[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-digital-health-clinical-asr-build":3,"mdc-eoltpa-key":33,"related-org-nvidia-digital-health-clinical-asr-build":2511,"related-repo-nvidia-digital-health-clinical-asr-build":2672},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":22,"repoUrl":23,"updatedAt":24,"license":25,"forks":26,"topics":27,"repo":28,"sourceUrl":31,"mdContent":32},"digital-health-clinical-asr-build","curate clinical ASR manifests","Stage 2 of the Clinical ASR Flywheel. Use when curating clinical terms, tagging IPA, and synthesizing a NeMo manifest. NOT for scoring (use \u002Fdigital-health-clinical-asr-eval).",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19],{"name":13,"slug":14,"type":15},"Healthcare","healthcare","tag",{"name":17,"slug":18,"type":15},"Transcription","transcription",{"name":20,"slug":21,"type":15},"Speech","speech",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:28:18.39979","Apache-2.0",281,[],{"repoUrl":23,"stars":22,"forks":26,"topics":29,"description":30},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fdigital-health-clinical-asr-build","---\nname: \"digital-health-clinical-asr-build\"\ndescription: \"Stage 2 of the Clinical ASR Flywheel. Use when curating clinical terms, tagging IPA, and synthesizing a NeMo manifest. NOT for scoring (use \u002Fdigital-health-clinical-asr-eval).\"\nversion: \"1.1.0\"\nauthor: \"Ben Randoing \u003Cbrandoing@nvidia.com>\"\ntags:\n  - clinical-asr\n  - dataset\n  - ipa\n  - magpie\n  - nemo-manifest\n  - flywheel\ntools:\n  - Read\n  - Write\n  - Bash\n  - Skill\nlicense: Apache-2.0\ncompatibility: \"NVIDIA_API_KEY (required) for hosted Magpie TTS via NVCF. DICTIONARY_API_KEY (optional) for Merriam-Webster Medical Dictionary lookup. Stage 1 (\u002Fdigital-health-clinical-asr-setup) must have been completed first. All TTS, IPA, and synthesis recipes are inlined — no sibling agent skill required.\"\nmetadata:\n  author: \"Ben Randoing \u003Cbrandoing@nvidia.com>\"\n  tags:\n    - clinical-asr\n    - flywheel\n    - dataset\n    - ipa\n    - magpie\n  team: healthcare-tme\n  domain: ai-ml\n  stage: 2\n  previous_skill: digital-health-clinical-asr-setup\n  next_skill: digital-health-clinical-asr-eval\n---\n\n\u003C!--\nSPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.\nSPDX-License-Identifier: Apache-2.0\n-->\n\n# Clinical ASR Flywheel — Stage 2 (Build the benchmark)\n\n> **⚠ Agent: read this entire SKILL.md before answering.** This stage is conversational and gated. Specifically: ask the user 1–2 specialty-aware clarifying questions **before** proposing terms (Step 2a), walk them through the two-tier IPA pipeline (override → merriam-webster → magpie_g2p) in Step 2c, hit the explicit QA-mode audition gate in Step 2d before full Cartesian synthesis, and name **KER** as the headline metric they'll see in Stage 3. Skipping any of these defeats the methodology.\n\nYou are the **curate-and-synthesize** stage. The user arrives from `\u002Fdigital-health-clinical-asr-setup` and leaves with a NeMo-format `manifest.jsonl` plus the audio it references — both ready for scoring at `\u002Fdigital-health-clinical-asr-eval`.\n\nBe conversational. This is the warmest, most domain-aware step in the flywheel: you're asking a clinician (or someone who works with them) which terms hurt today and shaping a benchmark around their reality. Ask short, focused questions. Show the user what's being added. Don't lecture.\n\n## Data leaves your environment — disclose this to the user before any term is sent\n\nThis stage transmits user-curated content to two external services. Surface this to the user before invoking either call:\n\n| Service | What gets sent | When |\n|---|---|---|\n| **Merriam-Webster** (`dictionaryapi.com` API or `merriam-webster.com` public site) | One HTTP request per term in the seed list — term goes in URL path | Step 2c — see MW path bullets below |\n| **NVIDIA NVCF Magpie TTS** (`grpc.nvcf.nvidia.com`) | Each generated clinical sentence (text, plus any SSML IPA wrappers) | Steps 2d and 2e, every synthesis call |\n\nBoth endpoints expect **non-PHI synthetic content** — the term list you curate, the sentences `\u002Fdata-designer` (or your fallback templates) generates from it. **Do not pass real patient records, real ASR transcripts, or any PHI through this skill.** If the term list itself is sensitive (proprietary drug codenames, unreleased product names, customer-confidential indications), confirm with the user that external-API transmission is acceptable under their organization's data-governance policy before proceeding.\n\nIf no MW transmission is acceptable: take Path C below (skip MW; pipeline falls through to Magpie G2P with reduced coverage on long-tail terms).\n\n## Purpose\n\nCurate a clinical-specialty term list, generate eval audio for it through Magpie TTS with a two-tier IPA pipeline, and write a NeMo-format manifest tagged with the clinical-extension fields (`term`, `entity_category`, `ipa_source`, `voice_id`, `noise_level`, `context_type`). The output is the input to Stage 3.\n\nBy the end the user has:\n\n```\n$EVAL_DIR\u002Fcycle\u003CN>\u002F\n├── audio\u002F\u003Cslug>.wav        synthesized clips\n├── manifest.jsonl          NeMo format + clinical extension\n├── term_seed.csv           the curated input\n└── pronunciation_overrides.csv   appendable across cycles\n```\n\n(`$EVAL_DIR` is the user's own choice — this skill does not impose a layout. The structure above is a recommendation, not a requirement.)\n\n## When to use this skill\n\nActivate on user phrases like:\n\n- \"Build a clinical ASR benchmark\"\n- \"Curate drug names \u002F procedure names for ASR eval\"\n- \"Generate eval audio for medical terms\"\n- \"Create a NeMo manifest from clinical terms\"\n- \"Add oncology \u002F cardiology \u002F ortho terms to my benchmark\"\n- \"Audition the TTS pronunciation for these drug names\"\n- \"Make me a cycle-N manifest\"\n\nDo **not** activate when (also: if the message mentions `auth`, `API key`, `gRPC`, `streaming`, `riva-build`, `NIM deploy`, `NGC`, or `Docker`, route per the bullets below and stop):\n\n- The user already has a manifest and wants to score it → `\u002Fdigital-health-clinical-asr-eval`\n- The user wants to fine-tune on an existing manifest → `\u002Fdigital-health-clinical-asr-finetune`\n- The user is asking generic TTS \u002F SSML \u002F voice-cloning \u002F voice-catalog questions → `\u002Fread-aloud` (or `\u002Friva-tts`)\n- TTS\u002FASR **auth \u002F API keys \u002F gRPC \u002F streaming** → `\u002Friva-tts` or `\u002Friva-asr`\n- **NIM deploy** or `riva-build` \u002F `riva-deploy` flags → `\u002Friva-asr-custom` or `\u002Friva-tts-custom`\n- **NGC \u002F Docker \u002F NVIDIA Container Toolkit** → `\u002Friva-nim-setup`\n- The user is asking generic synthetic-data questions → `\u002Fdata-designer`\n\n## Prerequisites\n\n- **`\u002Fdigital-health-clinical-asr-setup` completed** — `NVIDIA_API_KEY` exported, Python deps installed, the six upstream skills confirmed.\n- **`\u002Fread-aloud`** (or `\u002Friva-tts`) reachable. Hosted Magpie via NVCF is the default. Self-hosted Magpie NIM works but adds `\u002Friva-nim-setup` to the prerequisite chain.\n- **`\u002Fdata-designer`** reachable. Template fallback is acceptable for a first cycle if `\u002Fdata-designer` is unavailable, but tag those rows so future cycles can re-generate.\n- **A working directory** the user owns. The skill recommends `$EVAL_DIR\u002Fcycle\u003CN>\u002F` but does not enforce it.\n\n## Instructions\n\n### 2a. Specialty interview → `term_seed.csv`\n\nAsk **one question at a time**. The goal is to surface 4–10 candidate terms with the right `entity_category`, not to write a textbook.\n\nQuestions, in order:\n\n1. *What specialty \u002F workflow is this for?* (oncology dictation, ICU handoff, psych intake, ortho post-op, …)\n2. *What ASR failure modes have you seen?* — drug names, multi-word procedures, abbreviations, compound conditions.\n3. *Which terms come up daily vs which are the hard ones?* — daily-common terms become the sanity baseline; daily-hard terms become the signal.\n\nPropose 4–10 candidate terms with `entity_category`. Confirm with the user before writing. Then write `term_seed.csv`:\n\n```csv\nterm,entity_category\ncefazolin,drug\nacetabular reamer,procedure\ntibial plateau,anatomy\nfemoroacetabular impingement,condition\nhemoglobin a1c,lab\nrespiratory therapist,role\n```\n\n**The category vocabulary is fixed.** KER keys off it. Allowed values:\n\n```\ndrug | procedure | anatomy | condition | lab | role\n```\n\nIf the user proposes a new category, push back: either it maps to one of the six, or the methodology needs a deliberate extension (which is a future cycle's job, not a one-off ad-hoc add).\n\n### 2b. Sentence generation via `\u002Fdata-designer`\n\nBrief `\u002Fdata-designer` with:\n\n> For each row in `term_seed.csv`, generate one or more natural English sentences embedding `term` in a way that fits the row's `entity_category`. Output schema: `{term, entity_category, sentence, context_type}`. Generate 3–5 `context_type` variants per term. Initial `context_type` vocabulary: `dictation`, `handoff`, `chart_note`, `history`. Sentence length 10–30 words.\n\nThe output of this step is a per-term sentence variants file. Any filename is fine — pick one and use it consistently across the cycle directory.\n\n**Template fallback.** If `\u002Fdata-designer` is unavailable, use a 4-template fallback (one per `context_type`) and substitute `term` mechanically. Tag those rows in the manifest (`context_type` is set, the sentence is just less natural) so a future cycle can regenerate.\n\n### 2c. Two-tier IPA tagging (the load-bearing quality lever)\n\nEvery term passes through a 3-tier pipeline, in order:\n\n1. **Override** — `pronunciation_overrides.csv` carries verified IPA the team has audited. If `term` matches a row here, the override wins.\n2. **Merriam-Webster** — for un-overridden terms, fetch the MW respelling, convert to IPA, validate against Magpie's en-US phoneme set. If both succeed, the term is tagged `merriam-webster`.\n3. **Magpie G2P (fall-through)** — if neither override nor MW produces a valid IPA, the plain text is passed to Magpie's neural G2P at synthesis time. The row is tagged `magpie_g2p`.\n\nEvery manifest row carries the `ipa_source` tag (`override | merriam-webster | magpie_g2p`). The delta between `merriam-webster` and `magpie_g2p` rows in the Stage 3 leaderboard **is the proof** the pronunciation strategy is working — call it out explicitly when you produce the leaderboard.\n\n**Three MW lookup choices** — all tag `merriam-webster`. **A**: `dictionaryapi.com` JSON API + `DICTIONARY_API_KEY` (free at dictionaryapi.com) — recommended for standalone use. **B**: HTML scrape of `merriam-webster.com` — no key, brittle to site HTML changes; recipe inlined in `references\u002Fpronunciation-pipeline.md`. **C**: skip MW, fall through to Magpie G2P with weaker long-tail coverage. Both recipes + the full respelling→IPA table live in `references\u002Fpronunciation-pipeline.md`. The Path A function takes `api_key` as an arg (never reads `os.environ`); pass `None` to skip MW.\n\n`pronunciation_overrides.csv` schema:\n\n```csv\nterm,ipa,verified_by,verified_at,notes\ncefazolin,sɛfəˈzoʊlɪn,brandoing,2026-05-13,confirmed against MW respelling + ear test\n```\n\nAppend-only across cycles. Re-running the build later picks up new entries automatically.\n\n### 2d. QA-mode synthesis (do **not** skip this gate)\n\nBefore running the full Cartesian product, synthesize **one wav per term** with: first voice, clean noise, default context. Audition each clip with the user.