "contract-necromancer",
   "total": 26,
   "why": "Makes 'done' mean the user's own locked tests pass: a pre_verify gate that sends the agent back with the failing test named, file guards on frozen tests, allow-listed child env, about 390 tests, no network; the idea is a known one but the execution is the most fail-closed in the field.",
   "scores": {
    "useful": 4,
    "reliable": 5,
    "polish": 4,
    "creative": 3,
    "beyond": 5
   }
  },
  {
   "rank": 4,
   "pr": 153851,
   "name": "foreman",
   "total": 26,
   "why": "A deterministic no-model gate that blocks commits, pushes and PRs when the diff skips or deletes tests or carries a secret, never lets the agent merge, and asks a second model for review via ctx.llm; 383 tests passed here and the README's known-limits section is candid.",
   "scores": {
    "useful": 4,
    "reliable": 4,
    "polish": 4,
    "creative": 4,
    "beyond": 5
   }
  },
  {
   "rank": 5,
   "pr": 150644,
   "name": "hermes-canvas",
   "total": 26,
   "why": "A hand-drawn plan canvas beside the chat whose boxes are bound to real files and check exit codes, with seven hooks including pre_verify and transform_tool_result for mid-turn steering; 306 tests passed here, nothing leaves the machine, and it is already merged.",
   "scores": {
    "useful": 4,
    "reliable": 4,
    "polish": 4,
    "creative": 4,
    "beyond": 5
   }
  },
  {
   "rank": 6,
   "pr": 150596,
   "name": "ariadne",
   "total": 26,
   "why": "Records any session and replays it offline through llm_execution and tool middleware, with bisect, fork and hook-contract checks that found a real Hermes regression; the audience is Hermes and plugin developers rather than every user, and one Desktop route can block the event loop offline.",
   "scores": {
    "useful": 3,
    "reliable": 4,
    "polish": 4,
    "creative": 5,
    "beyond": 5
   }
  },
  {
   "rank": 7,
   "pr": 149738,
   "name": "trust-ladder",
   "total": 25,
   "why": "Compiles each user correction into the strongest provable guard (block, post-write check, injected rule, or review), verifies it against the mistake first, has TLA+ specs and 96 passing tests, and reads nothing it should not; no Desktop surface, but it is the plugin most users would keep without noticing it.",
   "scores": {
    "useful": 4,
    "reliable": 5,
    "polish": 4,
    "creative": 4,
    "beyond": 4
   }
  },
  {
   "rank": 8,
   "pr": 152691,
   "name": "grimoire",
   "total": 25,
   "why": "Locked-down DuckDB over CSV, JSON, Parquet and SQLite with a footer that traces every number in the answer to a real query, a Desktop Data page and a CI self-test; polished and useful, though its tests live in one selftest script and committed .pyc files slipped in.",
   "scores": {
    "useful": 4,
    "reliable": 4,
    "polish": 5,
    "creative": 4,
    "beyond": 4
   }
  },
  {
   "rank": 9,
   "pr": 150251,
   "name": "gemini-live",
   "total": 25,
   "why": "Full-duplex Gemini Live voice calls in Desktop and the dashboard that delegate work to your own agent, key taken from Hermes's secret store; polished with Mac and uninstall docs, but the opt-in run_readonly path executes voice-model commands outside Hermes's approval prompt and one shell-rule test failed here.",
   "scores": {
    "useful": 4,
    "reliable": 3,
    "polish": 4,
    "creative": 4,
    "beyond": 5
   }
  },
  {
   "rank": 10,
   "pr": 150800,
   "name": "hypnos",
   "total": 25,
   "why": "Real LoRA code (PEFT, SFT and DPO losses) that practises the day's corrections in a sandbox and routes mastered turns to a local brain through llm_execution middleware; the most original entry, but it needs a trainer environment and spends model tokens at night, so few users will keep it running.",
   "scores": {
    "useful": 3,
    "reliable": 3,
    "polish": 4,
    "creative": 5,
    "beyond": 5
   }
  },
  {
   "rank": 11,
   "pr": 153809,
   "name": "living-spaces",
   "total": 25,
   "why": "Persistent sandboxed Spaces that bots build in chat and keep current through Hermes cron, with 394 tests passing here; held back by macOS-only verification, private cron imports, minified vendored JS with admitted provenance gaps, and a README that calls itself not catalog-approved.",
   "scores": {
    "useful": 4,
    "reliable": 3,
    "polish": 3,
    "creative": 5,
    "beyond": 5
   }
  },
  {
   "rank": 12,
   "pr": 153955,
   "name": "shadow",
   "total": 25,
   "why": "Trials a new model on your own recent conversations, judged blind in both orders, then switches via llm_request middleware with undo; a genuinely new capability, but the acceptance checks skip off macOS, a catalog entry still holds a placeholder SHA, and an exclusive mode can block other sessions.",
