SamGu-NRX · Oct 14, 2026 · Oct 13, 2026 · Oct 13, 2026 · Oct 13, 2026 · Oct 13, 2026
diff --git a/Makefile b/Makefile
@@ -1,35 +1,37 @@
 # Suite targets follow their CI jobs where one exists. The viewer has no CI job yet.
 # `make check` runs server, web and ios, then scoring, measure-lab, evals, recon
-# meter-closeup and the integration viewer where the branch has them.
+# meter-closeup, verification and the integration viewer where the branch has them.
 # tests/makefile.sh checks this dispatch with stub suites.
 
 XCODEGEN ?= xcodegen
 # Must match XCODEGEN_VERSION in .github/workflows/ios.yml; other versions write a different project.
 XCODEGEN_VERSION := 2.46.0
 
-# The file each experiment suite's target needs. These suites arrive in separate pull requests, so
+# The file each optional suite's target needs. These suites arrive in separate pull requests, so
 # `check` includes one only when its file exists and the Makefile works before and after each
-# merge. Once all five are on main, list them in `check` directly so a deleted folder fails.
+# merge. Once all seven are on main, list them in `check` directly so a deleted folder fails.
 SCORING := experiments/scoring/pyproject.toml
 MEASURE_LAB := experiments/measure-lab/Makefile
 EVALS := experiments/evals/Makefile
 RECON := recon/Makefile
 METER_CLOSEUP := experiments/meter-closeup/Makefile
+VERIFICATION := verification/Makefile
 INTEGRATION_VIEWER := tools/integration-viewer/package.json
 
 # $(call present,FILE,TARGET) is TARGET when FILE exists, otherwise nothing.
 present = $(if $(wildcard $(1)),$(2))
 # $(call require,FILE,PR): stop before any command runs when the suite's file is missing.
 require = @test -f $(1) || { echo "make $@: $(1) is missing; the suite arrives with pull request $(2)." >&2; exit 1; }
 
-.PHONY: check server web ios ios-project smoke scoring measure-lab evals recon meter-closeup integration-viewer
+.PHONY: check server web ios ios-project smoke scoring measure-lab evals recon meter-closeup verification integration-viewer
 
 check: server web ios \
 	$(call present,$(SCORING),scoring) \
 	$(call present,$(MEASURE_LAB),measure-lab) \
 	$(call present,$(EVALS),evals) \
 	$(call present,$(RECON),recon) \
 	$(call present,$(METER_CLOSEUP),meter-closeup) \
+	$(call present,$(VERIFICATION),verification) \
 	$(call present,$(INTEGRATION_VIEWER),integration-viewer)
 
 server:
@@ -94,6 +96,12 @@ meter-closeup:
 	cd experiments/meter-closeup && uv sync --locked
 	$(MAKE) -C experiments/meter-closeup check
 
+# Match verification CI's locked sync so a stale lockfile cannot pass the local check.
+verification:
+	$(call require,$(VERIFICATION),#13)
+	cd verification && uv sync --locked
+	$(MAKE) -C verification test
+
 integration-viewer:
 	$(call require,$(INTEGRATION_VIEWER),#124)
 	cd tools/integration-viewer && node --test "test/*.test.js"

diff --git a/README.md b/README.md
@@ -52,6 +52,8 @@ The criteria (e.g. the distance the battery must be from the gas meter) exist in
 Below, one wall makes the whole trip, from photos to a checked spot. The models build the wall and name what's on it. The last step, the rule checks, is that deterministic evaluation.
 
 <picture>
+  <source media="(prefers-color-scheme: dark) and (prefers-reduced-motion: reduce)" srcset="docs/readme/rebuild-dark-still.webp">
+  <source media="(prefers-color-scheme: light) and (prefers-reduced-motion: reduce)" srcset="docs/readme/rebuild-light-still.webp">
   <source media="(prefers-color-scheme: dark)" srcset="docs/readme/rebuild-dark.webp">
   <img alt="An animated drawing of one wall in six steps. Photos arrive, each with the phone's position. Points are matched between them. A ruler marks feet out from the meter. Dots fill in the wall and the ground, while the left side nobody filmed stays hatched and reads Not seen. The window, meter, AC unit and gravel get labels. Last, a battery appears 12 ft right of the meter with its cable, beside a card where wall behind it, ground under it, open space in front and a 13.6 ft cable run all pass." src="docs/readme/rebuild-light.webp" width="100%">
 </picture>
@@ -75,6 +77,8 @@ Here's where each piece runs. The phone uploads the capture to Google Cloud Stor
 ## Unseen? Then Unsure
 
