Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
80 commits
Select commit Hold shift + click to select a range
b358af2
formal: a machine-checked theory of inverting box-filter rasterisation
claude Oct 10, 2026
6eb7801
trace: sub-pixel boundaries from exact column sums; the seam of a clo…
claude Oct 10, 2026
7729caf
docs: theory page -- inverting the box filter, proved, surveyed and m…
claude Oct 10, 2026
0c9e272
strokes: on by default, with merges and the faces that leave with a s…
claude Oct 10, 2026
35c0f1d
docs: the optimal vectorizer, informally, with a search-or-model diag…
claude Oct 10, 2026
6b3ab27
docs: two chains from pixels to the fewest coordinates (boundary, rep…
claude Oct 10, 2026
32169b7
gate: geometric match to the artist's file (geom)
claude Oct 10, 2026
eb38a57
changelog: the geometric match axis
claude Oct 10, 2026
7062834
docs: the boundary chain, informally -- area windows as the likelihoo…
claude Oct 10, 2026
e425b90
docs: the representation chain, informally -- the posterior of the de…
claude Oct 10, 2026
79c916a
Merge branch 'worktree-agent-a2797ed10afe7623e' into ccr-73a29992-0f58d2
claude Oct 10, 2026
f766a5c
docs: the negotiated chains, and keeping proofs and code from driftin…
claude Oct 10, 2026
3131d26
verified: checkers generated from the Lean definitions the theorems a…
claude Oct 10, 2026
08e6ab0
layers: each face completed behind what is painted over it (R4.1), ce…
claude Oct 10, 2026
c5fed8f
Merge branch 'ccr-73a29992-0f58d2' into worktree-agent-a514b1b680f4e6683
claude Oct 10, 2026
0757571
evidence: window likelihood of any description, its generated checker…
claude Oct 10, 2026
318b083
quality: budget tightened after the evidence milestone (test function…
claude Oct 10, 2026
81307fe
layers: painter's interval in Lean, span checker generated into inkve…
claude Oct 10, 2026
9789ce2
gate: a noisy condition, the screen set as a resized quality-80 JPEG
claude Oct 10, 2026
c128704
Merge boundary milestone 1; chain certificates for the painter-interv…
claude Oct 10, 2026
42d709c
Merge branch 'ccr-73a29992-0f58d2' into worktree-agent-a514b1b680f4e6683
claude Oct 10, 2026
3f38c65
gate: baseline for the noisy condition; where a real input's noise sits
claude Oct 10, 2026
d1e989f
gate: the web conditions themselves; noise theory page; residual nois…
claude Oct 10, 2026
5c9d102
layers: corners a blur cut off are restored under the cover; noise mo…
claude Oct 10, 2026
7e06623
Merge branch 'ccr-73a29992-0f58d2' into worktree-agent-a514b1b680f4e6683
claude Oct 10, 2026
c3ae3e4
evidence: noise measured at the edges, the blur forward model, the fi…
claude Oct 10, 2026
4ec9d47
layers: the seam a narrower stroke leaves is shown; rectangles on the…
claude Oct 10, 2026
bcdf2c8
Merge branch 'worktree-agent-a2797ed10afe7623e' into worktree-agent-a…
claude Oct 10, 2026
b8f4634
fast: a JPEG's halos go back to the inks they lie between (lossy inta…
claude Oct 10, 2026
e6b1737
evidence: the strip side test withheld; clean and web output byte-ide…
claude Oct 10, 2026
4fa1298
Merge branch 'worktree-agent-a2797ed10afe7623e' into worktree-agent-a…
claude Oct 10, 2026
5850de5
Merge boundary chain milestone 2: per-image noise, the blur forward m…
claude Oct 10, 2026
95de3df
layers: a smooth bridge under the cover; the seam counted where the u…
claude Oct 10, 2026
d65bc5c
cli: --restore auto and --sr auto act only on input that shows damage…
claude Oct 10, 2026
e07519c
trace: the opaque finisher's forwarding wrapper removed; lib.rs back …
claude Oct 10, 2026
74a7b22
docs: real inks inside the soft same-ink floor, measured and not shipped
claude Oct 10, 2026
a15fec5
Merge branch 'ccr-73a29992-0f58d2' into worktree-agent-a514b1b680f4e6683
claude Oct 10, 2026
0454ee4
gate: turning and parameters judged as distances from the artist's file
claude Oct 10, 2026
129263a
Merge branch 'ccr-73a29992-0f58d2' into worktree-agent-a514b1b680f4e6683
claude Oct 10, 2026
a876257
bench/theory: human_gap.py, a change read as distance to the artist's…
claude Oct 10, 2026
0bc5b89
layers: no completion makes a face the page
claude Oct 10, 2026
84545aa
layers: the L completes on a page; alone it would become the page
claude Oct 10, 2026
f913432
cli: a probe --restore auto / --sr auto keeps reports the plain trace…
