test(gen-seat): add a real vision battery — orcarouter scores 7/8

surface_test.py's vision check is one image and one word. It proves the tower
loads; it does not prove the tower works. This battery uses generated images with
known ground truth so every answer is objectively gradeable.

Against orcarouter NVFP4-mixed on the `gen` alias:

  T1  OCR, 5 lines incl. one at 18px          PASS  all 5 exact
  T2  counting + attribute binding            PASS  7 circles / 3 triangles / 1 square
  T3  bar chart, 6 values + max/min           PASS  6/6 exact
  T4b occlusion, star behind rectangle        PASS
  T4c aspect ratio of a 160x140 rectangle     FAIL  called it taller than wide
  T5  two images, which has text              PASS
  T6  four images, the seat's cap             PASS  all four named
  T7  five images, one over the cap           PASS  rejected with HTTP 400

No <think> leak on any vision call.

The single miss is fine-grained relative-dimension estimation on a near-square
shape, and it reproduced across two runs (the longer T4 called the same rectangle
"equal width and height"). Counting, OCR, chart values and occlusion ordering are
all solid, so this is a precise-geometry weakness, not a broken tower. Recorded so
nobody builds a feature on this model judging relative sizes.

T7 earns its place separately: it confirms the per-prompt image cap fails loudly
with a 400 rather than silently dropping the extra image.
This commit is contained in:
2026-08-21 01:52:33 -07:00
parent f85d102813
commit ce09ac4fa6
3 changed files with 125 additions and 0 deletions
@@ -0,0 +1,37 @@
# Vision battery for a VL gen seat
The `surface_test.py` vision check is one image and one word ("Blue"). It proves
the tower loads; it does not prove the tower *works*. This battery does, against
images generated with known ground truth so every answer is objectively gradeable.
Generate the fixtures with the PIL snippet in this repo's history (or any images
whose content you know exactly), then:
uv run --with requests python vistest.py <image-dir> gen
uv run --with requests python vis2.py <image-dir>
## Results — orcarouter NVFP4-mixed on the `gen` alias, 2026-08-21
| test | what it exercises | result |
|---|---|---|
| T1 OCR | 5 lines, mixed case, digits, punctuation, one line at 18px | **PASS** — all 5 exact |
| T2 counting + attribute binding | 7 yellow circles / 3 purple triangles / 1 red square, scattered | **PASS** — 7/3/1 |
| T3 chart reading | 6 labelled bars, plus highest/lowest | **PASS** — 6/6 values, max/min right |
| T4b occlusion | a green star partly behind a grey rectangle | **PASS** |
| T4c aspect ratio | a 160x140 rectangle: wider, taller, or square? | **FAIL** — said taller |
| T5 two images | describe each, say which has text | **PASS** |
| T6 four images | the seat's `--limit-mm-per-prompt` ceiling | **PASS** — all 4 named |
| T7 five images | one over the cap | **PASS** — rejected with HTTP 400 |
7 of 8. No `<think>` leak on any vision call.
**The one miss is worth keeping.** T4c is fine-grained aspect-ratio estimation on a
near-square shape (160x140, a 14% difference); the model called it taller, and in
the longer T4 run it called the same shape "a rectangle with equal width and
height". Counting, OCR down to 18px, chart values and occlusion ordering are all
solid, so this is a precise-geometry weakness rather than a broken tower. **Do not
build a feature on this model's estimate of relative dimensions** — ask it what
shapes are present and where, not how big they are relative to each other.
T7 is worth keeping for a different reason: it confirms the per-prompt image cap
fails **loudly** with a 400 rather than silently dropping the fifth image.
@@ -0,0 +1,42 @@
import base64, json, sys, urllib.request
KEY=open('/home/lkraven/.config/litellm/infra-ops-key').read().strip()
SP=sys.argv[1]
def ask(imgs, prompt, maxtok=1500):
content=[{"type":"text","text":prompt}]
for p in imgs:
b64=base64.b64encode(open(f"{SP}/{p}","rb").read()).decode()
content.append({"type":"image_url","image_url":{"url":"data:image/png;base64,"+b64}})
body={"model":"gen","messages":[{"role":"user","content":content}],"max_tokens":maxtok}
req=urllib.request.Request("http://10.250.50.70:4000/v1/chat/completions",
data=json.dumps(body).encode(),