 <picture>
+  <source media="(prefers-color-scheme: dark) and (prefers-reduced-motion: reduce)" srcset="docs/readme/unseen-dark-still.webp">
+  <source media="(prefers-color-scheme: light) and (prefers-reduced-motion: reduce)" srcset="docs/readme/unseen-light-still.webp">
   <source media="(prefers-color-scheme: dark)" srcset="docs/readme/unseen-dark.webp">
   <img alt="A wall with the left side hazed over. Three checks read Not seen. The haze sweeps away, the checks turn Unsure, then settle: wall and ground pass, clear space fails, and the spot reads Not here." src="docs/readme/unseen-light.webp" width="100%">
 </picture>

diff --git a/docs/how-it-works.html b/docs/how-it-works.html
@@ -11,6 +11,10 @@
     --sky:#e2edf8; --wall:#e8dcc4; --eave:#6b5a4a; --ground:#c9dcb0; --bush:#6f9a52;
     --tape:#f3cf4b; --tape-ink:#3d3106;
     --phone:#2a78d6; --server:#eb6834; --fog:#8d8d86; --battery:#2f3e4e;
+    /* Text in the phone and server colours. The fills above are 4.0:1 and 2.9:1 on --bg, too
+       faint for small text; these keep the hue with OKLCH lightness lowered until text clears
+       4.5:1 on --bg (4.63 and 4.61) and on --card (5.06 and 5.03). */
+    --phone-ink:#1e6ecb; --server-ink:#c44400;
     --meterbox:#6f7680; --meter:#a9d3f2; --gas:#b9b3a6; --glass:#bfe0f5;
     --zone:rgba(208,59,59,.14); --zone-line:#d03b3b;
     --pass:#107a2a; --fail:#b83232; --fail-ink:#a12b2b; --unsure:#fab219;
@@ -24,6 +28,8 @@
       --sky:#1c2734; --wall:#4a4033; --eave:#8a7562; --ground:#2f4128; --bush:#5d8a45;
       --tape:#7a6016; --tape-ink:#fff3c4;
       --phone:#3987e5; --server:#e8683a; --fog:#6e6e68; --battery:#cfd8e3;
+      /* The dark fills already read as text: 4.69:1 and 5.26:1 on --card. */
+      --phone-ink:#3987e5; --server-ink:#e8683a;
       --meterbox:#8f97a3; --meter:#5c8fb8; --gas:#7d786d; --glass:#6f9fbf;
       --zone:rgba(230,103,103,.22); --zone-line:#e66767;
       --pass:#107a2a; --fail:#b83232; --fail-ink:#f28b8b; --unsure:#fab219;
@@ -42,7 +48,7 @@
   .prose{max-width:58ch}
   .meta{font-family:var(--sans);color:var(--muted);font-size:15px}
   .lede{font-size:20px;line-height:1.45;max-width:60ch}
-  a{color:var(--phone);text-underline-offset:3px}
+  a{color:var(--phone-ink);text-underline-offset:3px}
   a:focus-visible,summary:focus-visible{outline:2px solid var(--phone);outline-offset:3px;border-radius:2px}
   code{font-family:var(--mono);font-size:.86em;background:var(--code-bg);padding:1px 5px;border-radius:3px;overflow-wrap:anywhere}
   ul,ol{padding-left:1.3em}
@@ -99,8 +105,8 @@
   .lab.s{font-size:16px}
   .lab.b{font-weight:600}
   /* Colour utilities come after .lab and .t so they win at equal specificity. */
-  .lab.blue,.t.blue{fill:var(--phone)}
-  .lab.orange,.t.orange{fill:var(--server)}
+  .lab.blue,.t.blue{fill:var(--phone-ink)}
+  .lab.orange,.t.orange{fill:var(--server-ink)}
   .lab.red,.t.red{fill:var(--fail-ink)}
   .lab.mu,.t.mu{fill:var(--muted)}
   .tp{font-family:var(--sans);font-size:15px;fill:var(--tape-ink);font-weight:600}
@@ -715,43 +721,48 @@ <h3>Unseen means unsure</h3>
 </figure>
 