claude Oct 10, 2026
29cf4ad
fast: the halo pass's two clippy warnings
claude Oct 10, 2026
e8c2dfd
Merge the session branch: the boundary chain's milestone 2 and the ga…
claude Oct 10, 2026
878aaf5
Merge front-end noise work: Fast's JPEG halos, `auto` restore/sr only…
claude Oct 10, 2026
d8a5646
gate: fast-web baseline re-recorded after the halo pass
claude Oct 10, 2026
6663eb0
cli: auto asks for the sign of damage first, the probe's residual onl…
claude Oct 10, 2026
87d2bde
layers: the canvas is the image's frame, not the painted faces' box
claude Oct 10, 2026
c62af06
Merge branch 'ccr-73a29992-0f58d2' into worktree-agent-a514b1b680f4e6683
claude Oct 10, 2026
c39ac16
docs: item 2 by the distance gate, two narrowings that separate nothi…
claude Oct 10, 2026
10207a0
cli: --restore auto without a sign of compression makes no probe; Fas…
claude Oct 10, 2026
92f2f65
gate: faster with the same numbers; human statistics of how a trace i…
claude Oct 10, 2026
c53b7db
Merge benchmark work: a faster gate with the same numbers, and human …
claude Oct 10, 2026
0982429
gradient: on a lossy intake the carve stage leaves the codec's band a…
claude Oct 10, 2026
0f2f21a
planar: the strip side test scales with the intake's noise on a soft …
claude Oct 10, 2026
4f6999f
docs: completion milestone, the gate and the amodal evaluation
claude Oct 11, 2026
44776e1
Merge branch 'ccr-73a29992-0f58d2' into worktree-agent-a514b1b680f4e6683
claude Oct 11, 2026
ca2892a
bench: design_stats.py absorbed into the human statistics battery
claude Oct 11, 2026
387212c
quality: on a lossy intake, the codec's variants, halos and thin vari…
claude Oct 11, 2026
0c28e5c
quality: a restored image skips the codec clean-up (ColorOptions::res…
claude Oct 11, 2026
b4abae7
Merge front-end noise work through the carve band: auto restore costs…
claude Oct 11, 2026
c1a28b3
Merge boundary chain: the strip reading's corner test scales with noi…
claude Oct 11, 2026
988fe21
Merge representation chain milestone 1: faces completed behind what c…
claude Oct 11, 2026
b8ebd30
cli: the completion stage's emitter glue in its own module; rustfmt o…
claude Oct 11, 2026
189f61e
fit: nodes at the curve's extrema, where the fit's own objective pref…
claude Oct 11, 2026
e2ee543
--editability: curves as drawing tools write them; G1 and node alignm…
claude Oct 11, 2026
ea49a44
Merge the session branch (b8ebd30): completion stage, carve band, sid…
claude Oct 11, 2026
2bbfc30
Merge the session branch at b8ebd30 under the Quality codec clean-up
claude Oct 11, 2026
67e7e46
seams: judge a cut against the lower face's own outline, not a coarse…
claude Oct 11, 2026
78fae36
trace: lib.rs back to its length budget after the merge
claude Oct 11, 2026
8441040
gate: Quality baselines re-recorded after the completion, side-test a…
claude Oct 11, 2026
7d55440
cli: --restore auto is the default in Quality mode
claude Oct 11, 2026
46aacb1
contract: bless the extrema and seam output; CHANGELOG: gate numbers …
claude Oct 11, 2026
0a7d00d
emit: counters cut out of their glyph, square rects in their paint's …
claude Oct 11, 2026
e8912a3
Merge front-end: Quality's JPEG clean-up, and --restore auto as Quali…
claude Oct 11, 2026
1f24602
Merge representation chain milestone 2: counters cut out of their gly…
claude Oct 11, 2026
3e4f0f3
Merge node placement: nodes at a curve's extrema, --editability as dr…
claude Oct 11, 2026
6d8f0c8
quality: budget tightened after the three merges (test functions 1189…
claude Oct 11, 2026
1323524
gate: baselines re-recorded after the front-end, encoder and node-pla…
claude Oct 11, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
2 changes: 1 addition & 1 deletion .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ jobs:
- name: gate's own tests
run: >-
python -m unittest bench/tests/test_gate_stats.py bench/tests/test_ci_gate.py
bench/tests/test_gate_scoring.py bench/tests/test_gt_diff.py
bench/tests/test_gate_scoring.py bench/tests/test_gt_diff.py bench/tests/test_design.py
bench/tests/test_cases_ratchet.py
- name: hard cases, Quality
run: python bench/cases.py --exe target/release/inkvec --ratchet quality
Expand Down
65 changes: 65 additions & 0 deletions .github/workflows/formal.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,65 @@
name: formal

