VTL Spatial — Robotics Success Verifier

The dated cards below are the chronological trail and can differ in tone; this block is the authoritative "where we are."

proven A deterministic two-view-parallax cross-check discriminates real 3D objects from flat prints on real pixels, across many object types, within a stated envelope. A real frontier VLM (Gemini Robotics-ER) is empirically fooled by flat decoys the instrument catches — 10 cases, some where it names the flatness and passes anyway. The synthetic mechanism is pre-registered, scaled (1200 scenes), with independence + null protocols passing.

scoped The instrument measures height-above-plane and nothing else. Out of envelope: transparent objects, and objects too short to parallax (tape roll, box lying flat). The decoys are constructed adversarial cases — this shows the failure mode exists and geometry catches it, not how often it occurs in the wild.

the edge, precisely A placement-precision test (real cup, offset ladder) came back a clean negative — the VLM localizes to ~0.5cm and judges placement sharply, so the instrument gains nothing on geometry the VLM can see. That agreement is expected (embodied systems localize well) and it is the calibration half of a working detector: a label-quality filter needs both agreement-when-fine (low false alarms — placement) and disagreement-when-wrong (the decoy catch). The instrument disagrees selectively, only when perception is fooled — which is what makes its disagreements trustworthy. The edge is perceptual independence, not out-measuring geometry.

open (1) Real-world value on natural silent failures in real VLA data — the central open question (the collaboration). (2) σ(Ω) temporal axis — needs video. (3) Two arguments still to write for the paper: the learned-judge counterargument (information limit + un-gameability, as two distinct claims) and the falsifiability frame (an information theorem plus a falsifiable monotonic-robustness claim).

new The cup shoot happened. Two-view parallax on an ordinary phone (two photos per scene, ~10 cm apart), 7 real scenes, one per class, ground truth read straight from the filenames. The synthetic mechanism reproduces on real pixels: the flat printed decoy reads flat, the real mug reads raised, and the raised decoy (on a box) fools the single depth surface — exactly as predicted.

Real mug → RAISED. The object (orange) shifts ~46px more than the table plane (blue) between the two camera positions, and reveals its side wall. That extra shift is real height.

Flat paper cup → FLAT. The object tracks the table plane to within ~9px and looks identical in both frames. A flat thing cannot fake the shift — the whole mechanism, in two pictures.

suggestive only The clear glass read raised (+19px, but only 4px over threshold and sign-flipped vs everything else — the shaky one). Tempting to call it "stereo beats active depth on transparency," but that is one anomalous data point, and batch-2 contradicts it: there the glass mostly detect-failed. Honest status — transparency is the instrument's hard case, not a win. Left in only as a hypothesis that would need its own test.

honest Caveats kept in view: n=1 per class — this proves the mechanism reproduces on real pixels, it is not an accuracy statistic. The glass is the shaky one (+19, only 4px over threshold, sign flipped vs the others — low-contrast object, harder to locate). Flat objects read ~−10, not exactly 0: that is the handheld noise floor. Full statistical set (≈6–8 pairs/class) is the next optional step; the mechanism and the hero image are already in hand. Reproduce: python3 -m benchmark.realframe_parallax · overlays in realframes/.

same day · honest Real-pixel VLM head-to-head. Ran the real Gemini Robotics-ER as an RGB-only judge on the same 7 frames (one view each, neutral presence prompt — not told decoys exist), lined up against parallax. python3 -m benchmark.realframe_vlm.

null this batch Gemini and parallax agreed on all 7, so this run does not demonstrate parallax's advantage. Reason: the flat cutout has a visible paper edge, so it is catchable on appearance — it is not an information-limit decoy (the synthetic one was RGB-identical by construction; this real one is not).

striking The raised decoy fooled both the frontier VLM and the 2-surface instrument on real pixels — the exact humility case that needs touch. fix for full set the flat decoy must be a borderless, matte, photo-quality print laid flush so appearance says "real cup" while geometry says "flat" — that disagreement is the hero the paper is built around.

same day · envelope Multi-object envelope test (Cups2). Pushed the tool to 19 object classes, ~113 scene-pairs — real mugs, a white paper cup, a Scotch tape box and roll, and a zoo of prints (cut-outs, full-page, flattened-to-look-3D) — to map where the instrument works and where it doesn't. Ground truth from filenames. Rebuilt the object-finder to locate objects by their edges (the white table + low-contrast paper prints broke the old brightness method); bare table now correctly reads no-object, and the low-contrast white cup reads raised 6/6. python3 -m benchmark.realframe_parallax Cups2.