 <h3>Coverage is a fixed zone</h3>
-<p class="prose">Not the whole house, just a zone around the meter. The app always knows exactly what's left.</p>
+<p class="prose">Not the whole house. The zone is every spot the cable can reach, plus everything the checks on those spots look at. The rules fix it before the scan starts, so the app always knows exactly what's left.</p>
 <figure class="narrow">
   <svg viewBox="0 0 480 300" role="img" aria-labelledby="cvt cvd">
     <title id="cvt">The zone a scan must cover, seen from above</title>
-    <desc id="cvd">A plan view of the house with the electric meter on its front wall. A dashed blue rectangle marks the zone the checks need: the wall 20 feet either side of the meter and the ground in front of it out to 13 feet. Everything outside it is grey haze labelled not needed. Inside, a bush stands in front of the wall, and the strip of wall behind it is unseen, so checks there are UNSURE. A dotted path shows the phone walking along the zone.</desc>
-    <!-- 1 ft = 8 px, meter at x=240, wall face at y=110 -->
+    <desc id="cvd">A plan view of the house with the electric meter on its front wall. An orange band along the wall marks the candidate battery spots, 20 feet of cable either side of the meter. A dashed blue outline marks the zone the checks need: that band grown by how far each check reaches, so it runs past the cable's reach at both ends and out over the ground in front, with rounded front corners. A bracket under it reads needed: all of it, past the cable's reach too. Everything outside is washed out and labelled no check looks here. Inside the zone, a bush stands in front of the wall, and the strip of wall behind it is unseen, so checks there are UNSURE. A dotted path shows the phone walking through the zone.</desc>
+    <!-- 1 ft = 7 px, meter at x=240, wall face at y=110. Candidate footprints (22 in deep, 13 px)
+         fill a band within 20 ft of cable, x 100 to 380. The zone is that band grown by the
+         farthest check. Its front sits at the 13 ft of ground the pool and driveway checks need
+         (docs/00-overview.md), y=201, so it grows the band by about 11 ft (78 px): the same
+         reach puts its ends at x 22 and 458 and rounds its front corners. -->
     <rect width="480" height="300" fill="var(--ground)"/>
-    <rect x="40" y="16" width="400" height="88" fill="var(--eave)"/>
-    <rect x="40" y="104" width="400" height="6" fill="var(--wall)"/>
-    <text class="lab b" x="240" y="48" text-anchor="middle">house, seen from above</text>
-    <!-- cable reach bracket -->
+    <rect x="0" y="16" width="480" height="88" fill="var(--eave)"/>
+    <rect x="0" y="104" width="480" height="6" fill="var(--wall)"/>
+    <text class="lab b" x="240" y="44" text-anchor="middle">house, seen from above</text>
+    <!-- candidate spots: cable reach bracket and the band of footprints -->
+    <g stroke="var(--server)" stroke-width="2.5">
+      <line x1="100" y1="88" x2="380" y2="88"/><line x1="100" y1="81" x2="100" y2="95"/><line x1="240" y1="81" x2="240" y2="95"/><line x1="380" y1="81" x2="380" y2="95"/>
+    </g>
+    <text class="lab s" x="240" y="74" text-anchor="middle">candidate spots: 20 ft of cable either side</text>
+    <rect x="100" y="110" width="280" height="13" fill="var(--server)" fill-opacity=".22" stroke="var(--server)" stroke-width="2"/>
+    <!-- outside the zone: washed out, not hazed, because haze means unseen and needed -->
+    <path d="M0 104 H480 V300 H0 Z M22 104 V123 A78 78 0 0 0 100 201 H380 A78 78 0 0 0 458 123 V104 Z" fill="var(--card)" fill-opacity=".6" fill-rule="evenodd"/>
+    <!-- the zone -->
+    <path d="M22 104 V123 A78 78 0 0 0 100 201 H380 A78 78 0 0 0 458 123 V104" fill="none" stroke="var(--phone)" stroke-width="2.5" stroke-dasharray="7 4"/>
     <g stroke="var(--phone)" stroke-width="2.5">
-      <line x1="80" y1="88" x2="400" y2="88"/><line x1="80" y1="81" x2="80" y2="95"/><line x1="240" y1="81" x2="240" y2="95"/><line x1="400" y1="81" x2="400" y2="95"/>
+      <line x1="22" y1="214" x2="458" y2="214"/><line x1="22" y1="207" x2="22" y2="221"/><line x1="458" y1="207" x2="458" y2="221"/>
     </g>
-    <text class="lab" x="240" y="76" text-anchor="middle">cable reach: 20 ft either side</text>
-    <!-- haze outside the needed zone -->
-    <rect x="0" y="102" width="80" height="198" fill="url(#fogpat)"/>
-    <rect x="400" y="102" width="80" height="198" fill="url(#fogpat)"/>
-    <rect x="80" y="214" width="320" height="86" fill="url(#fogpat)"/>
-    <text class="lab b" x="240" y="266" text-anchor="middle">not needed</text>
-    <!-- needed zone -->
-    <rect x="80" y="104" width="320" height="110" fill="none" stroke="var(--phone)" stroke-width="2.5" stroke-dasharray="7 4"/>
+    <text class="lab s b blue" x="240" y="238" text-anchor="middle">needed: all of it, past the cable's reach too</text>
+    <text class="lab mu" x="240" y="280" text-anchor="middle">no check looks here</text>
     <!-- meter -->
     <rect x="234" y="101" width="12" height="9" rx="1.5" fill="var(--meterbox)"/>
-    <text class="lab" x="252" y="130">meter</text>
+    <text class="lab" x="228" y="142" text-anchor="end">meter</text>
     <!-- bush hiding a strip of wall -->
-    <rect x="318" y="101" width="24" height="9" fill="url(#fogpat)" stroke="var(--fail-ink)" stroke-width="1.5"/>
-    <circle cx="330" cy="124" r="12" fill="var(--bush)"/>
-    <text class="lab" x="320" y="158" text-anchor="middle">wall behind the bush:</text>
-    <text class="lab" x="320" y="178" text-anchor="middle">unseen, so UNSURE</text>
+    <rect x="333" y="101" width="24" height="9" fill="url(#fogpat)" stroke="var(--fail-ink)" stroke-width="1.5"/>
+    <circle cx="345" cy="124" r="11" fill="var(--bush)"/>
+    <text class="lab s" x="345" y="152" text-anchor="middle">wall behind the bush:</text>
+    <text class="lab s" x="345" y="170" text-anchor="middle">unseen, so UNSURE</text>
     <!-- the phone's path -->
-    <line x1="96" y1="186" x2="384" y2="186" stroke="var(--phone)" stroke-width="3" stroke-dasharray="1 7" stroke-linecap="round"/>
-    <use href="#phone" x="140" y="173"/>
-    <text class="lab b blue" x="90" y="132">needed: all of it</text>
-    <text class="lab" x="392" y="208" text-anchor="end">ground out to 13 ft</text>
+    <line x1="70" y1="186" x2="410" y2="186" stroke="var(--phone)" stroke-width="3" stroke-dasharray="1 7" stroke-linecap="round"/>
+    <use href="#phone" x="150" y="173"/>
   </svg>
-  <figcaption>The zone: wall within 20 ft of cable each side, ground out to 6 to 13 ft, and headroom above. It reaches a little past 20 ft for clearances at the edge.</figcaption>
+  <figcaption>Each check reaches out from the spot it tests: 3 ft along the wall for gas meters, AC units, doors and windows, 5 ft for a driveway and 10 ft for a pool (both placeholders, with no public value), and up for headroom and wall equipment. So the zone runs past the 20 ft of cable at both ends, and over the ground out to 13 ft. Any part of it nobody saw keeps its checks UNSURE.</figcaption>
 </figure>
 