# Machine-checks formal/InkvecTheory (Lean 4 + Mathlib): every proof, then the axiom audit,
# which must report only Lean's three standard axioms, then that the checkers generated from
# the Lean definitions (crates/inkvec-verified/src/generated) are exactly what those
# definitions print, so the Rust cannot drift from the proofs. Runs when either changes; the
# Mathlib build comes from its prebuilt cache (`lake exe cache get`).

on:
push:
branches: [main, master]
paths: ["formal/**", "crates/inkvec-verified/**", ".github/workflows/formal.yml"]
pull_request:
paths: ["formal/**", "crates/inkvec-verified/**", ".github/workflows/formal.yml"]

permissions:
contents: read

concurrency:
group: formal-${{ github.workflow }}-${{ github.ref }}
cancel-in-progress: true

jobs:
lean:
runs-on: ubuntu-latest
defaults:
run:
working-directory: formal/InkvecTheory
steps:
- uses: actions/checkout@v4
- name: Install elan
run: |
curl -sSfL https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh \
| sh -s -- -y --default-toolchain none
echo "$HOME/.elan/bin" >> "$GITHUB_PATH"
- name: Fetch prebuilt Mathlib
run: lake exe cache get
- name: Build every proof
run: lake build
- name: Axiom audit
run: |
lake build InkvecTheory.Audit 2>&1 | tee audit.log
n=$(grep -c "depends on axioms" audit.log || true)
expected=$(grep -c "^#print axioms" InkvecTheory/Audit.lean)
if [ "$n" -ne "$expected" ]; then
echo "the audit printed $n axiom reports, expected $expected" >&2
exit 1
fi
if grep "depends on axioms" audit.log | grep -v "\[propext, Classical.choice, Quot.sound\]"; then
echo "a theorem depends on an axiom beyond Lean's standard three" >&2
exit 1
fi
if grep -rnE "\bsorry\b|^axiom |\badmit\b" InkvecTheory --include=*.lean | grep -v "Audit.lean"; then
echo "sorry, admit or a new axiom in the development" >&2
exit 1
fi
- name: Generated checkers match the Lean definitions
run: |
lake env lean InkvecTheory/Gen/Emit.lean
git diff --exit-code -- ../../crates/inkvec-verified/src/generated
untracked=$(git ls-files --others --exclude-standard -- ../../crates/inkvec-verified/src/generated)
if [ -n "$untracked" ]; then
echo "generated files not committed: $untracked" >&2
exit 1
fi
3 changes: 2 additions & 1 deletion .github/workflows/release.yml
Original file line number Diff line number Diff line change
Expand Up @@ -226,7 +226,8 @@ jobs:
Logolabs/inkvec-denoiser-001 (https://huggingface.co/Logolabs/inkvec-denoiser-001).
2. Place it next to this binary, or in a `models/` directory next to this
binary, or point INKVEC_RESTORE_ONNX at its path.
3. Run inkvec with --restore auto (or on).
3. Quality mode (the default) then runs it as --restore auto by itself; in
fast or balanced mode, pass --restore auto (or on).