\n\nFor every term tagged `magpie_g2p`, propose an IPA candidate using clinical suffix patterns and validate against Magpie's en-US phoneme set **before** suggesting:\n\n| Suffix | Stress pattern (example) |\n|---|---|\n| `-mycin` | …ˈmaɪsɪn (vancomycin, gentamicin) |\n| `-prazole` | …ˈpreɪzoʊl (esomeprazole, omeprazole) |\n| `-statin` | …ˈstætɪn (atorvastatin, rosuvastatin) |\n| `-sartan` | …ˈsɑːrtən (losartan, valsartan) |\n| `-azole` | …ˈeɪzoʊl (fluconazole, ketoconazole) |\n| `-cillin` | …ˈsɪlɪn (amoxicillin, piperacillin) |\n| `-parin` | …ˈpɛərɪn (enoxaparin, heparin) |\n\n**Phoneme-validation pattern** — live-probe Magpie's en-US neural G2P with a candidate IPA. If Magpie accepts the SSML, the IPA is in its inventory. Use the suffix patterns above as a *pre-filter* (cheap heuristic) and the live probe to confirm before committing to an override. The `magpie_validates_ipa(ipa, api_key, voice_id)` recipe — a minimal NVCF gRPC synthesis call that returns `True`\u002F`False` fail-closed — is in `references\u002Fpronunciation-pipeline.md`.\n\nCall it once per candidate IPA before showing it to the user. On user approval, append the verified IPA to `pronunciation_overrides.csv`. The row's `ipa_source` flips from `magpie_g2p` to `override` on the next manifest generation.\n\n**HITL audition gate before Step 2e — fail-closed.** Do not synthesize the full Cartesian product, do not promote any staged IPA candidate to `pronunciation_overrides.csv`, and do not advance to Stage 3 until **one of the following has happened explicitly in conversation**:\n\n1. **The user confirms they have auditioned the QA clips** and reports their verdict per clip (or per bucket: \"the MW set sounds fine\", \"fix `pembrolizumab`\", etc.). Provide the `afplay` (macOS) or `paplay`\u002F`aplay` (Linux) commands so the user can play them — then **halt and wait for their reply after listening**. Paper-only approval via an AskUserQuestion prompt — clicking \"Promote all\" or \"Lock in\" without auditioning — **does not satisfy this gate**. Magpie-validating an IPA proves it's in the phoneme inventory; it does not prove it matches the *intended* pronunciation. Only the user's ears do that.\n2. **The user explicitly opts to skip audition for this cycle**, in deliberate language (e.g. *\"skip audition, accept the risk that mispronunciations may dilute the Stage 3 KER signal — log it as a cycle-N caveat\"*), not as a side-effect of a single click-through. Record the skip in a cycle-level note (e.g. `eval\u002Fcycle\u003CN>\u002Fcycle_notes.md`) so a future operator can see the audition was deferred.\n\nMagpie NVCF rate-limits aggressively on >100-row jobs, and a do-over costs both API credits and clock time — but the larger risk is shipping a manifest with mispronounced reference audio that quietly corrupts the Stage 3 KER signal. Time spent auditioning is cheaper than re-running the cycle.\n\n### 2e. Full benchmark generation\n\nAfter pronunciations are locked, generate the full Cartesian product `|terms| × |voices| × |noise_levels| × |context_types|`. Defaults: 2–4 Magpie en-US voices (Mia\u002FJason\u002FRay), `[clean, snr_15db, snr_5db]`, `[dictation, handoff, chart_note, history]`.\n\nSelf-contained synthesis — no `\u002Fread-aloud` required. The `synthesize_row(row, all_overrides, out_dir, api_key)` recipe — opens an NVCF gRPC stream, wraps overrides into SSML via `render_sentence_with_overrides`, writes 16-bit mono PCM to `\u003Cout_dir>\u002Faudio\u002F\u003Cslug>.wav` — is in `references\u002Fpronunciation-pipeline.md` (§Synthesis call). Key invariant: `all_overrides` carries *every* entry from `pronunciation_overrides.csv` (including context-word overrides like `intravenously`) so the renderer wraps any override whose verbatim text appears in `row['text']`. Wrapping only `row['term']` silently drops context-word overrides.\n\nNoise-injection (clean → `snr_15db` → `snr_5db`) and the manifest schema (NeMo canonical fields + clinical extension, plus pre-flight schema and audio-existence checks) all live in `references\u002Fmanifest-schema.md`.\n\n**Warn when product > 100 rows.** Magpie NVCF rate-limits with ~5–10% `RESOURCE_EXHAUSTED` drops on big runs. Re-run the dropped rows.\n\n### Stage 2 completion checklist\n\nDon't consider Stage 2 done until all five sub-steps ran. Agents commonly stop after 2a or 2b; the goal is a synthesized manifest plus a hand-off:\n\n- **2a** — `term_seed.csv`, 4–10 terms, `entity_category ∈ {drug, procedure, anatomy, condition, lab, role}`\n- **2b** — 3–5 `context_type` sentence variants per term\n- **2c** — every term tagged `ipa_source ∈ {override, merriam-webster, magpie_g2p}`\n- **2d** — QA wavs auditioned, IPA overrides locked with explicit user approval\n- **2e** — `manifest.jsonl` + per-row audio for the Cartesian product\n- **Hand-off** — name `\u002Fdigital-health-clinical-asr-eval` as the next skill and **KER** as its headline metric\n\nWrites go only into the user-chosen `$EVAL_DIR\u002Fcycle\u003CN>\u002F`. Don't write elsewhere, modify env, or install packages — those belong to `\u002Fdigital-health-clinical-asr-setup`.\n\n## Examples\n\n**Scenario A — fresh oncology benchmark.** User: *\"We're seeing chemo drug names mistranscribed. Where do I start?\"* → Step 2a: confirm specialty is oncology, ask about which drugs (immunotherapy biologics, platinum agents, taxanes). Propose ~10 candidates: `cisplatin`, `paclitaxel`, `pembrolizumab`, `nivolumab`, `carboplatin`, `docetaxel`, `bevacizumab`, `trastuzumab`, `cetuximab`, `pemetrexed`. Write `term_seed.csv` with all `entity_category=drug`. Step 2b: brief `\u002Fdata-designer` for 4 context variants each = 40 sentences. Step 2c: MW lookup for each — biologics like `pembrolizumab` will likely fall to `magpie_g2p`; platinum agents likely hit MW. Step 2d: synthesize one QA wav per term, walk the user through the `pembrolizumab` etc. clips, propose IPA candidates with `-mab` suffix stress patterns. Step 2e: on approval, run 10 terms × 2 voices × 2 noise levels × 3 contexts = 120 rows.\n\n**Scenario B — appending to an existing cycle.** User: *\"I have a cycle-1 manifest and I want to add 5 more procedures.\"* → Re-run only Steps 2a (specialty interview just for the new terms), 2b (sentence gen for the additions), 2c (IPA pipeline for the additions), 2d (audition the new terms), and 2e (synthesize only the new term rows). Append to the existing `manifest.jsonl`. **Do not regenerate audio for existing terms** — cycle isolation is intentional so leaderboards diff cycle N vs cycle N+1 cleanly.\n\n## Artifacts produced\n\n- `term_seed.csv` — curated terms with `entity_category`\n- `pronunciation_overrides.csv` — verified IPA, **appendable across cycles**\n- `manifest.jsonl` — NeMo format with clinical extension fields (one JSON object per line)\n- `audio\u002F\u003Cslug>.wav` — synthesized clips, one per manifest row\n\n## Troubleshooting\n\n- **TTS rate-limit drops (`RESOURCE_EXHAUSTED`)** on >100-row generation → expected on Magpie NVCF. Confirm exponential backoff is active in `\u002Fread-aloud`; expect ~5–10% drops on big runs and re-run for the gaps.\n- **All `ipa_source` rows tagged `magpie_g2p`** → MW lookup is failing across the board, or candidate IPAs are failing phoneme validation. Re-verify whichever MW path you configured (`DICTIONARY_API_KEY` for A; HTTPS reachability + parser for B), then check candidates against Magpie's en-US phoneme inventory.\n- **Magpie mispronounces a term even with the IPA override** → first verify the IPA is in the Magpie en-US phoneme inventory and the SSML wrapping is syntactically valid. If both check out, the underlying TTS bug is owned by `\u002Fread-aloud` (`\u002Friva-tts`) — route there for diagnosis. This skill provides the override mechanism but does not own the neural G2P or SSML parser.\n- **Sentence variants from `\u002Fdata-designer` are bland \u002F template-like** → check the brief; the schema-only prompt sometimes produces stereotyped output. Add 1–2 in-context examples to the brief and re-run.\n- **Audio files exist but `manifest.jsonl` is short** → manifest writer skipped rows whose synthesis returned a NVCF error. Re-run the build with only the missing rows.\n\nFor anything not in this list, identify which upstream skill is implicated and route there. The `digital-health-clinical-asr-build` skill owns the methodology, not the TTS or DataDesigner internals.\n\n## Limitations\n\n- **English-only by default.** Magpie's en-US phoneme inventory is what the two-tier IPA pipeline validates against. Other locales need a different upstream phoneme set + override CSV format.\n- **Six fixed entity categories.** Extending `entity_category` is a deliberate methodology change, not a one-off tweak — KER breakdowns, leaderboard sections, and downstream finetune scripts all key off the vocabulary.\n- **Tiny first cycles.** Below ~20 terms, the by-`ipa_source` leaderboard split won't have enough rows in each bucket to be statistically meaningful. Build a meaningful cycle even if it costs a session.\n- **Magpie NVCF rate-limits.** ~5–10% drops on large jobs; budget a re-run pass.\n\n## Next steps\n\n- **Forward:** `\u002Fdigital-health-clinical-asr-eval` — transcribe the manifest, score WER\u002FCER\u002FKER\u002FSER, produce the five-section leaderboard.\n- **Back to setup** (if anything in the env is broken): `\u002Fdigital-health-clinical-asr-setup`.\n- **Lateral** for TTS-specific debugging: `\u002Fread-aloud` or `\u002Friva-tts`.\n\n## References\n\n- [`references\u002Fmanifest-schema.md`](references\u002Fmanifest-schema.md) — NeMo canonical fields + clinical extension; pre-flight schema and audio-existence checks; cross-cycle stability rules\n\n\n",{"data":34,"body":57},{"name":4,"description":6,"version":35,"author":36,"tags":37,"tools":44,"license":25,"compatibility":49,"metadata":50},"1.1.0","Ben Randoing \u003Cbrandoing@nvidia.com>",[38,39,40,41,42,43],"clinical-asr","dataset","ipa","magpie","nemo-manifest","flywheel",[45,46,47,48],"Read","Write","Bash","Skill","NVIDIA_API_KEY (required) for hosted Magpie TTS via NVCF. DICTIONARY_API_KEY (optional) for Merriam-Webster Medical Dictionary lookup. Stage 1 (\u002Fdigital-health-clinical-asr-setup) must have been completed first. All TTS, IPA, and synthesis recipes are inlined — no sibling agent skill required.",{"author":36,"tags":51,"team":52,"domain":53,"stage":54,"previous_skill":55,"next_skill":56},[38,43,39,40,41],"healthcare-tme","ai-ml",2,"digital-health-clinical-asr-setup","digital-health-clinical-asr-eval",{"type":58,"children":59},"root",[60,69,99,136,141,148,153,255,282,287,293,342,347,359,372,378,383,423,493,623,629,719,725,738,757,762,797,816,889,899,908,913,924,936,1016,1021,1059,1065,1070,1130,1171,1269,1279,1302,1307,1319,1331,1349,1490,1537,1571,1594,1684,1689,1695,1722,1811,1838,1856,1862,1867,1974,1992,1998,2136,2165,2171,2225,2231,2343,2355,2361,2418,2424,2482,2488,2505],{"type":61,"tag":62,"props":63,"children":65},"element","h1",{"id":64},"clinical-asr-flywheel-stage-2-build-the-benchmark",[66],{"type":67,"value":68},"text","Clinical ASR Flywheel — Stage 2 (Build the benchmark)",{"type":61,"tag":70,"props":71,"children":72},"blockquote",{},[73],{"type":61,"tag":74,"props":75,"children":76},"p",{},[77,83,85,90,92,97],{"type":61,"tag":78,"props":79,"children":80},"strong",{},[81],{"type":67,"value":82},"⚠ Agent: read this entire SKILL.md before answering.",{"type":67,"value":84}," This stage is conversational and gated. Specifically: ask the user 1–2 specialty-aware clarifying questions ",{"type":61,"tag":78,"props":86,"children":87},{},[88],{"type":67,"value":89},"before",{"type":67,"value":91}," proposing terms (Step 2a), walk them through the two-tier IPA pipeline (override → merriam-webster → magpie_g2p) in Step 2c, hit the explicit QA-mode audition gate in Step 2d before full Cartesian synthesis, and name ",{"type":61,"tag":78,"props":93,"children":94},{},[95],{"type":67,"value":96},"KER",{"type":67,"value":98}," as the headline metric they'll see in Stage 3. Skipping any of