docs/ADA_REFINEMENTS.md

# Ada-primary product — refinements → competitive translator

**Pivot (2026-07):** Python is **data factory only** (export gold/densify).  
**Ada/SPARK** is the shipping product binary (`pflt-Ada/bin/pflt_main.exe`).  
**Law master:** `I:\FSOT-Physical-Archive` (pin D1D38A). Sensory students: vision + audio (V5).

Ada forced quality rules Python bulk inject hid. V4 encodes competitive train mass.

## Refinements (exposed → fixed)

| Issue | Evidence | Fix |
|--------|----------|-----|
| Meta grammatical glosses | gold `singular of` / `dative` | `PFLT_Gloss_Quality` + export `clean_gloss` |
| Noisy compound definitions | `The compound water` | Clean + prefer short densify/seeds |
| English function words invented | `tell/about` mapped wrong | Closed-class identity pass |
| Open-set only exact | inflected forms missed | Reverse morph peels + tight edit-sim |
| Tiny densify-only open-set | V3 partial ~4.7% | **`train_mass.tsv` ~1.05M** keys |
| Edit-sim false positives | morph_resolve high, miss high | Dist-1 only; short-form ban; len gate |
| Paradigm leakage into eval | expand added held-out forms | Eval rewritten after expand if still absent |
| No Ada self-climb | Python-only inject | `pflt_main.exe inject FORM GLOSS` |

## Competitive tracks (beat the bar)

| Track | Competitor bar | Ada V4 attack |
|--------|----------------|---------------|
| Classical/visual form→gloss | Google weak/absent | train_mass + morph + quality gold deploy |
| Open-set partial | Internal ≥70% | Climb train densify + peels; honest held-out |
| Language catalog | Google ~249 / NLLB 200 | Focus langs: la grc egy ang en ar cu cop arc akk… |
| Offline binary + pin | Cloud MT | Native `pflt_main.exe` + D1D38A |
| Proven law | N/A | SPARK kernel + golden S |

## Live metrics (V4)

| Metric | Value |
|--------|--------|
| densify.tsv | ~96k |
| gold_core.tsv | ~1.02M quality rows |
| train_mass.tsv | ~1.02M (+ paradigm expand) |
| deploy map_entries | ~2.08M |
| honest open-set partial (n=8000, classical-weighted) | **~36.0%** (V3 densify-only ~4.7%) |
| honest open-set exact | **~32.8%** |
| converse `aqua lingua manus` | water language hand · map_rate=1.0 |
| converse `logos theos virtus` | word god virtue · map_rate=1.0 |
| GOLDEN law | PASS |

Eval is held-out forms absent from train_keys, classical-weighted (la 55% / grc 15% / ang / egy…).

## Commands

```powershell
cd C:\Users\damia\Desktop\pflt\pflt-Ada
python export_data_for_ada.py   # quality densify + gold + train_mass + eval
alr build
.\bin\pflt_main.exe eval
.\bin\pflt_main.exe converse "aqua lingua manus"
.\bin\pflt_main.exe inject aquarum water
.\bin\pflt_main.exe status
```

## Climb path to 70%+

1. Grow densify via `inject` + paradigm expand (export).  
2. More classical peels / sense banks for soft partial.  
3. Re-export after any Python factory densify growth.  
4. Keep eval held-out pure (no train key leakage).  
5. Own M5 classical/visual offline before modern BLEU (M6).

Python climb remains optional **fuel factory** only — product path is Ada.