See the project README for the other restorer backends (restore-burn,
restore-wgpu), which are not part of this pre-built archive.
Expand Down
65 changes: 65 additions & 0 deletions CHANGELOG.md

Large diffs are not rendered by default.

33 changes: 26 additions & 7 deletions CONTRIBUTING.md
Original file line number Diff line number Diff line change
Expand Up @@ -59,14 +59,23 @@ the installer refuses to replace a pre-push hook it did not write (unless `--for
## The regression gate

`bench/ci_gate.py` scores the committed 246-icon screen set (`bench/data`, kept in the
repository so the gate runs from a bare checkout) under six conditions: Quality and Fast mode,
each at 128 px, 512 px with transparency and 512 px flattened onto white (an opaque logo).
repository so the gate runs from a bare checkout) under eight conditions: Quality and Fast
mode, each at 128 px, 512 px with transparency, 512 px flattened onto white (an opaque logo)
and `web` (the same logo resized to 400 px and saved as a quality-80 JPEG: the noise real
inputs carry; `bench/build_web_tier.py`).
Each condition is compared icon by icon with the per-platform baseline
`bench/gate/baselines/<os>-<arch>.json`, on three axes:
`bench/gate/baselines/<os>-<arch>.json`, on four axes:

- **dE00** (colour error against the artist's file), margin 1%
- **turning** (control-polygon turning per unit length, read command by command), margin 2%
- **ratio** (parameter count versus the artist's), margin 3%
- **turning_gap** (distance from the artist's control-polygon turning, per canvas side),
margin 2%
- **ratio_gap** (distance from the artist's parameter count, `|ln(ratio)|`), margin 3%

Turning and parameters are judged as distances from the artist's file, not as raw values:
a trace that turns more, or writes more numbers, toward what the artist drew is not a
regression. The raw values are still reported.
- **geom** (geometric match to the artist's file: the mean distance between the trace's
edges and the artist's, in input pixels, read as the area between them), margin 2%

The verdict is statistical: a paired bootstrap interval for the change of each axis, and a
non-inferiority test that passes when the one-sided 95% upper bound stays below the margin.
Expand All @@ -75,15 +84,25 @@ the margin is floored at the smallest effect the set can detect, and such a pass
reported as "within-noise"; there is no "inconclusive" verdict. Details and the
literature are in `bench/README.md`.

Run it against a release build (about 30 min for all six conditions on a desktop; pass
Run it against a release build (about 40 min for all eight conditions on a desktop; pass
`--conditions` for a subset while iterating):

```
cargo build --release -p inkvec-cli
python bench/ci_gate.py --exe target/release/inkvec --workers 4
```

(needs `numpy`, `pillow`, `scikit-image`, `resvg_py`). `python tools/prepush.py --gate`
(needs `numpy`, `scipy`, `pillow`, `scikit-image`, `resvg_py`). Icons whose SVG is
byte-identical to the baseline's are traced but not scored again, so a change that moves few
icons costs little more than the traces. The gate also reports, without gating them, how far
each trace is from the artist's file in how it is *drawn* (segment kinds, smooth joins, nodes
at the extrema, the design grid, ...); to tune a build or a flag against those statistics:

```
python bench/human_stats.py --exe target/release/inkvec --extra-args "--editability" --base-args ""
```

`python tools/prepush.py --gate`
runs it with the hard cases. Only a demonstrable gain moves a baseline. If a change
regresses an axis, the gate fails; it can only be bypassed with
`--exe ... --bypass-gate "<justification>"`, and only with a strong, explicit
Expand Down
6 changes: 6 additions & 0 deletions Cargo.lock

