# Épée v0.3 · Signer-Robustness Benchmark

**What it shows:** signer variation is the core problem in sign-language AI, and the fix is
*more native signers and more data*, not a bigger model. Re-measured on the six v0.3 signers
with the exact v0.2 protocol, so the two releases are directly comparable.

The model is trained on the six signers in this release (ALPHA through FOXTROT) and tested on
**two native Deaf signers with no clips in this release** (Signer A and Signer B, pseudonymized,
from the broader CLERC corpus). The held-out signers recorded under the same protocol and share
prompts with the release signers. That is by design, not leakage: same content, brand-new body,
which is exactly the controlled comparison a signer-generalization test needs. No clip and no
recording is shared between train and test.

![benchmark](benchmark.png)

---

## Setup

| | |
|---|---|
| Model | small BiLSTM (1 layer, hidden 64, bidirectional, mean-pool → linear) |
| Input | keypoint sequences, hands + arms (48 pts, x/y), shoulder-normalized, resampled to 24 frames |
| Task | sign (gloss) classification on the **same 24 glosses** as the v0.2 measurement (vocabulary frozen, not recomputed) |
| Train | 1,393 gloss segments from the six release signers (858 in v0.2) |
| Test | Signer A: 1,961 segments · Signer B: 1,453 segments |
| Chance | 4.2% (1 / 24) · **8 seeds** per configuration, mean ± std |
| Metrics | top-1 accuracy **and macro-F1** (the task is imbalanced; macro-F1 is the honest headline) |

Everything pinned against v0.2: vocabulary and class order, held-out signers and their exact
test segments, model, features, seeds, aggregation. Only the training pool changed.

---

## Result 1 · More signers → better on a brand-new signer

Tested on Signer A, never seen in training. Mean over all signer combinations at each size,
as in v0.2 (v0.2 values in grey):

| # training signers | Accuracy | Macro-F1 | v0.2 accuracy |
|---|---|---|---|
| 1 | 29% ± 15 | 0.17 | 22% |
| 2 | 47% ± 12 | 0.30 | 40% |
| 3 | 57% ± 10 | 0.39 | 50% |
| 4 | 63% ± 8 | 0.47 | 59% |
| 5 | 65% ± 7 | 0.51 | - |
| **6** | **69% ± 5** | **0.57** | - |

The curve keeps rising through six signers, and the run-to-run spread narrows from ±15 to ±5
along the way. A signer the model *has* seen scores 73% (ceiling): **the stranger gap, 43 points
at one training signer, is down to 3 points at six.** Confirmed on Signer B: 58% (54% in v0.2).

Every point also beats its v0.2 counterpart at the same signer count (29 vs 22, 47 vs 40, 57 vs
50, 63 vs 59). Same protocol, so the lift is the data: 200 clips per signer instead of 150, and
a wider phrase grid.

## Result 2 · More data → keeps climbing, not saturated

Same held-out signer; training set grown by a clean random fraction of the six-signer pool:

| Training examples | 80 | 160 | 260 | 380 | 520 | 680 | 858 | 1,044 | 1,393 |
|---|---|---|---|---|---|---|---|---|---|
| Accuracy | 22% | 34% | 41% | 53% | 55% | 68% | 64% | 63% | **71%** |

From 22% to 71% across the pool, above the v0.2 curve at almost every budget, and still rising
at the largest size. The dip around 858 to 1,044 is within the ±5 run-to-run spread, not a trend.

---

## Takeaway

> Trained on one Deaf signer, the recognizer gets **29%** on a signer it has never seen. Trained
> on six, it reaches **69%**, three points under the 73% it scores on signers it knows. The
> stranger gap is closing, and it is the data doing it: more signers, more clips per signer. This
> is why CLERC builds the multi-signer data layer.

## Caveats

- 24-gloss shared subset (frozen from v0.2 for comparability); the full vocabulary will differ,
  but the *shape* (more signers/data → up) should hold.
- FOXTROT covers 18 of the 24 glosses: his prompt set is a different register (declarative
  sentences, not conversational phrases). On this conversational test signer, the five grid
  signers alone reach 70.5%, statistically indistinguishable from the full six at 69.4%. His
  contribution to the release is register and vocabulary breadth, not extra accuracy on this
  particular test. Combinations that include him at small k pull the low end of the spread down,
  which is part of why the k=1 std is wide.
- The 5→6 rise in the table is partly composition: at k=5, the mean averages mixes where FOXTROT
  replaces a grid signer (worst combo 58.4%). Read the trend, not any single step.
- Measures generalization across signer **bodies**, not across topics; cross-vocabulary
  generalization is a separate, harder test.
- Small BiLSTM by design; a bigger model does not close the 1-signer-vs-6-signer gap, which is a
  data property.
- A correction to the v0.2 card's wording: the held-out signers do share prompts with the release
  signers (same recording protocol). "No shared clips" is the true and relevant guarantee; the
  earlier "no shared phrases" claim was inaccurate and is retired.