File size: 7,397 Bytes
6c0bfac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
# Competitive roadmap — accurate status vs competition

**Law:** \(S=K(T_1+T_2+T_3)\) pin **D1D38A** (unchanged; never fitted to BLEU).  

**Measured:** 2026-07-21 SOTA push (`m6_sota_push_report.json` / v0.2.0 metrics).

---

## 1. Two competitions (do not mix them)

| Arena | Competitor definition of “win” | PFLT status |
|-------|--------------------------------|-------------|
| **A. Inventory / form→gloss** | Coverage + sense under offline densify | **Winning our catalog** (~99.99% on 113 langs) |
| **B. Sentence MT** | sacreBLEU / chrF / COMET on held-out sentences | **Chat: competitive**; **News: trailing DeepL-class** |

Google/DeepL/NLLB are judged almost entirely on **B**.  
We already win parts of **A** (especially classical / offline / FSOT). Beating them “accurately” means closing **B** without lying about A.

---

## 2. Scoreboard (measured)

### A — Catalog

| Metric | Us | Google-ish | DeepL-ish | NLLB-ish |
|--------|---:|-----------:|----------:|---------:|
| Lang count | **113** | ~249 | ~30–100 | ~200 |
| Form→gloss on our catalog | **~99.99%** | N/A (not their product metric) | N/A | N/A |
| Classical / dead / hieroglyph | **Strong** | Weak | Weak | Weak |
| FSOT law pin | **Yes** | No | No | No |

**Verdict A:** Competitive / leading on **our** product definition. Not “more languages than Google.”

### B1 — Chat / easy parallel (Tatoeba-style open-set)

| System path | sacreBLEU (mean) | Notes |
|-------------|-----------------:|-------|
| Neural best-of (opus/mul/NLLB) | **50.19** | Fair open-set |
| Hybrid oracle densify\|neural | **53.58** | Product upper bound |
| Staged internal bar | 45 | **Passed** |
| Strong open MT chat (rough) | ~45–65+ | We sit mid/high |

**Per-lang neural highs:** it 68 · es/pt 61 · de/ru ~57–58 · hi 55  
**Gaps:** ja 37 · zh 32 · la neural 13 (densify wins classical)

**Verdict B1:** **Competitive** on chat open-set. Not SOTA vs every commercial pair, but past mid bar.

### B2 — News (WMT14 de→en test, n=3003) — the hard public bar

| System | sacreBLEU | Gap to 40 mid | Gap to 48 stretch |
|--------|----------:|--------------:|------------------:|
| Densify-only | ~0.4 | — | — |
| **opus-mt-de-en beams=5** | **33.88** | **−6.1** | **−14.1** |
| NLLB-600M beams=5 | 33.37 | −6.6 | −14.6 |
| DeepL-class mid (staged) | ~40 | 0 | — |
| Strong commercial stretch | ~45–55 | — | 0 |

**Verdict B2:** **Near mid open-MT**, **not** DeepL-class yet. This is the main gap if “beat the competition” means news fluency.

---

## 3. Stage ladder (honest)

```text

S0 Dict          ████████████ DONE

S1 Phrase        ████████████ DONE

S2 Chat strong   ████████████ DONE  (neural ~50)

S3 Mid news      ████████░░░░ HERE  (~34 / need ~40)

S4 DeepL-class   ████░░░░░░░░       (need ~45–55)

```

---

## 4. What is left to be competitive / beat them (priority order)

| # | Lever | Closes | Effort | Expected gain |
|---|-------|--------|--------|----------------|
| **1** | **Ship hybrid router** (densify short/classical; neural long/news/CJK) | Product accuracy vs single-path | Low | Realizes measured **~53.6** chat hybrid; better UX |
| **2** | **WMT student finetune** (opus or NLLB on WMT train; law fixed) | News −6 gap | Med–High | Often **+2–6+** sacre if done right |
| **3** | **Stronger student** (NLLB-1.3B / more beams / length / ensembling) | News + chat CJK | Med | **+0.5–3** sacre (varies) |
| **4** | **CJK order** always neural SPM path (no EN-dep densify for ja/zh) | ja/zh chat | Low | Lift ja/zh from ~32–37 toward EU band |
| **5** | **FLORES** same-file public bar | Comparable claim vs NLLB papers | Blocked until Hub access | Credibility, not training |
| **6** | **Breadth** more Kaikki langs toward 150–200 | Catalog vs Google count | Med | Count game only |
| **7** | **COMET / human** spot checks | Beyond BLEU honesty | Med | Avoid BLEU-only overclaim |