these defeats the methodology.",{"type":61,"tag":74,"props":100,"children":101},{},[102,104,109,111,118,120,126,128,134],{"type":67,"value":103},"You are the ",{"type":61,"tag":78,"props":105,"children":106},{},[107],{"type":67,"value":108},"curate-and-synthesize",{"type":67,"value":110}," stage. The user arrives from ",{"type":61,"tag":112,"props":113,"children":115},"code",{"className":114},[],[116],{"type":67,"value":117},"\u002Fdigital-health-clinical-asr-setup",{"type":67,"value":119}," and leaves with a NeMo-format ",{"type":61,"tag":112,"props":121,"children":123},{"className":122},[],[124],{"type":67,"value":125},"manifest.jsonl",{"type":67,"value":127}," plus the audio it references — both ready for scoring at ",{"type":61,"tag":112,"props":129,"children":131},{"className":130},[],[132],{"type":67,"value":133},"\u002Fdigital-health-clinical-asr-eval",{"type":67,"value":135},".",{"type":61,"tag":74,"props":137,"children":138},{},[139],{"type":67,"value":140},"Be conversational. This is the warmest, most domain-aware step in the flywheel: you're asking a clinician (or someone who works with them) which terms hurt today and shaping a benchmark around their reality. Ask short, focused questions. Show the user what's being added. Don't lecture.",{"type":61,"tag":142,"props":143,"children":145},"h2",{"id":144},"data-leaves-your-environment-disclose-this-to-the-user-before-any-term-is-sent",[146],{"type":67,"value":147},"Data leaves your environment — disclose this to the user before any term is sent",{"type":61,"tag":74,"props":149,"children":150},{},[151],{"type":67,"value":152},"This stage transmits user-curated content to two external services. Surface this to the user before invoking either call:",{"type":61,"tag":154,"props":155,"children":156},"table",{},[157,181],{"type":61,"tag":158,"props":159,"children":160},"thead",{},[161],{"type":61,"tag":162,"props":163,"children":164},"tr",{},[165,171,176],{"type":61,"tag":166,"props":167,"children":168},"th",{},[169],{"type":67,"value":170},"Service",{"type":61,"tag":166,"props":172,"children":173},{},[174],{"type":67,"value":175},"What gets sent",{"type":61,"tag":166,"props":177,"children":178},{},[179],{"type":67,"value":180},"When",{"type":61,"tag":182,"props":183,"children":184},"tbody",{},[185,225],{"type":61,"tag":162,"props":186,"children":187},{},[188,215,220],{"type":61,"tag":189,"props":190,"children":191},"td",{},[192,197,199,205,207,213],{"type":61,"tag":78,"props":193,"children":194},{},[195],{"type":67,"value":196},"Merriam-Webster",{"type":67,"value":198}," (",{"type":61,"tag":112,"props":200,"children":202},{"className":201},[],[203],{"type":67,"value":204},"dictionaryapi.com",{"type":67,"value":206}," API or ",{"type":61,"tag":112,"props":208,"children":210},{"className":209},[],[211],{"type":67,"value":212},"merriam-webster.com",{"type":67,"value":214}," public site)",{"type":61,"tag":189,"props":216,"children":217},{},[218],{"type":67,"value":219},"One HTTP request per term in the seed list — term goes in URL path",{"type":61,"tag":189,"props":221,"children":222},{},[223],{"type":67,"value":224},"Step 2c — see MW path bullets below",{"type":61,"tag":162,"props":226,"children":227},{},[228,245,250],{"type":61,"tag":189,"props":229,"children":230},{},[231,236,237,243],{"type":61,"tag":78,"props":232,"children":233},{},[234],{"type":67,"value":235},"NVIDIA NVCF Magpie TTS",{"type":67,"value":198},{"type":61,"tag":112,"props":238,"children":240},{"className":239},[],[241],{"type":67,"value":242},"grpc.nvcf.nvidia.com",{"type":67,"value":244},")",{"type":61,"tag":189,"props":246,"children":247},{},[248],{"type":67,"value":249},"Each generated clinical sentence (text, plus any SSML IPA wrappers)",{"type":61,"tag":189,"props":251,"children":252},{},[253],{"type":67,"value":254},"Steps 2d and 2e, every synthesis call",{"type":61,"tag":74,"props":256,"children":257},{},[258,260,265,267,273,275,280],{"type":67,"value":259},"Both endpoints expect ",{"type":61,"tag":78,"props":261,"children":262},{},[263],{"type":67,"value":264},"non-PHI synthetic content",{"type":67,"value":266}," — the term list you curate, the sentences ",{"type":61,"tag":112,"props":268,"children":270},{"className":269},[],[271],{"type":67,"value":272},"\u002Fdata-designer",{"type":67,"value":274}," (or your fallback templates) generates from it. ",{"type":61,"tag":78,"props":276,"children":277},{},[278],{"type":67,"value":279},"Do not pass real patient records, real ASR transcripts, or any PHI through this skill.",{"type":67,"value":281}," If the term list itself is sensitive (proprietary drug codenames, unreleased product names, customer-confidential indications), confirm with the user that external-API transmission is acceptable under their organization's data-governance policy before proceeding.",{"type":61,"tag":74,"props":283,"children":284},{},[285],{"type":67,"value":286},"If no MW transmission is acceptable: take Path C below (skip MW; pipeline falls through to Magpie G2P with reduced coverage on long-tail terms).",{"type":61,"tag":142,"props":288,"children":290},{"id":289},"purpose",[291],{"type":67,"value":292},"Purpose",{"type":61,"tag":74,"props":294,"children":295},{},[296,298,304,306,312,313,319,320,326,327,333,334,340],{"type":67,"value":297},"Curate a clinical-specialty term list, generate eval audio for it through Magpie TTS with a two-tier IPA pipeline, and write a NeMo-format manifest tagged with the clinical-extension fields (",{"type":61,"tag":112,"props":299,"children":301},{"className":300},[],[302],{"type":67,"value":303},"term",{"type":67,"value":305},", ",{"type":61,"tag":112,"props":307,"children":309},{"className":308},[],[310],{"type":67,"value":311},"entity_category",{"type":67,"value":305},{"type":61,"tag":112,"props":314,"children":316},{"className":315},[],[317],{"type":67,"value":318},"ipa_source",{"type":67,"value":305},{"type":61,"tag":112,"props":321,"children":323},{"className":322},[],[324],{"type":67,"value":325},"voice_id",{"type":67,"value":305},{"type":61,"tag":112,"props":328,"children":330},{"className":329},[],[331],{"type":67,"value":332},"noise_level",{"type":67,"value":305},{"type":61,"tag":112,"props":335,"children":337},{"className":336},[],[338],{"type":67,"value":339},"context_type",{"type":67,"value":341},"). The output is the input to Stage 3.",{"type":61,"tag":74,"props":343,"children":344},{},[345],{"type":67,"value":346},"By the end the user has:",{"type":61,"tag":348,"props":349,"children":353},"pre",{"className":350,"code":352,"language":67},[351],"language-text","$EVAL_DIR\u002Fcycle\u003CN>\u002F\n├── audio\u002F\u003Cslug>.wav        synthesized clips\n├── manifest.jsonl          NeMo format + clinical extension\n├── term_seed.csv           the curated input\n└── pronunciation_overrides.csv   appendable across cycles\n",[354],{"type":61,"tag":112,"props":355,"children":357},{"__ignoreMap":356},"",[358],{"type":67,"value":352},{"type":61,"tag":74,"props":360,"children":361},{},[362,364,370],{"type":67,"value":363},"(",{"type":61,"tag":112,"props":365,"children":367},{"className":366},[],[368],{"type":67,"value":369},"$EVAL_DIR",{"type":67,"value":371}," is the user's own choice — this skill does not impose a layout. The structure above is a recommendation, not a requirement.)",{"type":61,"tag":142,"props":373,"children":375},{"id":374},"when-to-use-this-skill",[376],{"type":67,"value":377},"When to use this skill",{"type":61,"tag":74,"props":379,"children":380},{},[381],{"type":67,"value":382},"Activate on user phrases like:",{"type":61,"tag":384,"props":385,"children":386},"ul",{},[387,393,398,403,408,413,418],{"type":61,"tag":388,"props":389,"children":390},"li",{},[391],{"type":67,"value":392},"\"Build a clinical ASR benchmark\"",{"type":61,"tag":388,"props":394,"children":395},{},[396],{"type":67,"value":397},"\"Curate drug names \u002F procedure names for ASR eval\"",{"type":61,"tag":388,"props":399,"children":400},{},[401],{"type":67,"value":402},"\"Generate eval audio for medical terms\"",{"type":61,"tag":388,"props":404,"children":405},{},[406],{"type":67,"value":407},"\"Create a NeMo manifest from clinical terms\"",{"type":61,"tag":388,"props":409,"children":410},{},[411],{"type":67,"value":412},"\"Add oncology \u002F cardiology \u002F ortho terms to my benchmark\"",{"type":61,"tag":388,"props":414,"children":415},{},[416],{"type":67,"value":417},"\"Audition the TTS pronunciation for these drug names\"",{"type":61,"tag":388,"props":419,"children":420},{},[421],{"type":67,"value":422},"\"Make me a cycle-N manifest\"",{"type":61,"tag":74,"props":424,"children":425},{},[426,428,433,435,441,442,448,449,455,456,462,463,469,470,476,477,483,485,491],{"type":67,"value":427},"Do ",{"type":61,"tag":78,"props":429,"children":430},{},[431],{"type":67,"value":432},"not",{"type":67,"value":434}," activate when (also: if the message mentions ",{"type":61,"tag":112,"props":436,"children":438},{"className":437},[],[439],{"type":67,"value":440},"auth",{"type":67,"value":305},{"type":61,"tag":112,"props":443,"children":445},{"className":444},[],[446],{"type":67,"value":447},"API key",{"type":67,"value":305},{"type":61,"tag":112,"props":450,"children":452},{"className":451},[],[453],{"type":67,"value":454},"gRPC",{"type":67,"value":305},{"type":61,"tag":112,"props":457,"children":459},{"className":458},[],[460],{"type":67,"value":461},"streaming",{"type":67,"value":305},{"type":61,"tag":112,"props":464,"children":466},{"className":465},[],[467],{"type":67,"value":468},"riva-build",{"type":67,"value":305},{"type":61,"tag":112,"props":471,"children":473},{"className":472},[],[474],{"type":67,"value":475},"NIM deploy",{"type":67,"value":305},{"type":61,"tag":112,"props":478,"children":480},{"className":479},[],[481],{"type":67,"value":482},"NGC",{"type":67,"value":484},", or ",{"type":61,"tag":112,"props":486,"children":488},{"className":487},[],[489],{"type":67,"value":490},"Docker",{"type":67,"value":492},", route per the bullets below and stop):",{"type":61,"tag":384,"props":494,"children":495},{},[496,506,517,537,562,598,613],{"type":61,"tag":388,"props":497,"children":498},{},[499,501],{"type":67,"value":500},"The user already has a manifest and wants to score it → ",{"type":61,"tag":112,"props":502,"children":504},{"className":503},[],[505],{"type":67,"value":133},{"type":61,"tag":388,"props":507,"children":508},{},[509,511],{"type":67,"value":510},"The user wants to fine-tune on an existing manifest → ",{"type":61,"tag":112,"props":512,"children":514},{"className":513},[],[515],{"type":67,"value":516},"\u002Fdigital-health-clinical-asr-finetune",{"type":61,"tag":388,"props":518,"children":519},{},[520,522,528,530,536],{"type":67,"value":521},"The user is asking generic TTS \u002F SSML \u002F voice-cloning \u002F voice-catalog questions → ",{"type":61,"tag":112,"props":523,"children":525},{"className":524},[],[526],{"type":67,"value":527},"\u002Fread-aloud",{"type":67,"value":529}," (or ",{"type":61,"tag":112,"props":531,"children":533},{"className":532},[],[534],{"type":67,"value":535},"\u002Friva-tts",{"type":67,"value":244},{"type":61,"tag":388,"props":538,"children":539},{},[540,542,547,549,554,556],{"type":67,"value":541},"TTS\u002FASR ",{"type":61,"tag":78,"props":543,"children":544},{},[545],{"type":67,"value":546},"auth \u002F API keys \u002F gRPC \u002F streaming",{"type":67,"value":548}," → ",{"type":61,"tag":112,"props":550,"children":552},{"className":551},[],[553],{"type":67,"value":535},{"type":67,"value":555}," or ",{"type":61,"tag":112,"props":557,"children":559},{"className":558},[],[560],{"type":67,"value":561},"\u002Friva-asr",{"type":61,"tag":388,"props":563,"children":564},{},[565,569,570,575,577,583,585,591,592],{"type":61,"tag":78,"props":566,"children":567},{},[568],{"type":67,"value":475},{"type":67,"value":555},{"type":61,"tag":112,"props":571,"children":573},{"className":572},[],[574],{"type":67,"value":468},{"type":67,"value":576}," \u002F ",{"type":61,"tag":112,"props":578,"children":580},{"className":579},[],[581],{"type":67,"value":582},"riva-deploy",{"type":67,"value":584}," flags → ",{"type":61,"tag":112,"props":586,"children":588},{"className":587},[],[589],{"type":67,"value":590},"\u002Friva-asr-custom",{"type":67,"value":555},{"type":61,"tag":112,"props":593,"children":595},{"className":594},[],[596],{"type":67,"value":597},"\u002Friva-tts-custom",{"type":61,"tag":388,"props":599,"children":600},{},[601,606,607],{"type":61,"tag":78,"props":602,"children":603},{},[604],{"type":67,"value":605},"NGC \u002F Docker \u002F NVIDIA Container Toolkit",{"type":67,"value":548},{"type":61,"tag":112,"props":608,"children":610},{"className":609},[],[611],{"type":67,"value":612},"\u002Friva-nim-setup",{"type":61,"tag":388,"props":614,"children":615},{},[616,618],{"type":67,"value":617},"The user is asking generic synthetic-data questions → ",{"type":61,"tag":112,"props":619,"children":621},{"className":620},[],[622],{"type":67,"value":272},{"type":61,"tag":142,"props":624,"children":626},{"id":625},"prerequisites",[627],{"type":67,"value":628},"Prerequisites",{"type":61,"tag":384,"props":630,"children":631},{},[632,655,681,701],{"type":61,"tag":388,"props":633,"children":634},{},[635,645,647,653],{"type":61,"tag":78,"props":636,"children":637},{},[638,643],{"type":61,"tag":112,"props":639,"children":641},{"className":640},[],[642],{"type":67,"value":117},{"type":67,"value":644}," completed",{"type":67,"value":646}," — ",{"type":61,"tag":112,"props":648,"children":650},{"className":649},[],[651],{"type":67,"value":652},"NVIDIA_API_KEY",{"type":67,"value":654}," exported, Python deps installed, the six upstream skills confirmed.",{"type":61,"tag":388,"props":656,"children":657},{},[658,666,667,672,674,679],{"type":61,"tag":78,"props":659,"children":660},{},[661],{"type":61,"tag":112,"props":662,"children":664},{"className":663},[],[665],{"type":67,"value":527},{"type":67,"value":529},{"type":61,"tag":112,"props":668,"children":670},{"className":669},[],[671],{"type":67,"value":535},{"type":67,"value":673},") reachable. Hosted Magpie via NVCF is the default. Self-hosted Magpie NIM works but adds ",{"type":61,"tag":112,"props":675,"children":677},{"className":676},[],[678],{"type":67,"value":612},{"type":67,"value":680}," to the prerequisite chain.",{"type":61,"tag":388,"props":682,"children":683},{},[684,692,694,699],{"type":61,"tag":78,"props":685,"children":686},{},[687],{"type":61,"tag":112,"props":688,"children":690},{"className":689},[],[691],{"type":67,"value":272},{"type":67,"value":693}," reachable. Template fallback is acceptable for a first cycle if ",{"type":61,"tag":112,"props":695,"children":697},{"className":696},[],[698],{"type":67,"value":272},{"type":67,"value":700}," is unavailable, but tag those rows so future cycles can re-generate.",{"type":61,"tag":388,"props":702,"children":703},{},[704,709,711,717],{"type":61,"tag":78,"props":705,"children":706},{},[707],{"type":67,"value":708},"A working directory",{"type":67,"value":710}," the user owns. The skill recommends ",{"type":61,"tag":112,"props":712,"children":714},{"className":713},[],[715],{"type":67,"value":716},"$EVAL_DIR\u002Fcycle\u003CN>\u002F",{"type":67,"value":718}," but does not enforce it.",{"type":61,"tag":142,"props":720,"children":722},{"id":721},"instructions",[723],{"type":67,"value":724},"Instructions",{"type":61,"tag":726,"props":727,"children":729},"h3",{"id":728},"_2a-specialty-interview-term_seedcsv",[730,732],{"type":67,"value":731},"2a. Specialty interview → ",{"type":61,"tag":112,"props":733,"children":735},{"className":734},[],[736],{"type":67,"value":737},"term_seed.csv",{"type":61,"tag":74,"props":739,"children":740},{},[741,743,748,750,755],{"type":67,"value":742},"Ask ",{"type":61,"tag":78,"props":744,"children":745},{},[746],{"type":67,"value":747},"one question at a time",{"type":67,"value":749},". The goal is to surface 4–10 candidate terms with the right ",{"type":61,"tag":112,"props":751,"children":753},{"className":752},[],[754],{"type":67,"value":311},{"type":67,"value":756},", not to write a textbook.",{"type":61,"tag":74,"props":758,"children":759},{},[760],{"type":67,"value":761},"Questions, in order:",{"type":61,"tag":763,"props":764,"children":765},"ol",{},[766,777,787],{"type":61,"tag":388,"props":767,"children":768},{},[769,775],{"type":61,"tag":770,"props":771,"children":772},"em",{},[773],{"type":67,"value":774},"What specialty \u002F workflow is this for?",{"type":67,"value":776}," (oncology dictation, ICU handoff, psych intake, ortho post-op, …)",{"type":61,"tag":388,"props":778,"children":779},{},[780,785],{"type":61,"tag":770,"props":781,"children":782},{},[783],{"type":67,"value":784},"What ASR failure modes have you seen?",{"type":67,"value":786}," — drug names, multi-word procedures, abbreviations, compound conditions.",{"type":61,"tag":388,"props":788,"children":789},{},[790,795],{"type":61,"tag":770,"props":791,"children":792},{},[793],{"type":67,"value":794},"Which terms come up daily vs which are the hard ones?",{"type":67,"value":796}," — daily-common terms become the sanity baseline; daily-hard terms become the signal.",{"type":61,"tag":74,"props":798,"children":799},{},[800,802,807,809,814],{"type":67,"value":801},"Propose 4–10 candidate terms with ",{"type":61,"tag":112,"props":803,"children":805},{"className":804},[],[806],{"type":67,"value":311},{"type":67,"value":808},". Confirm with the user before writing. Then write ",{"type":61,"tag":112,"props":810,"children":812},{"className":811},[],[813],{"type":67,"value":737},{"type":67,"value":815},":",{"type":61,"tag":348,"props":817,"children":821},{"className":818,"code":819,"language":820,"meta":356,"style":356},"language-csv shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","term,entity_category\ncefazolin,drug\nacetabular reamer,procedure\ntibial plateau,anatomy\nfemoroacetabular impingement,condition\nhemoglobin a1c,lab\nrespiratory therapist,role\n","csv",[822],{"type":61,"tag":112,"props":823,"children":824},{"__ignoreMap":356},[825,836,844,853,862,871,880],{"type":61,"tag":826,"props":827,"children":830},"span",{"class":828,"line":829},"line",1,[831],{"type":61,"tag":826,"props":832,"children":833},{},[834],{"type":67,"value":835},"term,entity_category\n",{"type":61,"tag":826,"props":837,"children":838},{"class":828,"line":54},[839],{"type":61,"tag":826,"props":840,"children":841},{},[842],{"type":67,"value":843},"cefazolin,drug\n",{"type":61,"tag":826,"props":845,"children":847},{"class":828,"line":846},3,[848],{"type":61,"tag":826,"props":849,"children":850},{},[851],{"type":67,"value":852},"acetabular reamer,procedure\n",{"type":61,"tag":826,"props":854,"children":856},{"class":828,"line":855},4,[857],{"type":61,"tag":826,"props":858,"children":859},{},[860],{"type":67,"value":861},"tibial plateau,anatomy\n",{"type":61,"tag":826,"props":863,"children":865},{"class":828,"line":864},5,[866],{"type":61,"tag":826,"props":867,"children":868},{},[869],{"type":67,"value":870},"femoroacetabular impingement,condition\n",{"type":61,"tag":826,"props":872,"children":874},{"class":828,"line":873},6,[875],{"type":61,"tag":826,"props":876,"children":877},{},[878],{"type":67,"value":879},"hemoglobin a1c,lab\n",{"type":61,"tag":826,"props":881,"children":883},{"class":828,"line":882},7,[884],{"type":61,"tag":826,"props":885,"children":886},{},[887],{"type":67,"value":888},"respiratory therapist,role\n",{"type":61,"tag":74,"props":890,"children":891},{},[892,897],{"type":61,"tag":78,"props":893,"children":894},{},[895],{"type":67,"value":896},"The category vocabulary is fixed.",{"type":67,"value":898}," KER keys off it. Allowed values:",{"type":61,"tag":348,"props":900,"children":903},{"className":901,"code":902,"language":67},[351],"drug | procedure | anatomy | condition | lab | role\n",[904],{"type":61,"tag":112,"props":905,"children":906},{"__ignoreMap":356},[907],{"type":67,"value":902},{"type":61,"tag":74,"props":909,"children":910},{},[911],{"type":67,"value":912},"If the user proposes a new category, push back: either it maps to one of the six, or the methodology needs a deliberate extension (which is a future cycle's job, not a one-off ad-hoc add).",{"type":61,"tag":726,"props":914,"children":916},{"id":915},"_2b-sentence-generation-via-data-designer",[917,919],{"type":67,"value":918},"2b. Sentence generation via ",{"type":61,"tag":112,"props":920,"children":922},{"className":921},[],[923],{"type":67,"value":272},{"type":61,"tag":74,"props":925,"children":926},{},[927,929,934],{"type":67,"value":928},"Brief ",{"type":61,"tag":112,"props":930,"children":932},{"className":931},[],[933],{"type":67,"value":272},{"type":67,"value":935}," with:",{"type":61,"tag":70,"props":937,"children":938},{},[939],{"type":61,"tag":74,"props":940,"children":941},{},[942,944,949,951,956,958,963,965,971,973,978,980,985,987,993,994,1000,1001,1007,1008,1014],{"type":67,"value":943},"For each row in ",{"type":61,"tag":112,"props":945,"children":947},{"className":946},[],[948],{"type":67,"value":737},{"type":67,"value":950},", generate one or more natural English sentences embedding ",{"type":61,"tag":112,"props":952,"children":954},{"className":953},[],[955],{"type":67,"value":303},{"type":67,"value":957}," in a way that fits the row's ",{"type":61,"tag":112,"props":959,"children":961},{"className":960},[],[962],{"type":67,"value":311},{"type":67,"value":964},". Output schema: ",{"type":61,"tag":112,"props":966,"children":968},{"className":967},[],[969],{"type":67,"value":970},"{term, entity_category, sentence, context_type}",{"type":67,"value":972},". Generate 3–5 ",{"type":61,"tag":112,"props":974,"children":976},{"className":975},[],[977],{"type":67,"value":339},{"type":67,"value":979}," variants per term. Initial ",{"type":61,"tag":112,"props":981,"children":983},{"className":982},[],[984],{"type":67,"value":339},{"type":67,"value":986}," vocabulary: ",{"type":61,"tag":112,"props":988,"children":990},{"className":989},[],[991],{"type":67,"value":992},"dictation",{"type":67,"value":305},{"type":61,"tag":112,"props":995,"children":997},{"className":996},[],[998],{"type":67,"value":999},"handoff",{"type":67,"value":305},{"type":61,"tag":112,"props":1002,"children":1004},{"className":1003},[],[1005],{"type":67,"value":1006},"chart_note",{"type":67,"value":305},{"type":61,"tag":112,"props":1009,"children":1011},{"className":1010},[],[1012],{"type":67,"value":1013},"history",{"type":67,"value":1015},". Sentence length 10–30 words.",{"type":61,"tag":74,"props":1017,"children":1018},{},[1019],{"type":67,"value":1020},"The output of this step is a per-term sentence variants file. Any filename is fine — pick one and use it consistently across the cycle directory.",{"type":61,"tag":74,"props":1022,"children":1023},{},[1024,1029,1031,1036,1038,1043,1045,1050,1052,1057],{"type":61,"tag":78,"props":1025,"children":1026},{},[1027],{"type":67,"value":1028},"Template fallback.",{"type":67,"value":1030}," If ",{"type":61,"tag":112,"props":1032,"children":1034},{"className":1033},[],[1035],{"type":67,"value":272},{"type":67,"value":1037}," is unavailable, use a 4-template fallback (one per ",{"type":61,"tag":112,"props":1039,"children":1041},{"className":1040},[],[1042],{"type":67,"value":339},{"type":67,"value":1044},") and substitute ",{"type":61,"tag":112,"props":1046,"children":1048},{"className":1047},[],[1049],{"type":67,"value":303},{"type":67,"value":1051}," mechanically. Tag those rows in the manifest (",{"type":61,"tag":112,"props":1053,"children":1055},{"className":1054},[],[1056],{"type":67,"value":339},{"type":67,"value":1058}," is set, the sentence is just less natural) so a future cycle can