docs(wiki): a signpost Home page and a curated sidebar
Home was still the auto-generated one-line stub, which reads as abandoned to
the one visitor the wiki actually gets: someone already on the repo who
clicked the Wiki tab.
It cannot be deleted - GitHub recreates it as the wiki's landing page - so
the choice was only ever what it says. It now carries: what this wiki is and
how it relates to gps.csutil.com/blog, an explicit drafts warning (GitHub
hides the status field, so a visitor cannot otherwise tell these are drafts),
four entry points chosen by reader intent rather than by topic, and where to
send a correction.
_Sidebar.md puts six themed groups on every page. Deliberately a subset -
eighteen of thirty - since GitHub's own page list already enumerates
everything; what it cannot do is say where to start.
Sized at one screen each on purpose: GitHub serves this wiki with
x-robots-tag: none, so it earns no search traffic and is worth investing in
only for readers who are already here. Search indexing needs 500+ stars per
GitHub's documented rule; this repo has 2.
b7dffb2
fix(blog): links, edit seams, and errors in my own corrections
Three cold reviews. The largest finding was that all thirty articles
contained three outbound links between them - none to the repo, npm or any
demo, and none to each other despite a dozen prose cross-references. A
reader convinced by an article had nowhere to go. Fixed with links to the
deployed demos, the repo, npm, and the articles that prose already pointed
at.
A re-verification pass found real errors in yesterday's corrections, which
is the more useful half:
- The convergence checkpoints are measured in OBSERVATIONS, not seconds. My
"double-digit metres at twenty seconds" was wrong: at ~40 observations,
which is roughly twenty seconds, the median is about 7 m. Corrected in the
one article whose stated point is publishing the uncomfortable number.
- The permission count. My retitle to "four dialogs" was wrong on the only
platform that runs this: the motion prompt is iOS-only and iOS does not
expose WebXR AR at all. The article no longer hard-codes a count.
- The framework DOES route the probe outcome into distinct messages; only one
helper is narrow. My correction generalised a function-scoped gap into a
framework-scoped accusation.
- The on-screen first-placement exemption applies only to anchors created
with bootstrapping skipped, not to every anchor's first placement.
- Planar PnP disambiguates by reprojection error alone - there is no temporal
tiebreak, which is why the flicker it describes can still occur.
- QR size precedence is the opposite of what I wrote: an authored size WINS
and the depth measurement is the fallback.
- A float32 step at earth radius is half a metre, not 0.4 - the 0.38 figure
is relative precision, not a step.
- The 30 fps camera-delivery figure is measured nowhere; replaced with the
distinction that actually matters (render rate is not sensor rate).
Editorial seams from the correction pass: a heading swap orphaned "the
opposite of what?", a "general lesson" section that had grown a second
lesson, a bolded claim retracted four lines below itself, an inserted
paragraph that broke the bold-led list it landed in, and two descriptions
still describing titles that had changed.
2d84465
fix(blog): the blur gate with its real numbers, and the rest of the fact-check
The blur-gate article now publishes the measurement instead of asserting a
benefit: 156 hand-labelled frames from one indoor session, four metrics
compared, a threshold sweep replayed through the real gate. The headline
finding is a negative one and it is the most useful part - no metric
separates cleanly, every AUC below 0.9, so there is no threshold that keeps
the good frames and drops the bad ones. k=0.5 rejects 7% and never rejected
a sharp frame but catches only a fifth of the blurry ones; k=0.8 rejects 30%
at ~0.98 precision and catches 59%. The default is now enabled at 0.8 per
the owner's decision, and the article states plainly what was NOT measured:
whether the retry produces a better frame, and anything at all about
outdoor scenes.
COLMAP: the export is posed, not georeferenced, and the article says so. It
also now carries the constraint that would bite anyone who tried - do not
export earth-centred coordinates, because splat trainers store geometry in
float32 and one step at earth radius is ~0.4 m.
PWA: the article is retitled "Shipping AR as static files" and the offline
claims are split - the computation genuinely has no server to be offline
from, the page has no service worker yet.