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

2 changes: 2 additions & 0 deletions Cargo.toml
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,7 @@
resolver = "2"
members = [
"crates/inkvec-core",
"crates/inkvec-verified",
"crates/inkvec-fit",
"crates/inkvec-trace",
"crates/inkvec-cli",
Expand Down Expand Up @@ -43,6 +44,7 @@ too_many_lines = "warn"
# Internal crates carry a version as well as a path: crates.io refuses a published crate
# whose dependency is a bare path.
inkvec-core = { path = "crates/inkvec-core", version = "0.2.7" }
inkvec-verified = { path = "crates/inkvec-verified", version = "0.2.7" }
inkvec-fit = { path = "crates/inkvec-fit", version = "0.2.7" }
inkvec-trace = { path = "crates/inkvec-trace", version = "0.2.7" }
inkvec-sr = { path = "crates/inkvec-sr", version = "0.2.7" }
Expand Down
4 changes: 2 additions & 2 deletions PIPELINE_EXPLANATION.md
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ Inkvec implements a dual-tier neural pre-processing architecture. These two neur
| **Extreme Clamping / Snapping** | **`SNAP_LEVELS = 6`** (`crates/inkvec-restore/src/lib.rs:101`) | Channel-wise affine **`clean::match_flats`** (`crates/inkvec-sr/src/clean.rs:206`) |
| **Stochastic Routing Fix** | N/A (deterministic feed-forward CNN) | **Pinned Seed `ROUTING_SEED = 0x5641_4331`** ("VAC1", `tools/inkvec_sr/model.py:40`) |
| **Damage Gating** | `inkvec_restore::decide` (`crates/inkvec-restore/src/lib.rs:568`) | `inkvec_sr::decide` (`crates/inkvec-sr/src/lib.rs:171`) |
| **CLI Flag** | `inkvec --restore <auto\|on\|off>` (default: `off`) | `inkvec --sr <auto\|on\|off>` (default: `off`) |
| **CLI Flag** | `inkvec --restore <auto\|on\|off>` (default: `auto` in Quality mode when the build can restore, else `off`) | `inkvec --sr <auto\|on\|off>` (default: `off`) |
| **Tracer Intake Interaction** | Forces tracer **`--lossy on`** (soft intake) when restoration triggers | Runs before intake normalization; downsampled to target scale |

> **Critical Architectural Clarification (The Arrhenius Attribution):**
Expand Down Expand Up @@ -885,7 +885,7 @@ Crucially, Inkvec derives the **honest posterior covariance** of every measureme
$$\sigma_{\text{pos}} = \frac{\sigma_\alpha}{\|\nabla \alpha\|} = \frac{\sigma_{\text{pixel}}}{\|F - B\| \cdot \|\nabla \alpha\|}$$

#### 3. Quantization Noise Floor (`sigma_floor`)
On inputs with discretized alpha channels (e.g. 16-level alpha), uniform quantization noise creates artificial gradient steps. The floor that uniform quantization over a unit interval would suggest is shown below; Inkvec does not apply it: `SIGMA_FLOOR` is zero, because coverage-derived positions can be more precise than whole pixel coordinates and a universal $1/\sqrt{12}$ px floor oversmooths clean small artwork (`crates/inkvec-trace/src/contour.rs:318-327`, applied on the planar path at `crates/inkvec-trace/src/planar.rs:1253-1255`):
On inputs with discretized alpha channels (e.g. 16-level alpha), uniform quantization noise creates artificial gradient steps. The floor that uniform quantization over a unit interval would suggest is shown below; Inkvec does not apply it: `SIGMA_FLOOR` is zero, because coverage-derived positions can be more precise than whole pixel coordinates and a universal $1/\sqrt{12}$ px floor oversmooths clean small artwork (`crates/inkvec-trace/src/contour.rs:318-327`, applied on the planar path at `crates/inkvec-trace/src/planar.rs:1214-1216`):

$$\sigma_{\text{floor}} \ge \frac{1}{\sqrt{12}} \approx 0.2887\text{ px}$$

Expand Down
6 changes: 3 additions & 3 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -117,7 +117,7 @@ Per-case coordinate count over the same 21 cases, from the same script and data