The separation, honestly. Flat prints (green) cluster low, real 3D objects (red) high, with a small overlap band (amber) where curled cut-outs and short objects live. The out-of-envelope glass/short objects (grey) scatter across the whole axis — no clean read — which is exactly why they are excluded, not hidden.

out of envelope Named, not hidden: transparent glass (see-through defeats a centroid read) and short objects — the tape roll and the box lying flat are ~1–2 cm tall, so their real parallax sits at the ~10px noise floor. The same box standing reads RAISED cleanly. Little height = little parallax; a physical limit, not a photo flaw. This is the operating-envelope discipline: a narrow, deterministic instrument that states its wall rather than pretending it has none.

standing Across a real variety of objects, the tool separates flat prints (5–10px) from real 3D (26–67px) with a clean gap, and catches the built-to-look-3D flat as flat. The 3 in-envelope misses were 2 curled cut-outs (a lifted edge reads raised) and 1 borderline. Verdict: findings look solid inside the stated envelope; next is the Gemini head-to-head on the strong flat decoys.

the hero split · real pixels Gemini head-to-head, better decoys. Curated set — strong flat prints (full-page, scotch-box-flat, fake-3d) + real controls + empty — one RGB frame each to Gemini Robotics-ER, neutral object-named prompt, vs the parallax instrument. This is the test batch-1 failed (that flat cut-out was too obviously paper). python3 -m benchmark.realframe_vlm Cups2.

The discriminator: fake-3d. A flattened box drawn to read as a 3D box. Gemini: "a box of Scotch tape is present." The instrument: object tracks the table plane → FLAT. Appearance fooled, geometry not.

the mechanism, exactly On two of the ten, Gemini explicitly names the flatness / the paper and still returns PASS. That is the thesis in the raw: an appearance judge conflates looks-like-X with X-is-really-there. The parallax surface does not — a flat thing cannot fake the shift, so it read FLAT on all 18 prints and RAISED on all 9 real objects.

honest nuance Not every print fooled it: the rounded mug cut-out was caught ("a flat sticker or image of a mug"), and the full-page mug only 1/3. A distinctly 3D-shaped object is hard to fake in print; flat-faced product prints (the tape box) and the fake-3d fooled it every time (3/3). So decoy quality matters, and it is now empirically mapped. Standing: the real-pixel advantage — unproven after batch 1's null — is demonstrated, within the stated envelope. A skeptic's fair caveat remains: these are constructed adversarial decoys, so this shows the failure mode exists and geometry catches it, not its frequency in the wild (that needs real VLA data).

validation note Honest bound on the "27/27": it used the first 3 scenes/class (by filename, not by result). On the full 6/class set the flat and raised bands overlap ~13–20px — curled cut-outs read falsely high, short/borderline real objects read low. The clean, threshold-robust heroes are fake-3d (1.9–5.1px, Gemini fooled 3/3) and scotch-box-flat (5.3–12.3px, 3/3); lead with those and their margins, not a blanket score. The overlap band is the same envelope boundary named above, not a new problem.

negative result · kept Placement precision — the test that should have favored us, and didn't. The decoy tests presence; this tests precision (the "off by millimeters" failure). A real cup at a ladder of offsets from a green target; the instrument reports the exact gap + correction vector; Gemini judges "on target?" and points to cup and target. Two shoots: a gross-miss batch (8–32cm) and a near-miss batch (0.6–28cm, after a circle-fit fix that recovers the target centre when the cup occludes it). python3 -m benchmark.placement_precision Placement2.

The instrument works. Green target centre (circle-fit, occlusion-robust), mug centre, correction vector, dots. It turns a placement into an exact centimetre gap and a move-this-way vector, deterministically.

the finding Gemini is not fooled here. Its verdict boundary is sharp — PASS at the dead-on 0.6cm, FAIL from 2.2cm out — with zero lenient passes and zero self-inconsistency. And it localizes as precisely as the instrument measures:

Gemini's implied gap (from its own two points) tracks the measured gap to a mean of 0.55cm. There is no measurement-vs-approximation gap on placement — a frontier VLM sees geometry it can perceive just fine.

dialectic — I tried to break the null Flipped it every way: (a) sub-cm resolution — Gemini's ~0.5cm (max 2.2cm) pointing noise means the instrument's pixel-precise determinism could win for sub-centimetre-tolerance tasks, but that produced no wrong verdict here, so it is a hypothesis, not a result; (b) prompt sensitivity — the strict "centred-on-target" prompt drove Gemini strict; a looser prompt might surface leniency (real caveat, not chased); (c) on scene 00 Gemini's points said on-target while its verdict correctly said FAIL — its holistic judgment beat its own localization, which is pro-Gemini, not anti. Nothing overturns the null.