Also corrected against the code: capabilityMessage returns ONE templated
sentence and never sees the probe outcome, so the article's four-message
design is now described as absent rather than shipped; the Chromium
workaround is unconditional, not version-scoped; the permission order is AR
essentials then camera then orientation then geolocation LAST, which is the
opposite of what the article recommended and was decided by field testing;
HUD elements must be DOM DESCENDANTS of the overlay root, which is a nesting
rule and not a z-index one; the physics demo collides against a trimesh, not
the AABB list; the AR layer is a module singleton rather than an injected
seam; the QR demo does not replay recordings; H3 has twelve pentagons per
resolution; and the retry-after-timeout advice no longer promises it works.
The "fifteen seconds" rule of thumb is gone from every article, replaced
where relevant by the corpus measurement it contradicts: a median of about
two minutes across 51 recordings.
ed30eaf
fix(blog): the QR family, as tested parts rather than a shipped feature
Per the owner's decision, the marker-anchoring articles now say plainly what
is true: detection, the pose solver, size-from-depth, the stability window
and the vote builder exist and are tested, and no app wires them together -
the QR demo is deliberately geo-less and casts no GPS vote. Four articles
leaned on the assembled capability; each now offers the parts and says the
assembly is yours.
The plausibility check was described as something better than it is. It
takes the marker's centre POINT, not its pose, so it cannot compare the
marker's plane against the wall - which means it does not catch the two-fold
ambiguity flip, the exact failure the article credited it with. And it does
not use voxel ray traversal; that exists but carves free space into the
occupancy grid instead. The article's real thesis - a second sensor is worth
more for detecting contradictions than for reducing noise - survives intact.
The detection scheduler is fixed-rate with no mode switch, so the adaptive
search/tracking/idle rates are now the argument for having a scheduler
object rather than a description of one. Also corrected: the quality gate is
not in front of the detector today, and the camera does not hand you sixty
usable frames a second.
Smaller: VPS accuracy is decimetre-class not centimetre; VPS coverage is
large and growing rather than "increasingly not much else"; iOS does not
expose WebXR AR, so ARKit is the native counterpart rather than a runtime
this reaches; the QR size field still exists and is honoured, so measurement
makes it optional rather than absent; and "fifteen seconds" no longer appears
as a fact anywhere.
4d7cf8d
fix(blog): correct the scene-graph inversion and the removed anchor bypass
The fact-check found five articles stating things that are the inverse of
what the code does. All five share one root cause and it was ours: the
framework README described a large-jump bypass that gps-anchor.ts.md records
as REMOVED in June, and it describes the scene graph in a way three articles
then paraphrased backwards.
The scene-graph correction is the one a developer would have acted on. The
scene ROOT is GPS-aligned world space and arWorldGroup - which holds the
camera - is a group inside it carrying the alignment. So an object at the
scene root keeps its world pose perfectly and appears to float, because the
camera moves under corrections while it stands still; an object at a fixed
transform inside the group stays visually nailed down and drifts in GPS; and
createGpsAnchor lives inside the group but re-derives its local position each
tick, which is how it gets both. The articles had the first two swapped, and
described the minimal example's "floater" as the local one when it is in fact
the world-correct one.
The anchor correction replaces an invented rule with the real reason: there
is no large-correction escape hatch, because smoothing lives once on the
group's lerped matrix, so a big alignment change is already eased across the
whole view before per-anchor logic runs. A per-anchor on-screen snap did not
rescue anything - it manufactured the hard jump the shared lerp existed to
remove.
Also corrected: the median bootstrap only rejects multipath when the caller
supplies a GPS sampler, since the default derives from the object's own pose;
and the "fifteen seconds" rule of thumb no longer appears as a fact.
8058d6c
docs(blog): the first batch of thirty drafts
Every page is gated at `status: draft` - the site build renders only pages
marked published, and an unrecognised or missing status renders as draft.
Nothing here reaches gps.csutil.com/blog/ until a marker is flipped.
The first three (why-outdoor-webxr-drifts, gps-is-a-distribution-not-a-
position, what-you-give-up-by-not-using-a-vps) are the style probe: written
first, deliberately different in register, so an edit pass has an obvious
entry point and the corrections can propagate to the other 27.
Topics and the suggested publication order are in the private docs repo's
2026-08-20-0812-marketing-content-backlog.md. Verified against the real
renderer: 31 files in, 0 published, 31 withheld, each with a reason.
758ba8a