   "scores": {
    "useful": 4,
    "reliable": 3,
    "polish": 3,
    "creative": 5,
    "beyond": 5
   }
  },
  {
   "rank": 13,
   "pr": 149640,
   "name": "hermes-ark",
   "total": 24,
   "why": "Encrypted, deduplicated backups of memory, SOUL, skills, config and sessions with item-level restore, S3 off-site, self-defense middleware against the agent and a Desktop page; a day-one install for many, but a backup rather than a new capability, and the 'never writes config.yaml' claim is overstated.",
   "scores": {
    "useful": 5,
    "reliable": 4,
    "polish": 4,
    "creative": 3,
    "beyond": 4
   }
  },
  {
   "rank": 14,
   "pr": 150556,
   "name": "hermes-ui-cards",
   "total": 24,
   "why": "Seventeen interactive answer cards rendered in Desktop from one JSON schema with parity tests, undo timers and a real e2e run; finished like a product, but it is a presentation layer rather than a new ability for the agent.",
   "scores": {
    "useful": 4,
    "reliable": 4,
    "polish": 5,
    "creative": 3,
    "beyond": 4
   }
  },
  {
   "rank": 15,
   "pr": 150840,
   "name": "kata",
   "total": 24,
   "why": "Mines repeated multi-step tool routines and, only after the user approves, compiles each into one deterministic tool with risk re-classification at run time; 58 tests passed here and nothing registers unapproved, though the plugin folder's README is five lines and the repo README echoes the judging criteria.",
   "scores": {
    "useful": 4,
    "reliable": 4,
    "polish": 4,
    "creative": 4,
    "beyond": 4
   }
  },
  {
   "rank": 16,
   "pr": 150353,
   "name": "fluent",
   "total": 24,
   "why": "Reads the installed programs' help and man pages and checks every terminal command before it runs, fixing invented flags and sending destructive subcommands to Hermes's approval; 454 tests passed here and the adopt step is tightly bounded, but there is no Desktop surface and only Linux was exercised.",
   "scores": {
    "useful": 4,
    "reliable": 4,
    "polish": 4,
    "creative": 4,
    "beyond": 4
   }
  },
  {
   "rank": 17,
   "pr": 151887,
   "name": "cmi",
   "total": 24,
   "why": "A deterministic zero-LLM context engine that folds old tool results into recallable digests while keeping a byte-stable prompt-cache prefix; 498 tests passed here, but the catalogued folder has no README, the repo README is about a Claude Code proxy, and CI runs standalone rather than on Hermes.",
   "scores": {
    "useful": 4,
    "reliable": 3,
    "polish": 3,
    "creative": 4,
    "beyond": 5
   }
  },
  {
   "rank": 18,
   "pr": 152651,
   "name": "workroom",
   "total": 24,
   "why": "Thirty-one tools, tripwire snapshots and approval-only edits that keep Word, PowerPoint and Excel decks in step with their source numbers; impressive and well tested on Windows, but the surface is huge, every tool call scans up to 200 MB, and macOS and Linux Desktop are unverified.",
   "scores": {
    "useful": 3,
    "reliable": 3,
    "polish": 4,
    "creative": 4,
    "beyond": 5
   }
  },
  {
   "rank": 19,
   "pr": 153020,
   "name": "breadcrumbs",
   "total": 24,
   "why": "A memory provider plus context engine with content-addressed provenance and a zoom level that opens any session, 669 tests passing here; original, but it starts a thread at import, leans on private Hermes imports, ships a Desktop bundle without its sources and reads like a research notebook.",
   "scores": {
    "useful": 3,
    "reliable": 3,
    "polish": 3,
    "creative": 5,
    "beyond": 5
   }
  },
  {
   "rank": 20,
   "pr": 153844,
   "name": "fine-print",
   "total": 23,
   "why": "Shows where Hermes silently truncates or swaps what the model sees, with per-turn cards and a limits sheet; fail-soft with 487 tests passing here, but the audience is narrow and it adds no tools, UI or new ability.",
   "scores": {
    "useful": 3,
    "reliable": 4,
    "polish": 4,
    "creative": 4,
    "beyond": 4
   }
  }
 ],
 "method": "Each pinned commit fetched (git fetch --depth 1), run through Nous's hermes plugins validate, enable and doctor via tryit.py inside a fresh Hermes checkout (all 20 passed), own test suites run with pytest where collectable, README, manifest, PR text and the working code read. Suites that could not be collected here (workroom, hypnos, hermes-ui-cards, housemate, shadow: missing yaml, lxml or aiohttp in the venv; hermes-ark and contract-necromancer: relative-import invocation) were judged by reading. No plugin was found to address an AI judge."
}