### What will **not** beat DeepL alone

- More form→gloss densify on chat templates (product ceiling ≠ news SOTA)  
- Refitting the FSOT law scalar to BLEU (forbidden / false)  
- Claiming product densify BLEU-4 ~83 as open-set MT

---

## 5. Definition of “we beat the competition” (accurate)

Pick a claim level:

| Claim level | Criteria | Distance now |
|-------------|----------|--------------|
| **L1 Product unique** | Offline FSOT + classical + catalog depth | **Met** |
| **L2 Chat competitive** | Mean chat sacre ≥45 open-set multi-lang | **Met** (~50) |
| **L3 News mid-parity** | WMT14 de-en sacre ≥40 | **~6 pts short** |
| **L4 News strong** | WMT14 de-en sacre ≥45–48 | **~11–14 pts short** |
| **L5 Broad DeepL-class** | Many pairs FLORES/WMT mid-40s+ | **Not started** (FLORES gated; multi-pair news unrun) |

**Accurate one-liner:**  
We **beat / own** unique offline+classical+law. We **match mid open-MT on chat**. We **do not yet beat** commercial systems on **news full-sentence fluency** — need roughly **+6 sacre** for mid-parity and **+14** for stretch SOTA on de→en.

---

## 6. Immediate execution plan (this push)

1. Implement **product hybrid router** + measure non-oracle (feature rules, not ref peeking)  
2. **WMT decode push**: higher beams, dual-system sentence-level ensemble upper bound  
3. Write measured gaps into `reports/COMPETITIVE_PUSH.md`  
4. Keep law pin D1D38A fixed  

Optional next session: WMT finetune loop under densify law wrap.

---

## 7. Measured this push (2026-07-22 competitive run)

| Track | Result | Notes |
|-------|-------:|-------|
| Product hybrid chat sacreBLEU | **40.64** | No ref peeking; densify-heavy (2800) + neural CJK (400) |
| WMT opus-mt beams=8 | **33.79** | Slightly under prior beams=5 33.88 |
| WMT NLLB-600M beams=8 | **33.49** | |
| **WMT oracle dual ensemble** | **37.61** | Best of opus/nllb per sentence (upper bound) |
| Gap ensemble → mid 40 | **2.39** | Was ~6.1 for single student |
| Gap ensemble → stretch 48 | **10.39** | Still needs finetune / larger model |

**Insight:** Beams alone do not close news. **Dual-student ensemble** recovers ~+3.7 sacre toward bar 40 without finetune. Remaining ~2.4 points need quality model upgrade or WMT finetune.

---

## 8. Beat levers run (2026-07-22)

| Lever | Result | Clears bar? |
|-------|-------:|:------------|
| L2 Neural-first hybrid chat | **48.74** sacre | **Yes** (≥45) |
| L1 Product NLL ensemble WMT | **34.54** | No (gap 5.46 to 40) |
| L1 Oracle ensemble upper | **37.13** | No (gap 2.87) |
| L3 Aggressive FT | 32.2 | No (regressed) |
| L3b Safe FT freeze-enc | 33.86 | No (flat) |
| Base opus-mt-de-en | 33.88 | No |

**Still required to beat news mid-parity (40):** better ensemble selection (~2.6 pts headroom to oracle), multi-epoch WMT FT with validation early-stop, or larger NLLB.

---

## 9. News push v3 (2026-07-22)

| System | sacreBLEU |
|--------|----------:|
| Product cross-NLL ensemble | **34.11** |
| Base opus | 33.88 |
| FT v3 test / best val | 33.41 / 35.76 |
| Oracle upper | 37.13 |
| Gap product → 40 | **5.89** |

Chat neural-first hybrid remains **48.74** (mid bar met).