regenerate.",{"type":61,"tag":726,"props":1060,"children":1062},{"id":1061},"_2c-two-tier-ipa-tagging-the-load-bearing-quality-lever",[1063],{"type":67,"value":1064},"2c. Two-tier IPA tagging (the load-bearing quality lever)",{"type":61,"tag":74,"props":1066,"children":1067},{},[1068],{"type":67,"value":1069},"Every term passes through a 3-tier pipeline, in order:",{"type":61,"tag":763,"props":1071,"children":1072},{},[1073,1097,1113],{"type":61,"tag":388,"props":1074,"children":1075},{},[1076,1081,1082,1088,1090,1095],{"type":61,"tag":78,"props":1077,"children":1078},{},[1079],{"type":67,"value":1080},"Override",{"type":67,"value":646},{"type":61,"tag":112,"props":1083,"children":1085},{"className":1084},[],[1086],{"type":67,"value":1087},"pronunciation_overrides.csv",{"type":67,"value":1089}," carries verified IPA the team has audited. If ",{"type":61,"tag":112,"props":1091,"children":1093},{"className":1092},[],[1094],{"type":67,"value":303},{"type":67,"value":1096}," matches a row here, the override wins.",{"type":61,"tag":388,"props":1098,"children":1099},{},[1100,1104,1106,1112],{"type":61,"tag":78,"props":1101,"children":1102},{},[1103],{"type":67,"value":196},{"type":67,"value":1105}," — for un-overridden terms, fetch the MW respelling, convert to IPA, validate against Magpie's en-US phoneme set. If both succeed, the term is tagged ",{"type":61,"tag":112,"props":1107,"children":1109},{"className":1108},[],[1110],{"type":67,"value":1111},"merriam-webster",{"type":67,"value":135},{"type":61,"tag":388,"props":1114,"children":1115},{},[1116,1121,1123,1129],{"type":61,"tag":78,"props":1117,"children":1118},{},[1119],{"type":67,"value":1120},"Magpie G2P (fall-through)",{"type":67,"value":1122}," — if neither override nor MW produces a valid IPA, the plain text is passed to Magpie's neural G2P at synthesis time. The row is tagged ",{"type":61,"tag":112,"props":1124,"children":1126},{"className":1125},[],[1127],{"type":67,"value":1128},"magpie_g2p",{"type":67,"value":135},{"type":61,"tag":74,"props":1131,"children":1132},{},[1133,1135,1140,1142,1148,1150,1155,1157,1162,1164,1169],{"type":67,"value":1134},"Every manifest row carries the ",{"type":61,"tag":112,"props":1136,"children":1138},{"className":1137},[],[1139],{"type":67,"value":318},{"type":67,"value":1141}," tag (",{"type":61,"tag":112,"props":1143,"children":1145},{"className":1144},[],[1146],{"type":67,"value":1147},"override | merriam-webster | magpie_g2p",{"type":67,"value":1149},"). The delta between ",{"type":61,"tag":112,"props":1151,"children":1153},{"className":1152},[],[1154],{"type":67,"value":1111},{"type":67,"value":1156}," and ",{"type":61,"tag":112,"props":1158,"children":1160},{"className":1159},[],[1161],{"type":67,"value":1128},{"type":67,"value":1163}," rows in the Stage 3 leaderboard ",{"type":61,"tag":78,"props":1165,"children":1166},{},[1167],{"type":67,"value":1168},"is the proof",{"type":67,"value":1170}," the pronunciation strategy is working — call it out explicitly when you produce the leaderboard.",{"type":61,"tag":74,"props":1172,"children":1173},{},[1174,1179,1181,1186,1188,1193,1195,1200,1202,1208,1210,1215,1217,1222,1224,1230,1231,1236,1238,1243,1245,1251,1253,1259,1261,1267],{"type":61,"tag":78,"props":1175,"children":1176},{},[1177],{"type":67,"value":1178},"Three MW lookup choices",{"type":67,"value":1180}," — all tag ",{"type":61,"tag":112,"props":1182,"children":1184},{"className":1183},[],[1185],{"type":67,"value":1111},{"type":67,"value":1187},". ",{"type":61,"tag":78,"props":1189,"children":1190},{},[1191],{"type":67,"value":1192},"A",{"type":67,"value":1194},": ",{"type":61,"tag":112,"props":1196,"children":1198},{"className":1197},[],[1199],{"type":67,"value":204},{"type":67,"value":1201}," JSON API + ",{"type":61,"tag":112,"props":1203,"children":1205},{"className":1204},[],[1206],{"type":67,"value":1207},"DICTIONARY_API_KEY",{"type":67,"value":1209}," (free at dictionaryapi.com) — recommended for standalone use. ",{"type":61,"tag":78,"props":1211,"children":1212},{},[1213],{"type":67,"value":1214},"B",{"type":67,"value":1216},": HTML scrape of ",{"type":61,"tag":112,"props":1218,"children":1220},{"className":1219},[],[1221],{"type":67,"value":212},{"type":67,"value":1223}," — no key, brittle to site HTML changes; recipe inlined in ",{"type":61,"tag":112,"props":1225,"children":1227},{"className":1226},[],[1228],{"type":67,"value":1229},"references\u002Fpronunciation-pipeline.md",{"type":67,"value":1187},{"type":61,"tag":78,"props":1232,"children":1233},{},[1234],{"type":67,"value":1235},"C",{"type":67,"value":1237},": skip MW, fall through to Magpie G2P with weaker long-tail coverage. Both recipes + the full respelling→IPA table live in ",{"type":61,"tag":112,"props":1239,"children":1241},{"className":1240},[],[1242],{"type":67,"value":1229},{"type":67,"value":1244},". The Path A function takes ",{"type":61,"tag":112,"props":1246,"children":1248},{"className":1247},[],[1249],{"type":67,"value":1250},"api_key",{"type":67,"value":1252}," as an arg (never reads ",{"type":61,"tag":112,"props":1254,"children":1256},{"className":1255},[],[1257],{"type":67,"value":1258},"os.environ",{"type":67,"value":1260},"); pass ",{"type":61,"tag":112,"props":1262,"children":1264},{"className":1263},[],[1265],{"type":67,"value":1266},"None",{"type":67,"value":1268}," to skip MW.",{"type":61,"tag":74,"props":1270,"children":1271},{},[1272,1277],{"type":61,"tag":112,"props":1273,"children":1275},{"className":1274},[],[1276],{"type":67,"value":1087},{"type":67,"value":1278}," schema:",{"type":61,"tag":348,"props":1280,"children":1282},{"className":818,"code":1281,"language":820,"meta":356,"style":356},"term,ipa,verified_by,verified_at,notes\ncefazolin,sɛfəˈzoʊlɪn,brandoing,2026-05-13,confirmed against MW respelling + ear test\n",[1283],{"type":61,"tag":112,"props":1284,"children":1285},{"__ignoreMap":356},[1286,1294],{"type":61,"tag":826,"props":1287,"children":1288},{"class":828,"line":829},[1289],{"type":61,"tag":826,"props":1290,"children":1291},{},[1292],{"type":67,"value":1293},"term,ipa,verified_by,verified_at,notes\n",{"type":61,"tag":826,"props":1295,"children":1296},{"class":828,"line":54},[1297],{"type":61,"tag":826,"props":1298,"children":1299},{},[1300],{"type":67,"value":1301},"cefazolin,sɛfəˈzoʊlɪn,brandoing,2026-05-13,confirmed against MW respelling + ear test\n",{"type":61,"tag":74,"props":1303,"children":1304},{},[1305],{"type":67,"value":1306},"Append-only across cycles. Re-running the build later picks up new entries automatically.",{"type":61,"tag":726,"props":1308,"children":1310},{"id":1309},"_2d-qa-mode-synthesis-do-not-skip-this-gate",[1311,1313,1317],{"type":67,"value":1312},"2d. QA-mode synthesis (do ",{"type":61,"tag":78,"props":1314,"children":1315},{},[1316],{"type":67,"value":432},{"type":67,"value":1318}," skip this gate)",{"type":61,"tag":74,"props":1320,"children":1321},{},[1322,1324,1329],{"type":67,"value":1323},"Before running the full Cartesian product, synthesize ",{"type":61,"tag":78,"props":1325,"children":1326},{},[1327],{"type":67,"value":1328},"one wav per term",{"type":67,"value":1330}," with: first voice, clean noise, default context. Audition each clip with the user.",{"type":61,"tag":74,"props":1332,"children":1333},{},[1334,1336,1341,1343,1347],{"type":67,"value":1335},"For every term tagged ",{"type":61,"tag":112,"props":1337,"children":1339},{"className":1338},[],[1340],{"type":67,"value":1128},{"type":67,"value":1342},", propose an IPA candidate using clinical suffix patterns and validate against Magpie's en-US phoneme set ",{"type":61,"tag":78,"props":1344,"children":1345},{},[1346],{"type":67,"value":89},{"type":67,"value":1348}," suggesting:",{"type":61,"tag":154,"props":1350,"children":1351},{},[1352,1368],{"type":61,"tag":158,"props":1353,"children":1354},{},[1355],{"type":61,"tag":162,"props":1356,"children":1357},{},[1358,1363],{"type":61,"tag":166,"props":1359,"children":1360},{},[1361],{"type":67,"value":1362},"Suffix",{"type":61,"tag":166,"props":1364,"children":1365},{},[1366],{"type":67,"value":1367},"Stress pattern (example)",{"type":61,"tag":182,"props":1369,"children":1370},{},[1371,1388,1405,1422,1439,1456,1473],{"type":61,"tag":162,"props":1372,"children":1373},{},[1374,1383],{"type":61,"tag":189,"props":1375,"children":1376},{},[1377],{"type":61,"tag":112,"props":1378,"children":1380},{"className":1379},[],[1381],{"type":67,"value":1382},"-mycin",{"type":61,"tag":189,"props":1384,"children":1385},{},[1386],{"type":67,"value":1387},"…ˈmaɪsɪn (vancomycin, gentamicin)",{"type":61,"tag":162,"props":1389,"children":1390},{},[1391,1400],{"type":61,"tag":189,"props":1392,"children":1393},{},[1394],{"type":61,"tag":112,"props":1395,"children":1397},{"className":1396},[],[1398],{"type":67,"value":1399},"-prazole",{"type":61,"tag":189,"props":1401,"children":1402},{},[1403],{"type":67,"value":1404},"…ˈpreɪzoʊl (esomeprazole, omeprazole)",{"type":61,"tag":162,"props":1406,"children":1407},{},[1408,1417],{"type":61,"tag":189,"props":1409,"children":1410},{},[1411],{"type":61,"tag":112,"props":1412,"children":1414},{"className":1413},[],[1415],{"type":67,"value":1416},"-statin",{"type":61,"tag":189,"props":1418,"children":1419},{},[1420],{"type":67,"value":1421},"…ˈstætɪn (atorvastatin, rosuvastatin)",{"type":61,"tag":162,"props":1423,"children":1424},{},[1425,1434],{"type":61,"tag":189,"props":1426,"children":1427},{},[1428],{"type":61,"tag":112,"props":1429,"children":1431},{"className":1430},[],[1432],{"type":67,"value":1433},"-sartan",{"type":61,"tag":189,"props":1435,"children":1436},{},[1437],{"type":67,"value":1438},"…ˈsɑːrtən (losartan, valsartan)",{"type":61,"tag":162,"props":1440,"children":1441},{},[1442,1451],{"type":61,"tag":189,"props":1443,"children":1444},{},[1445],{"type":61,"tag":112,"props":1446,"children":1448},{"className":1447},[],[1449],{"type":67,"value":1450},"-azole",{"type":61,"tag":189,"props":1452,"children":1453},{},[1454],{"type":67,"value":1455},"…ˈeɪzoʊl (fluconazole, ketoconazole)",{"type":61,"tag":162,"props":1457,"children":1458},{},[1459,1468],{"type":61,"tag":189,"props":1460,"children":1461},{},[1462],{"type":61,"tag":112,"props":1463,"children":1465},{"className":1464},[],[1466],{"type":67,"value":1467},"-cillin",{"type":61,"tag":189,"props":1469,"children":1470},{},[1471],{"type":67,"value":1472},"…ˈsɪlɪn (amoxicillin, piperacillin)",{"type":61,"tag":162,"props":1474,"children":1475},{},[1476,1485],{"type":61,"tag":189,"props":1477,"children":1478},{},[1479],{"type":61,"tag":112,"props":1480,"children":1482},{"className":1481},[],[1483],{"type":67,"value":1484},"-parin",{"type":61,"tag":189,"props":1486,"children":1487},{},[1488],{"type":67,"value":1489},"…ˈpɛərɪn (enoxaparin, heparin)",{"type":61,"tag":74,"props":1491,"children":1492},{},[1493,1498,1500,1505,1507,1513,1515,1521,1523,1529,1531,1536],{"type":61,"tag":78,"props":1494,"children":1495},{},[1496],{"type":67,"value":1497},"Phoneme-validation pattern",{"type":67,"value":1499}," — live-probe Magpie's en-US neural G2P with a candidate IPA. If Magpie accepts the SSML, the IPA is in its inventory. Use the suffix patterns above as a ",{"type":61,"tag":770,"props":1501,"children":1502},{},[1503],{"type":67,"value":1504},"pre-filter",{"type":67,"value":1506}," (cheap heuristic) and the live probe to confirm before committing to an override. The ",{"type":61,"tag":112,"props":1508,"children":1510},{"className":1509},[],[1511],{"type":67,"value":1512},"magpie_validates_ipa(ipa, api_key, voice_id)",{"type":67,"value":1514}," recipe — a minimal NVCF gRPC synthesis call that returns ",{"type":61,"tag":112,"props":1516,"children":1518},{"className":1517},[],[1519],{"type":67,"value":1520},"True",{"type":67,"value":1522},"\u002F",{"type":61,"tag":112,"props":1524,"children":1526},{"className":1525},[],[1527],{"type":67,"value":1528},"False",{"type":67,"value":1530}," fail-closed — is in ",{"type":61,"tag":112,"props":1532,"children":1534},{"className":1533},[],[1535],{"type":67,"value":1229},{"type":67,"value":135},{"type":61,"tag":74,"props":1538,"children":1539},{},[1540,1542,1547,1549,1554,1556,1561,1563,1569],{"type":67,"value":1541},"Call it once per candidate IPA before showing it to the user. On user approval, append the verified IPA to ",{"type":61,"tag":112,"props":1543,"children":1545},{"className":1544},[],[1546],{"type":67,"value":1087},{"type":67,"value":1548},". The row's ",{"type":61,"tag":112,"props":1550,"children":1552},{"className":1551},[],[1553],{"type":67,"value":318},{"type":67,"value":1555}," flips from ",{"type":61,"tag":112,"props":1557,"children":1559},{"className":1558},[],[1560],{"type":67,"value":1128},{"type":67,"value":1562}," to ",{"type":61,"tag":112,"props":1564,"children":1566},{"className":1565},[],[1567],{"type":67,"value":1568},"override",{"type":67,"value":1570}," on the next manifest generation.",{"type":61,"tag":74,"props":1572,"children":1573},{},[1574,1579,1581,1586,1588,1593],{"type":61,"tag":78,"props":1575,"children":1576},{},[1577],{"type":67,"value":1578},"HITL audition gate before Step 2e — fail-closed.",{"type":67,"value":1580}," Do not synthesize the full Cartesian product, do not promote any staged IPA candidate to ",{"type":61,"tag":112,"props":1582,"children":1584},{"className":1583},[],[1585],{"type":67,"value":1087},{"type":67,"value":1587},", and do not advance to Stage 3 until ",{"type":61,"tag":78,"props":1589,"children":1590},{},[1591],{"type":67,"value":1592},"one of the following has happened explicitly in conversation",{"type":67,"value":815},{"type":61,"tag":763,"props":1595,"children":1596},{},[1597,1659],{"type":61,"tag":388,"props":1598,"children":1599},{},[1600,1605,1607,1613,1615,1621,1623,1629,1630,1636,1638,1643,1645,1650,1652,1657],{"type":61,"tag":78,"props":1601,"children":1602},{},[1603],{"type":67,"value":1604},"The user confirms they have auditioned the QA clips",{"type":67,"value":1606}," and reports their verdict per clip (or per bucket: \"the MW set sounds fine\", \"fix ",{"type":61,"tag":112,"props":1608,"children":1610},{"className":1609},[],[1611],{"type":67,"value":1612},"pembrolizumab",{"type":67,"value":1614},"\", etc.). Provide the ",{"type":61,"tag":112,"props":1616,"children":1618},{"className":1617},[],[1619],{"type":67,"value":1620},"afplay",{"type":67,"value":1622}," (macOS) or ",{"type":61,"tag":112,"props":1624,"children":1626},{"className":1625},[],[1627],{"type":67,"value":1628},"paplay",{"type":67,"value":1522},{"type":61,"tag":112,"props":1631,"children":1633},{"className":1632},[],[1634],{"type":67,"value":1635},"aplay",{"type":67,"value":1637}," (Linux) commands so the user can play them — then ",{"type":61,"tag":78,"props":1639,"children":1640},{},[1641],{"type":67,"value":1642},"halt and wait for their reply after listening",{"type":67,"value":1644},". Paper-only approval via an AskUserQuestion prompt — clicking \"Promote all\" or \"Lock in\" without auditioning — ",{"type":61,"tag":78,"props":1646,"children":1647},{},[1648],{"type":67,"value":1649},"does not satisfy this gate",{"type":67,"value":1651},". Magpie-validating an IPA proves it's in the phoneme inventory; it does not prove it matches the ",{"type":61,"tag":770,"props":1653,"children":1654},{},[1655],{"type":67,"value":1656},"intended",{"type":67,"value":1658}," pronunciation. Only the user's ears do that.",{"type":61,"tag":388,"props":1660,"children":1661},{},[1662,1667,1669,1674,1676,1682],{"type":61,"tag":78,"props":1663,"children":1664},{},[1665],{"type":67,"value":1666},"The user explicitly opts to skip audition for this cycle",{"type":67,"value":1668},", in deliberate language (e.g. ",{"type":61,"tag":770,"props":1670,"children":1671},{},[1672],{"type":67,"value":1673},"\"skip audition, accept the risk that mispronunciations may dilute the Stage 3 KER signal — log it as a cycle-N caveat\"",{"type":67,"value":1675},"), not as a side-effect of a single click-through. Record the skip in a cycle-level note (e.g. ",{"type":61,"tag":112,"props":1677,"children":1679},{"className":1678},[],[1680],{"type":67,"value":1681},"eval\u002Fcycle\u003CN>\u002Fcycle_notes.md",{"type":67,"value":1683},") so a future operator can see the audition was deferred.",{"type":61,"tag":74,"props":1685,"children":1686},{},[1687],{"type":67,"value":1688},"Magpie NVCF rate-limits aggressively on >100-row jobs, and a do-over costs both API credits and clock time — but the larger risk is shipping a manifest with mispronounced reference audio that quietly corrupts the Stage 3 KER signal. Time spent auditioning is cheaper than re-running the cycle.",{"type":61,"tag":726,"props":1690,"children":1692},{"id":1691},"_2e-full-benchmark-generation",[1693],{"type":67,"value":1694},"2e. Full benchmark generation",{"type":61,"tag":74,"props":1696,"children":1697},{},[1698,1700,1706,1708,1714,1715,1721],{"type":67,"value":1699},"After pronunciations are locked, generate the full Cartesian product ",{"type":61,"tag":112,"props":1701,"children":1703},{"className":1702},[],[1704],{"type":67,"value":1705},"|terms| × |voices| × |noise_levels| × |context_types|",{"type":67,"value":1707},". Defaults: 2–4 Magpie en-US voices (Mia\u002FJason\u002FRay), ",{"type":61,"tag":112,"props":1709,"children":1711},{"className":1710},[],[1712],{"type":67,"value":1713},"[clean, snr_15db, snr_5db]",{"type":67,"value":305},{"type":61,"tag":112,"props":1716,"children":1718},{"className":1717},[],[1719],{"type":67,"value":1720},"[dictation, handoff, chart_note, history]",{"type":67,"value":135},{"type":61,"tag":74,"props":1723,"children":1724},{},[1725,1727,1732,1734,1740,1742,1748,1750,1756,1758,1763,1765,1771,1773,1778,1780,1785,1787,1793,1795,1801,1803,1809],{"type":67,"value":1726},"Self-contained synthesis — no ",{"type":61,"tag":112,"props":1728,"children":1730},{"className":1729},[],[1731],{"type":67,"value":527},{"type":67,"value":1733}," required. The ",{"type":61,"tag":112,"props":1735,"children":1737},{"className":1736},[],[1738],{"type":67,"value":1739},"synthesize_row(row, all_overrides, out_dir, api_key)",{"type":67,"value":1741}," recipe — opens an NVCF gRPC stream, wraps overrides into SSML via ",{"type":61,"tag":112,"props":1743,"children":1745},{"className":1744},[],[1746],{"type":67,"value":1747},"render_sentence_with_overrides",{"type":67,"value":1749},", writes 16-bit mono PCM to ",{"type":61,"tag":112,"props":1751,"children":1753},{"className":1752},[],[1754],{"type":67,"value":1755},"\u003Cout_dir>\u002Faudio\u002F\u003Cslug>.wav",{"type":67,"value":1757}," — is in ",{"type":61,"tag":112,"props":1759,"children":1761},{"className":1760},[],[1762],{"type":67,"value":1229},{"type":67,"value":1764}," (§Synthesis call). Key invariant: ",{"type":61,"tag":112,"props":1766,"children":1768},{"className":1767},[],[1769],{"type":67,"value":1770},"all_overrides",{"type":67,"value":1772}," carries ",{"type":61,"tag":770,"props":1774,"children":1775},{},[1776],{"type":67,"value":1777},"every",{"type":67,"value":1779}," entry from ",{"type":61,"tag":112,"props":1781,"children":1783},{"className":1782},[],[1784],{"type":67,"value":1087},{"type":67,"value":1786}," (including context-word overrides like ",{"type":61,"tag":112,"props":1788,"children":1790},{"className":1789},[],[1791],{"type":67,"value":1792},"intravenously",{"type":67,"value":1794},") so the renderer wraps any override whose verbatim text appears in ",{"type":61,"tag":112,"props":1796,"children":1798},{"className":1797},[],[1799],{"type":67,"value":1800},"row['text']",{"type":67,"value":1802},". Wrapping only ",{"type":61,"tag":112,"props":1804,"children":1806},{"className":1805},[],[1807],{"type":67,"value":1808},"row['term']",{"type":67,"value":1810}," silently drops context-word overrides.",{"type":61,"tag":74,"props":1812,"children":1813},{},[1814,1816,1822,1823,1829,1831,1837],{"type":67,"value":1815},"Noise-injection (clean → ",{"type":61,"tag":112,"props":1817,"children":1819},{"className":1818},[],[1820],{"type":67,"value":1821},"snr_15db",{"type":67,"value":548},{"type":61,"tag":112,"props":1824,"children":1826},{"className":1825},[],[1827],{"type":67,"value":1828},"snr_5db",{"type":67,"value":1830},") and the manifest schema (NeMo canonical fields + clinical extension, plus pre-flight schema and audio-existence checks) all live in ",{"type":61,"tag":112,"props":1832,"children":1834},{"className":1833},[],[1835],{"type":67,"value":1836},"references\u002Fmanifest-schema.md",{"type":67,"value":135},{"type":61,"tag":74,"props":1839,"children":1840},{},[1841,1846,1848,1854],{"type":61,"tag":78,"props":1842,"children":1843},{},[1844],{"type":67,"value":1845},"Warn when product > 100 rows.",{"type":67,"value":1847}," Magpie NVCF rate-limits with ~5–10% ",{"type":61,"tag":112,"props":1849,"children":1851},{"className":1850},[],[1852],{"type":67,"value":1853},"RESOURCE_EXHAUSTED",{"type":67,"value":1855}," drops on big runs. Re-run the dropped rows.",{"type":61,"tag":726,"props":1857,"children":1859},{"id":1858},"stage-2-completion-checklist",[1860],{"type":67,"value":1861},"Stage 2 completion checklist",{"type":61,"tag":74,"props":1863,"children":1864},{},[1865],{"type":67,"value":1866},"Don't consider Stage 2 done until all five sub-steps ran. Agents commonly stop after 2a or 2b; the goal is a synthesized manifest plus a hand-off:",{"type":61,"tag":384,"props":1868,"children":1869},{},[1870,1892,1909,1925,1935,1951],{"type":61,"tag":388,"props":1871,"children":1872},{},[1873,1878,1879,1884,1886],{"type":61,"tag":78,"props":1874,"children":1875},{},[1876],{"type":67,"value":1877},"2a",{"type":67,"value":646},{"type":61,"tag":112,"props":1880,"children":1882},{"className":1881},[],[1883],{"type":67,"value":737},{"type":67,"value":1885},", 4–10 terms, ",{"type":61,"tag":112,"props":1887,"children":1889},{"className":1888},[],[1890],{"type":67,"value":1891},"entity_category ∈ {drug, procedure, anatomy, condition, lab, role}",{"type":61,"tag":388,"props":1893,"children":1894},{},[1895,1900,1902,1907],{"type":61,"tag":78,"props":1896,"children":1897},{},[1898],{"type":67,"value":1899},"2b",{"type":67,"value":1901}," — 3–5 ",{"type":61,"tag":112,"props":1903,"children":1905},{"className":1904},[],[1906],{"type":67,"value":339},{"type":67,"value":1908}," sentence variants per term",{"type":61,"tag":388,"props":1910,"children":1911},{},[1912,1917,1919],{"type":61,"tag":78,"props":1913,"children":1914},{},[1915],{"type":67,"value":1916},"2c",{"type":67,"value":1918}," — every term tagged ",{"type":61,"tag":112,"props":1920,"children":1922},{"className":1921},[],[1923],{"type":67,"value":1924},"ipa_source ∈ {override, merriam-webster, magpie_g2p}",{"type":61,"tag":388,"props":1926,"children":1927},{},[1928,1933],{"type":61,"tag":78,"props":1929,"children":1930},{},[1931],{"type":67,"value":1932},"2d",{"type":67,"value":1934}," — QA wavs auditioned, IPA overrides locked with explicit user approval",{"type":61,"tag":388,"props":1936,"children":1937},{},[1938,1943,1944,1949],{"type":61,"tag":78,"props":1939,"children":1940},{},[1941],{"type":67,"value":1942},"2e",{"type":67,"value":646},{"type":61,"tag":112,"props":1945,"children":1947},{"className":1946},[],[1948],{"type":67,"value":125},{"type":67,"value":1950}," + per-row audio for the Cartesian product",{"type":61,"tag":388,"props":1952,"children":1953},{},[1954,1959,1961,1966,1968,1972],{"type":61,"tag":78,"props":1955,"children":1956},{},[1957],{"type":67,"value":1958},"Hand-off",{"type":67,"value":1960}," — name ",{"type":61,"tag":112,"props":1962,"children":1964},{"className":1963},[],[1965],{"type":67,"value":133},{"type":67,"value":1967}," as the next skill and ",{"type":61,"tag":78,"props":1969,"children":1970},{},[1971],{"type":67,"value":96},{"type":67,"value":1973}," as its headline metric",{"type":61,"tag":74,"props":1975,"children":1976},{},[1977,1979,1984,1986,1991],{"type":67,"value":1978},"Writes go only into the user-chosen ",{"type":61,"tag":112,"props":1980,"children":1982},{"className":1981},[],[1983],{"type":67,"value":716},{"type":67,"value":1985},". Don't write elsewhere, modify env, or install packages — those belong to ",{"type":61,"tag":112,"props":1987,"children":1989},{"className":1988},[],[1990],{"type":67,"value":117},{"type":67,"value":135},{"type":61,"tag":142,"props":1993,"children":1995},{"id":1994},"examples",[1996],{"type":67,"value":1997},"Examples",{"type":61,"tag":74,"props":1999,"children":2000},{},[2001,2006,2008,2013,2015,2021,2022,2028,2029,2034,2035,2041,2042,2048,2049,2055,2056,2062,2063,2069,2070,2076,2077,2083,2085,2090,2092,2098,2100,2105,2107,2112,2114,2119,2121,2126,2128,2134],{"type":61,"tag":78,"props":2002,"children":2003},{},[2004],{"type":67,"value":2005},"Scenario A — fresh oncology benchmark.",{"type":67,"value":2007}," User: ",{"type":61,"tag":770,"props":2009,"children":2010},{},[2011],{"type":67,"value":2012},"\"We're seeing chemo drug names mistranscribed. Where do I start?