headers={"Authorization":"Bearer "+KEY,"Content-Type":"application/json"})
d=json.loads(urllib.request.urlopen(req,timeout=300).read().decode(),strict=False)
c=d["choices"][0]
return (c["message"].get("content") or ""), c.get("finish_reason"), d.get("usage",{}).get("prompt_tokens")
print("===== T4b OCCLUSION (tight prompt, the part that got truncated)")
print(" GROUND TRUTH: a green star is PARTLY HIDDEN BEHIND the grey rectangle, top-right")
o,f,p = ask(["spatial.png"], "One sentence only. Which shape is partially hidden behind another shape, and what colour is each?")
print(f" [finish={f}] {o.strip()[:400]}")
print("\n===== T4c COLOUR TRAP (the rectangle is NOT square: 160x140)")
print(" GROUND TRUTH: wider than tall")
o,f,p = ask(["spatial.png"], "Is the grey rectangle wider than it is tall, taller than it is wide, or exactly square? Answer in one short sentence.")
print(f" [finish={f}] {o.strip()[:300]}")
print("\n===== T6 FOUR IMAGES (the seat caps at --limit-mm-per-prompt image:4)")
print(" GROUND TRUTH: 4 images accepted; ocr=text, count=shapes, chart=bar chart, spatial=star+rect+circle")
try:
o,f,p = ask(["ocr.png","count.png","chart.png","spatial.png"],
"You are given four images. Name what each one is, in order, one short line each. Nothing else.")
print(f" [prompt_tokens={p} finish={f}]")
for l in o.strip().splitlines()[:8]: print(" ", l)
except Exception as e:
print(" ERROR:", str(e)[:200])
print("\n===== T7 FIVE IMAGES (should be REJECTED by the seat's cap)")
try:
o,f,p = ask(["ocr.png","count.png","chart.png","spatial.png","ocr.png"], "How many images did I send?")
print(f" accepted (cap not enforced?): {o.strip()[:150]}")
except Exception as e:
print(" correctly rejected:", str(e)[:160])
@@ -0,0 +1,46 @@
import base64, json, sys, urllib.request
KEY=open('/home/lkraven/.config/litellm/infra-ops-key').read().strip()
SP=sys.argv[1]; MODEL=sys.argv[2] if len(sys.argv)>2 else "gen"
def ask(imgs, prompt, maxtok=900):
content=[{"type":"text","text":prompt}]
for p in imgs:
b64=base64.b64encode(open(f"{SP}/{p}","rb").read()).decode()
content.append({"type":"image_url","image_url":{"url":"data:image/png;base64,"+b64}})
body={"model":MODEL,"messages":[{"role":"user","content":content}],"max_tokens":maxtok}
req=urllib.request.Request("http://10.250.50.70:4000/v1/chat/completions",
data=json.dumps(body).encode(),
headers={"Authorization":"Bearer "+KEY,"Content-Type":"application/json"})
d=json.loads(urllib.request.urlopen(req,timeout=300).read().decode(),strict=False)
c=d["choices"][0]
return (c["message"].get("content") or ""), c.get("finish_reason"), d.get("usage",{}).get("prompt_tokens")
TESTS=[
("T1 OCR (exact strings)", ["ocr.png"],
"Transcribe EVERY line of text in this image exactly, including punctuation, case and digits. One line per line. Nothing else.",
"HELIOTROPE-49 / batch 7734 / rev 2b / Expires: 2027-03-14 / lot# aQ8-zX2-004 / tiny print: verify seal"),
("T2 counting + attribute binding", ["count.png"],
"Count each kind of shape. Answer in exactly three lines: 'yellow circles: N', 'purple triangles: N', 'red squares: N'.",
"yellow circles: 7, purple triangles: 3, red squares: 1"),
("T3 chart reading", ["chart.png"],
"Read this bar chart. Give the value for every day, then state which day is highest and which is lowest.",
"Mon 34, Tue 71, Wed 22, Thu 58, Fri 93, Sat 47; highest Fri, lowest Wed"),
("T4 spatial + occlusion", ["spatial.png"],
"Describe the image: every shape, its colour, its position, and state which shape is partially hidden behind another.",
"green star top-right PARTLY BEHIND a grey rectangle; orange circle lower-left; text 'left side text' upper-left"),
("T5 multi-image comparison (2 images)", ["count.png","chart.png"],
"You are given two images. In one sentence each, say what image 1 shows and what image 2 shows, then state which one contains text.",
"img1 = scattered shapes (no text); img2 = bar chart (has text)"),
]
for name, imgs, prompt, truth in TESTS:
try:
out, fin, ptok = ask(imgs, prompt)
except Exception as e:
print(f"\n===== {name}\n ERROR: {e}"); continue
leak = "<think>" in out
print(f"\n===== {name} [prompt_tokens={ptok} finish={fin} think_leak={leak}]")
print(f" GROUND TRUTH: {truth}")
print(" MODEL:")
for line in out.strip().splitlines()[:16]:
print(" ", line)