why the negative is an asset Two real-pixel tests together: on placement (geometry it can see) the VLM is good, instrument gains nothing; on decoys (perception fooled) the VLM is caught 10/18, instrument catches all. That is the signature of a working detector: agreement-when-fine (low false alarms — placement) plus disagreement-when-wrong (the decoy catch). The agreement is the calibration half, not a loss; it is what makes the disagreements trustworthy. So the edge is perceptual independence — catching when the policy's perception is wrong — not out-measuring on geometry. Consistent with ProbeAct (spatial signal recoverable, R²=0.968; spatial is not where learned systems fail). Running the test that should have favoured us, finding it doesn't, and keeping it, is what makes the decoy result credible.

investigation · mostly negative Do the composition kernel's other dimensions do robotic work? Prompted by review. First I got it wrong — called six of the eight "degenerate." Correction: they are a live composition fingerprint (on compositional scenes r_v spreads 0.62→0.89, ρ_r 12→16.6, μ ~10×, x_p 0.48→0.73; the four composition modes are literally defined by void + peripheral). They only go quiet on clean single-object robotic scenes, because there is no arrangement to measure and real-vs-flat is a depth question a 2D metric can't answer. So the robotic verdict rests on displacement + the parallax depth surface; the rest is not dead, just off-task here.

three probes (1) 2D→3D via frame-to-frame relationship change (the sharp idea — does the full fingerprint's change between the two views carry 3D beyond the centroid shift?): directionally yes, but a whisper. Real objects change more in delta_y (5.3×) and SDI (2.1×) between views — they reveal vertical structure a flat print doesn't — but magnitudes are tiny (n=3), a faint echo of the displacement signal we already use, not a hidden channel. (2) Reinforcer (do dims cross-confirm placement?): no — x_p/SDI wander with position, not monotonically with the gap; too muted on a white single-object scene. (3) Material (μ, d_s): partial — μ flags the transparent glass (~5× the gradient of opaque objects), so μ is faintly material-sensitive as hypothesised, but only isolates the outlier; d_s is dead-flat across every material.

standing: open, not closed These probes used the wrong stimulus for composition metrics (clean, single-object), by design — so usage is unresolved, not disproven. Nothing here is report-worthy for the current experiments, and it is correctly left out of the paper's results. The genuine untapped thread is a different task family — scene-level outcomes (clearance via r_v, consolidation via ρ_r, alignment via θ) on dense multi-object scenes, verified as a fingerprint delta before/after a manipulation. Named as future scope; not claimed.

Robot-learning made its judge learnable along with its policy. A judge a policy can fool is a structural flaw. We bring an un-learnable judge, and the parallax twist: success is the agreement of N genuinely independent reads; their disagreement (Ω) is the honest uncertainty.

Ω (omega) = mean pairwise distance between the N reads' object centroids, in normalized image coordinates. Ω→0: the surfaces agree (coherent, trustworthy). Ω large: they disagree (contested — trust no single read). Identical construct to the Parallax program's central coordinate.

proven The flat-decoy blind spot is real in a frontier VLM and a deterministic guard closes it. Independence + null protocols pass. The mechanism is robust to n (1200, tight CIs) and pre-registered.

scoped This is a mechanism / existence result on synthetic scenes, not a real-world frequency or superiority claim. Aggregate accuracy is class-mix-dependent and not the headline.

first pass done (1) Realism — real photographed scenes with a literally printed decoy are now shot and analyzed: 7/7 agreement, flat decoy caught on real pixels (see the 2026-08-13 update at top). Still open: the full statistical set (≈6–8 pairs/class). open (2) σ(Ω) / temporal axis — the program's strongest signal ("distribution beats mean") needs video, not single frames.

closed Third observable — proprioception (grasp). Adding the canonical robotics touch surface takes the verifier from 70% → 100% on the adversarial benchmark, 40 → 0 false passes. It rescues transparent/dropout (real objects the depth camera was blind to — touch still feels them) and catches raised_decoy (a box has the wrong grasp profile). Direct confirmation of the N-independence thesis; see THIRD_SURFACE.md.

built LiDAR-free geometric surface (vtl_spatial/stereo.py): two-view parallax recovers the independent depth read from two ordinary iPhone photos — no depth sensor needed. Synthetic test: raised object → +14px relative parallax (above plane); flat decoy → 0px (at plane). Capture protocol written (CAPTURE_PROTOCOL.md); realframes ingestion is the next run.