\"",{"type":67,"value":2014}," → Step 2a: confirm specialty is oncology, ask about which drugs (immunotherapy biologics, platinum agents, taxanes). Propose ~10 candidates: ",{"type":61,"tag":112,"props":2016,"children":2018},{"className":2017},[],[2019],{"type":67,"value":2020},"cisplatin",{"type":67,"value":305},{"type":61,"tag":112,"props":2023,"children":2025},{"className":2024},[],[2026],{"type":67,"value":2027},"paclitaxel",{"type":67,"value":305},{"type":61,"tag":112,"props":2030,"children":2032},{"className":2031},[],[2033],{"type":67,"value":1612},{"type":67,"value":305},{"type":61,"tag":112,"props":2036,"children":2038},{"className":2037},[],[2039],{"type":67,"value":2040},"nivolumab",{"type":67,"value":305},{"type":61,"tag":112,"props":2043,"children":2045},{"className":2044},[],[2046],{"type":67,"value":2047},"carboplatin",{"type":67,"value":305},{"type":61,"tag":112,"props":2050,"children":2052},{"className":2051},[],[2053],{"type":67,"value":2054},"docetaxel",{"type":67,"value":305},{"type":61,"tag":112,"props":2057,"children":2059},{"className":2058},[],[2060],{"type":67,"value":2061},"bevacizumab",{"type":67,"value":305},{"type":61,"tag":112,"props":2064,"children":2066},{"className":2065},[],[2067],{"type":67,"value":2068},"trastuzumab",{"type":67,"value":305},{"type":61,"tag":112,"props":2071,"children":2073},{"className":2072},[],[2074],{"type":67,"value":2075},"cetuximab",{"type":67,"value":305},{"type":61,"tag":112,"props":2078,"children":2080},{"className":2079},[],[2081],{"type":67,"value":2082},"pemetrexed",{"type":67,"value":2084},". Write ",{"type":61,"tag":112,"props":2086,"children":2088},{"className":2087},[],[2089],{"type":67,"value":737},{"type":67,"value":2091}," with all ",{"type":61,"tag":112,"props":2093,"children":2095},{"className":2094},[],[2096],{"type":67,"value":2097},"entity_category=drug",{"type":67,"value":2099},". Step 2b: brief ",{"type":61,"tag":112,"props":2101,"children":2103},{"className":2102},[],[2104],{"type":67,"value":272},{"type":67,"value":2106}," for 4 context variants each = 40 sentences. Step 2c: MW lookup for each — biologics like ",{"type":61,"tag":112,"props":2108,"children":2110},{"className":2109},[],[2111],{"type":67,"value":1612},{"type":67,"value":2113}," will likely fall to ",{"type":61,"tag":112,"props":2115,"children":2117},{"className":2116},[],[2118],{"type":67,"value":1128},{"type":67,"value":2120},"; platinum agents likely hit MW. Step 2d: synthesize one QA wav per term, walk the user through the ",{"type":61,"tag":112,"props":2122,"children":2124},{"className":2123},[],[2125],{"type":67,"value":1612},{"type":67,"value":2127}," etc. clips, propose IPA candidates with ",{"type":61,"tag":112,"props":2129,"children":2131},{"className":2130},[],[2132],{"type":67,"value":2133},"-mab",{"type":67,"value":2135}," suffix stress patterns. Step 2e: on approval, run 10 terms × 2 voices × 2 noise levels × 3 contexts = 120 rows.",{"type":61,"tag":74,"props":2137,"children":2138},{},[2139,2144,2145,2150,2152,2157,2158,2163],{"type":61,"tag":78,"props":2140,"children":2141},{},[2142],{"type":67,"value":2143},"Scenario B — appending to an existing cycle.",{"type":67,"value":2007},{"type":61,"tag":770,"props":2146,"children":2147},{},[2148],{"type":67,"value":2149},"\"I have a cycle-1 manifest and I want to add 5 more procedures.\"",{"type":67,"value":2151}," → Re-run only Steps 2a (specialty interview just for the new terms), 2b (sentence gen for the additions), 2c (IPA pipeline for the additions), 2d (audition the new terms), and 2e (synthesize only the new term rows). Append to the existing ",{"type":61,"tag":112,"props":2153,"children":2155},{"className":2154},[],[2156],{"type":67,"value":125},{"type":67,"value":1187},{"type":61,"tag":78,"props":2159,"children":2160},{},[2161],{"type":67,"value":2162},"Do not regenerate audio for existing terms",{"type":67,"value":2164}," — cycle isolation is intentional so leaderboards diff cycle N vs cycle N+1 cleanly.",{"type":61,"tag":142,"props":2166,"children":2168},{"id":2167},"artifacts-produced",[2169],{"type":67,"value":2170},"Artifacts produced",{"type":61,"tag":384,"props":2172,"children":2173},{},[2174,2189,2204,2214],{"type":61,"tag":388,"props":2175,"children":2176},{},[2177,2182,2184],{"type":61,"tag":112,"props":2178,"children":2180},{"className":2179},[],[2181],{"type":67,"value":737},{"type":67,"value":2183}," — curated terms with ",{"type":61,"tag":112,"props":2185,"children":2187},{"className":2186},[],[2188],{"type":67,"value":311},{"type":61,"tag":388,"props":2190,"children":2191},{},[2192,2197,2199],{"type":61,"tag":112,"props":2193,"children":2195},{"className":2194},[],[2196],{"type":67,"value":1087},{"type":67,"value":2198}," — verified IPA, ",{"type":61,"tag":78,"props":2200,"children":2201},{},[2202],{"type":67,"value":2203},"appendable across cycles",{"type":61,"tag":388,"props":2205,"children":2206},{},[2207,2212],{"type":61,"tag":112,"props":2208,"children":2210},{"className":2209},[],[2211],{"type":67,"value":125},{"type":67,"value":2213}," — NeMo format with clinical extension fields (one JSON object per line)",{"type":61,"tag":388,"props":2215,"children":2216},{},[2217,2223],{"type":61,"tag":112,"props":2218,"children":2220},{"className":2219},[],[2221],{"type":67,"value":2222},"audio\u002F\u003Cslug>.wav",{"type":67,"value":2224}," — synthesized clips, one per manifest row",{"type":61,"tag":142,"props":2226,"children":2228},{"id":2227},"troubleshooting",[2229],{"type":67,"value":2230},"Troubleshooting",{"type":61,"tag":384,"props":2232,"children":2233},{},[2234,2257,2286,2309,2326],{"type":61,"tag":388,"props":2235,"children":2236},{},[2237,2248,2250,2255],{"type":61,"tag":78,"props":2238,"children":2239},{},[2240,2242,2247],{"type":67,"value":2241},"TTS rate-limit drops (",{"type":61,"tag":112,"props":2243,"children":2245},{"className":2244},[],[2246],{"type":67,"value":1853},{"type":67,"value":244},{"type":67,"value":2249}," on >100-row generation → expected on Magpie NVCF. Confirm exponential backoff is active in ",{"type":61,"tag":112,"props":2251,"children":2253},{"className":2252},[],[2254],{"type":67,"value":527},{"type":67,"value":2256},"; expect ~5–10% drops on big runs and re-run for the gaps.",{"type":61,"tag":388,"props":2258,"children":2259},{},[2260,2277,2279,2284],{"type":61,"tag":78,"props":2261,"children":2262},{},[2263,2265,2270,2272],{"type":67,"value":2264},"All ",{"type":61,"tag":112,"props":2266,"children":2268},{"className":2267},[],[2269],{"type":67,"value":318},{"type":67,"value":2271}," rows tagged ",{"type":61,"tag":112,"props":2273,"children":2275},{"className":2274},[],[2276],{"type":67,"value":1128},{"type":67,"value":2278}," → MW lookup is failing across the board, or candidate IPAs are failing phoneme validation. Re-verify whichever MW path you configured (",{"type":61,"tag":112,"props":2280,"children":2282},{"className":2281},[],[2283],{"type":67,"value":1207},{"type":67,"value":2285}," for A; HTTPS reachability + parser for B), then check candidates against Magpie's en-US phoneme inventory.",{"type":61,"tag":388,"props":2287,"children":2288},{},[2289,2294,2296,2301,2302,2307],{"type":61,"tag":78,"props":2290,"children":2291},{},[2292],{"type":67,"value":2293},"Magpie mispronounces a term even with the IPA override",{"type":67,"value":2295}," → first verify the IPA is in the Magpie en-US phoneme inventory and the SSML wrapping is syntactically valid. If both check out, the underlying TTS bug is owned by ",{"type":61,"tag":112,"props":2297,"children":2299},{"className":2298},[],[2300],{"type":67,"value":527},{"type":67,"value":198},{"type":61,"tag":112,"props":2303,"children":2305},{"className":2304},[],[2306],{"type":67,"value":535},{"type":67,"value":2308},") — route there for diagnosis. This skill provides the override mechanism but does not own the neural G2P or SSML parser.",{"type":61,"tag":388,"props":2310,"children":2311},{},[2312,2324],{"type":61,"tag":78,"props":2313,"children":2314},{},[2315,2317,2322],{"type":67,"value":2316},"Sentence variants from ",{"type":61,"tag":112,"props":2318,"children":2320},{"className":2319},[],[2321],{"type":67,"value":272},{"type":67,"value":2323}," are bland \u002F template-like",{"type":67,"value":2325}," → check the brief; the schema-only prompt sometimes produces stereotyped output. Add 1–2 in-context examples to the brief and re-run.",{"type":61,"tag":388,"props":2327,"children":2328},{},[2329,2341],{"type":61,"tag":78,"props":2330,"children":2331},{},[2332,2334,2339],{"type":67,"value":2333},"Audio files exist but ",{"type":61,"tag":112,"props":2335,"children":2337},{"className":2336},[],[2338],{"type":67,"value":125},{"type":67,"value":2340}," is short",{"type":67,"value":2342}," → manifest writer skipped rows whose synthesis returned a NVCF error. Re-run the build with only the missing rows.",{"type":61,"tag":74,"props":2344,"children":2345},{},[2346,2348,2353],{"type":67,"value":2347},"For anything not in this list, identify which upstream skill is implicated and route there. The ",{"type":61,"tag":112,"props":2349,"children":2351},{"className":2350},[],[2352],{"type":67,"value":4},{"type":67,"value":2354}," skill owns the methodology, not the TTS or DataDesigner internals.",{"type":61,"tag":142,"props":2356,"children":2358},{"id":2357},"limitations",[2359],{"type":67,"value":2360},"Limitations",{"type":61,"tag":384,"props":2362,"children":2363},{},[2364,2374,2391,2408],{"type":61,"tag":388,"props":2365,"children":2366},{},[2367,2372],{"type":61,"tag":78,"props":2368,"children":2369},{},[2370],{"type":67,"value":2371},"English-only by default.",{"type":67,"value":2373}," Magpie's en-US phoneme inventory is what the two-tier IPA pipeline validates against. Other locales need a different upstream phoneme set + override CSV format.",{"type":61,"tag":388,"props":2375,"children":2376},{},[2377,2382,2384,2389],{"type":61,"tag":78,"props":2378,"children":2379},{},[2380],{"type":67,"value":2381},"Six fixed entity categories.",{"type":67,"value":2383}," Extending ",{"type":61,"tag":112,"props":2385,"children":2387},{"className":2386},[],[2388],{"type":67,"value":311},{"type":67,"value":2390}," is a deliberate methodology change, not a one-off tweak — KER breakdowns, leaderboard sections, and downstream finetune scripts all key off the vocabulary.",{"type":61,"tag":388,"props":2392,"children":2393},{},[2394,2399,2401,2406],{"type":61,"tag":78,"props":2395,"children":2396},{},[2397],{"type":67,"value":2398},"Tiny first cycles.",{"type":67,"value":2400}," Below ~20 terms, the by-",{"type":61,"tag":112,"props":2402,"children":2404},{"className":2403},[],[2405],{"type":67,"value":318},{"type":67,"value":2407}," leaderboard split won't have enough rows in each bucket to be statistically meaningful. Build a meaningful cycle even