### Damaged input: JPEG, WebP, AI-decoder output

Inkvec ships an optional trained restorer that removes compression and decode damage before tracing. It is **off by default** (`--restore off`; `--restore auto` enables it only when the input looks damaged). The weights that ship — `restorer.onnx` from [`Logolabs/inkvec-denoiser-001`](https://huggingface.co/Logolabs/inkvec-denoiser-001), pulled on first use or via `tools/pull_model.py` — have **no published benchmark in this repository**, and no LPIPS or ablation figure for them is reported anywhere here. The figures that previously appeared in this section were measured during development on a different checkpoint and are not reproduced for the shipped weights; the VectorArk and StarVector numbers were those projects' own published results, measured under their protocols, not here.
Inkvec ships an optional trained restorer that removes compression and decode damage before tracing. In Quality mode it is **`auto` by default** in a build that carries it (`--restore auto`: it restores only when the input shows compression and its trace disagrees with it; a clean PNG pays the ringing score alone); Fast and balanced keep it off unless `--restore auto` or `on` is given, and a build without it traces as `--restore off`. The weights that ship — `restorer.onnx` from [`Logolabs/inkvec-denoiser-001`](https://huggingface.co/Logolabs/inkvec-denoiser-001), pulled on first use or via `tools/pull_model.py` — have **no published benchmark in this repository**, and no LPIPS or ablation figure for them is reported anywhere here. The figures that previously appeared in this section were measured during development on a different checkpoint and are not reproduced for the shipped weights; the VectorArk and StarVector numbers were those projects' own published results, measured under their protocols, not here.

The same pre-pass runs **in the browser**, on the same `restorer.onnx`: see [In the browser](#in-the-browser).

Expand Down Expand Up @@ -194,7 +194,7 @@ inkvec <input> [-o <output.svg>] [OPTIONS]
| Flag | Default | What it does |
|---|---|---|
| `--mode <quality\|fast\|balanced>` | quality | `quality` is the full engine: best fidelity, fewest parameters. `fast` is a Potrace-class fit on the same planar map, with flat fills: several times faster, a little less faithful (see [Speed](#speed-quality-and-fast)). `balanced` is `fast` plus eight iterations of Quality's boundary solve and a fit a quarter finer, on inputs up to 1024 px on the longer side (larger ones trace as `fast`). Options that only steer Quality stages are ignored in Fast and Balanced, and the report names them. |
| `--restore <auto\|on\|off>` | off | Trained-network cleanup for JPEG/WebP/AI-decoder damage. `auto` only restores if it looks damaged, and traces directly when no restorer is available (`on` is an error then). |
| `--restore <auto\|on\|off>` | auto in Quality mode when the build can restore, else off | Trained-network cleanup for JPEG/WebP/AI-decoder damage. `auto` only restores if it looks damaged, and traces directly when no restorer is available (`on` is an error then). |
| `--sr <auto\|on\|off>` | off | Super-resolution pre-pass (2-4× upscale), complementary to `--restore`. |
| `--lossy <auto\|on\|off>` | auto | Whether to trust the file as clean or trace with noise-aware intake. |
| `--max-dim <px>` | 2048 | Cap on the longer side; SVG is written at the original size. |
Expand Down Expand Up @@ -381,7 +381,7 @@ Full stage reference: [`docs/algorithm/`](docs/algorithm/) (start at `docs/algor

## Benchmark and CI gate

`bench/` scores against a source SVG with CIEDE2000, DISTS, LPIPS, DINOv2/v3, parameter ratio, bytes and time. CI runs `cargo test` on Linux, Windows and macOS, then `bench/ci_gate.py` against the 246-icon screen set in Quality and Fast mode at 128 and 512 px, failing unless a paired bootstrap shows colour error stays within 1%, anchor turning within 2% and the parameter ratio within 3% of the per-platform baseline.
`bench/` scores against a source SVG with CIEDE2000, DISTS, LPIPS, DINOv2/v3, parameter ratio, bytes and time. CI runs `cargo test` on Linux, Windows and macOS, then `bench/ci_gate.py` against the 246-icon screen set in Quality and Fast mode at 128 and 512 px and as a resized, quality-80 JPEG (the noise real inputs carry), failing unless a paired bootstrap shows colour error stays within 1%, the distance from the artist's control-polygon turning within 2%, the distance from the artist's parameter count within 3% and the geometric match to the artist's file (mean edge displacement, by area) within 2% of the per-platform baseline.

---

Expand Down
Loading
Loading