if it costs a session.",{"type":61,"tag":388,"props":2409,"children":2410},{},[2411,2416],{"type":61,"tag":78,"props":2412,"children":2413},{},[2414],{"type":67,"value":2415},"Magpie NVCF rate-limits.",{"type":67,"value":2417}," ~5–10% drops on large jobs; budget a re-run pass.",{"type":61,"tag":142,"props":2419,"children":2421},{"id":2420},"next-steps",[2422],{"type":67,"value":2423},"Next steps",{"type":61,"tag":384,"props":2425,"children":2426},{},[2427,2444,2460],{"type":61,"tag":388,"props":2428,"children":2429},{},[2430,2435,2437,2442],{"type":61,"tag":78,"props":2431,"children":2432},{},[2433],{"type":67,"value":2434},"Forward:",{"type":67,"value":2436}," ",{"type":61,"tag":112,"props":2438,"children":2440},{"className":2439},[],[2441],{"type":67,"value":133},{"type":67,"value":2443}," — transcribe the manifest, score WER\u002FCER\u002FKER\u002FSER, produce the five-section leaderboard.",{"type":61,"tag":388,"props":2445,"children":2446},{},[2447,2452,2454,2459],{"type":61,"tag":78,"props":2448,"children":2449},{},[2450],{"type":67,"value":2451},"Back to setup",{"type":67,"value":2453}," (if anything in the env is broken): ",{"type":61,"tag":112,"props":2455,"children":2457},{"className":2456},[],[2458],{"type":67,"value":117},{"type":67,"value":135},{"type":61,"tag":388,"props":2461,"children":2462},{},[2463,2468,2470,2475,2476,2481],{"type":61,"tag":78,"props":2464,"children":2465},{},[2466],{"type":67,"value":2467},"Lateral",{"type":67,"value":2469}," for TTS-specific debugging: ",{"type":61,"tag":112,"props":2471,"children":2473},{"className":2472},[],[2474],{"type":67,"value":527},{"type":67,"value":555},{"type":61,"tag":112,"props":2477,"children":2479},{"className":2478},[],[2480],{"type":67,"value":535},{"type":67,"value":135},{"type":61,"tag":142,"props":2483,"children":2485},{"id":2484},"references",[2486],{"type":67,"value":2487},"References",{"type":61,"tag":384,"props":2489,"children":2490},{},[2491],{"type":61,"tag":388,"props":2492,"children":2493},{},[2494,2503],{"type":61,"tag":2495,"props":2496,"children":2497},"a",{"href":1836},[2498],{"type":61,"tag":112,"props":2499,"children":2501},{"className":2500},[],[2502],{"type":67,"value":1836},{"type":67,"value":2504}," — NeMo canonical fields + clinical extension; pre-flight schema and audio-existence checks; cross-cycle stability rules",{"type":61,"tag":2506,"props":2507,"children":2508},"style",{},[2509],{"type":67,"value":2510},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":2512,"total":2671},[2513,2531,2549,2560,2572,2586,2599,2613,2626,2637,2651,2660],{"slug":2514,"name":2514,"fn":2515,"description":2516,"org":2517,"tags":2518,"stars":2528,"repoUrl":2529,"updatedAt":2530},"nemoclaw-user-guide","retrieve NemoClaw documentation and configuration","Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2519,2522,2525],{"name":2520,"slug":2521,"type":15},"Documentation","documentation",{"name":2523,"slug":2524,"type":15},"MCP","mcp",{"name":2526,"slug":2527,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":2532,"name":2532,"fn":2533,"description":2534,"org":2535,"tags":2536,"stars":2546,"repoUrl":2547,"updatedAt":2548},"mcore-build-and-dependency","manage Megatron-LM development environments","Container-based dev environment setup and dependency management for Megatron-LM. Covers acquiring and launching the CI container, uv package management, and updating uv.lock.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2537,2540,2543],{"name":2538,"slug":2539,"type":15},"Containers","containers",{"name":2541,"slug":2542,"type":15},"Deployment","deployment",{"name":2544,"slug":2545,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":2550,"name":2550,"fn":2551,"description":2552,"org":2553,"tags":2554,"stars":2546,"repoUrl":2547,"updatedAt":2559},"mcore-bump-base-image","update NVIDIA PyTorch base images","Bump the NVIDIA PyTorch base image (`nvcr.io\u002Fnvidia\u002Fpytorch:YY.MM-py3`) used by Megatron-LM CI. Covers the two pin sites (GitHub CI in `docker\u002F.ngc_version.dev` and GitLab CI in `.gitlab\u002Fstages\u002F01.build.yml`), the post-bump CI loop (re-run functional tests, refresh golden values, mark broken tests), and the gotchas that bit PRs",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2555,2558],{"name":2556,"slug":2557,"type":15},"CI\u002FCD","ci-cd",{"name":2541,"slug":2542,"type":15},"2026-07-14T05:25:59.97109",{"slug":2561,"name":2561,"fn":2562,"description":2563,"org":2564,"tags":2565,"stars":2546,"repoUrl":2547,"updatedAt":2571},"mcore-cicd","manage CI\u002FCD pipelines for Megatron-LM","CI\u002FCD reference for Megatron-LM. Covers CI pipeline structure, PR scope labels, triggering internal GitLab CI (which force-pushes the current branch to a pull-request\u002FBRANCH ref — always dry-run and verify the destination first; never run against shared or protected branches), and CI failure investigation.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2566,2567,2568],{"name":2556,"slug":2557,"type":15},{"name":2541,"slug":2542,"type":15},{"name":2569,"slug":2570,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":2573,"name":2573,"fn":2574,"description":2575,"org":2576,"tags":2577,"stars":2546,"repoUrl":2547,"updatedAt":2585},"mcore-create-issue","investigate CI failures and create issues","Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2578,2581,2582],{"name":2579,"slug":2580,"type":15},"Debugging","debugging",{"name":2569,"slug":2570,"type":15},{"name":2583,"slug":2584,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":2587,"name":2587,"fn":2588,"description":2589,"org":2590,"tags":2591,"stars":2546,"repoUrl":2547,"updatedAt":2598},"mcore-linting-and-formatting","lint and format Megatron-LM code","Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2592,2595],{"name":2593,"slug":2594,"type":15},"Best Practices","best-practices",{"name":2596,"slug":2597,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":2600,"name":2600,"fn":2601,"description":2602,"org":2603,"tags":2604,"stars":2546,"repoUrl":2547,"updatedAt":2612},"mcore-migrate-gpt-to-hybrid","migrate Megatron-LM models to HybridModel","Migration guide for moving Megatron Core GPTModel checkpoints, model providers, training commands, and layer mappings to HybridModel.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2605,2608,2611],{"name":2606,"slug":2607,"type":15},"Machine Learning","machine-learning",{"name":2609,"slug":2610,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":2614,"name":2614,"fn":2615,"description":2616,"org":2617,"tags":2618,"stars":2546,"repoUrl":2547,"updatedAt":2625},"mcore-onboard-gb200-1node-tests","onboard functional tests for GB200","Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2619,2622],{"name":2620,"slug":2621,"type":15},"QA","qa",{"name":2623,"slug":2624,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":2627,"name":2627,"fn":2628,"description":2629,"org":2630,"tags":2631,"stars":2546,"repoUrl":2547,"updatedAt":2636},"mcore-run-on-slurm","launch distributed training jobs on SLURM","How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2632,2633],{"name":2541,"slug":2542,"type":15},{"name":2634,"slug":2635,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":2638,"name":2638,"fn":2639,"description":2640,"org":2641,"tags":2642,"stars":2546,"repoUrl":2547,"updatedAt":2650},"mcore-split-pr","split pull requests to reduce review load","Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2643,2646,2647],{"name":2644,"slug":2645,"type":15},"Code Review","code-review",{"name":2569,"slug":2570,"type":15},{"name":2648,"slug":2649,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":2652,"name":2652,"fn":2653,"description":2654,"org":2655,"tags":2656,"stars":2546,"repoUrl":2547,"updatedAt":2659},"mcore-testing","run and manage Megatron-LM tests","Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2657,2658],{"name":2620,"slug":2621,"type":15},{"name":2623,"slug":2624,"type":15},"2026-07-14T05:25:54.928983",{"slug":2661,"name":2661,"fn":2662,"description":2663,"org":2664,"tags":2665,"stars":2546,"repoUrl":2547,"updatedAt":2670},"nightly-sync","manage nightly main-to-dev sync workflows","Domain knowledge for the nightly main-to-dev sync workflow. Covers merge strategy, CI architecture, failure investigation, and known issues.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2666,2669],{"name":2667,"slug":2668,"type":15},"Automation","automation",{"name":2556,"slug":2557,"type":15},"2026-07-30T05:29:03.275638",496,{"items":2673,"total":2768},[2674,2691,2701,2715,2725,2740,2754],{"slug":2675,"name":2675,"fn":2676,"description":2677,"org":2678,"tags":2679,"stars":22,"repoUrl":23,"updatedAt":2690},"accelerated-computing-cudf","accelerate data processing with cuDF","Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV\u002FParquet I\u002FO, nullable semantics, and multi-GPU DataFrame workloads.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2680,2683,2686,2687],{"name":2681,"slug":2682,"type":15},"Data Analysis","data-analysis",{"name":2684,"slug":2685,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":2688,"slug":2689,"type":15},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":2692,"name":2692,"fn":2693,"description":2694,"org":2695,"tags":2696,"stars":22,"repoUrl":23,"updatedAt":2700},"aiq-deploy","deploy and manage NVIDIA AI-Q infrastructure","Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2697,2698,2699],{"name":2541,"slug":2542,"type":15},{"name":2634,"slug":2635,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":2702,"name":2702,"fn":2703,"description":2704,"org":2705,"tags":2706,"stars":22,"repoUrl":23,"updatedAt":2714},"aiq-research","conduct deep research with AI-Q","Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2707,2710,2711],{"name":2708,"slug":2709,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":2712,"slug":2713,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":2716,"name":2716,"fn":2717,"description":2718,"org":2719,"tags":2720,"stars":22,"repoUrl":23,"updatedAt":2724},"amc-run-sample-calibration","run AMC sample dataset calibration","Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2721,2722,2723],{"name":2681,"slug":2682,"type":15},{"name":9,"slug":8,"type":15},{"name":2623,"slug":2624,"type":15},"2026-07-17T05:29:03.913266",{"slug":2726,"name":2726,"fn":2727,"description":2728,"org":2729,"tags":2730,"stars":22,"repoUrl":23,"updatedAt":2739},"amc-run-video-calibration","calibrate video datasets with AutoMagicCalib","Calibrate a new dataset from pre-recorded video files via the AutoMagicCalib REST API. Use when user has local MP4s and says 'calibrate my videos', 'run AMC on these videos', or similar. For RTSP\u002Flive streams, use amc-run-rtsp-calibration instead.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2731,2732,2735,2736],{"name":2667,"slug":2668,"type":15},{"name":2733,"slug":2734,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":2737,"slug":2738,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":2741,"name":2741,"fn":2742,"description":2743,"org":2744,"tags":2745,"stars":22,"repoUrl":23,"updatedAt":2753},"amc-setup-calibration-stack","deploy AutoMagicCalib microservice with Docker","Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2746,2747,2749,2750],{"name":2541,"slug":2542,"type":15},{"name":490,"slug":2748,"type":15},"docker",{"name":9,"slug":8,"type":15},{"name":2751,"slug":2752,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":2755,"name":2755,"fn":2756,"description":2757,"org":2758,"tags":2759,"stars":22,"repoUrl":23,"updatedAt":2767},"cudaq-guide","develop quantum applications with CUDA-Q","CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2760,2761,2764],{"name":9,"slug":8,"type":15},{"name":2762,"slug":2763,"type":15},"Quantum Computing","quantum-computing",{"name":2765,"slug":2766,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]