 <h3>LiDAR helps but isn't required</h3>

diff --git a/docs/readme/rebuild-dark-still.webp b/docs/readme/rebuild-dark-still.webp
diff --git a/docs/readme/rebuild-light-still.webp b/docs/readme/rebuild-light-still.webp
diff --git a/docs/readme/unseen-dark-still.webp b/docs/readme/unseen-dark-still.webp
diff --git a/docs/readme/unseen-light-still.webp b/docs/readme/unseen-light-still.webp
diff --git a/experiments/drift-anatomy/METHODS.md b/experiments/drift-anatomy/METHODS.md
@@ -8,8 +8,8 @@ The server allows 0.16 ft of error per foot along the wall. That allowance domin
 
 - **A1, return to the meter.** When a walk comes back within 0.3 m of an earlier point more than 30 s later, ARKit's closure residual r is the error in the displacement between the two visits. Does r bound the largest error between the visits?
 - **A2, scale.** Is ARKit's scale one number per walk? This measures the spread of scale over sliding 5 m windows within a walk.
-- **A3, the wall plane.** What share of the error runs across the direction of travel, which a wall plane would remove, compared with along it?
-- **A4, a range to the meter.** A second phone left at the meter ranges the walking phone by UWB (σ 10 cm, simulated) at every window end. What is the p90 left over?
+- **A3, along and across travel.** What share of the error runs across the direction of travel, compared with along it? A wall plane removes the across component only when travel is parallel to that wall; these walks do not annotate that orientation.
+- **A4, a range to the meter.** A second phone left at the meter ranges the walking phone by UWB (σ 10 cm, simulated) at every window end. What is the p90 left over? Each window is one session that starts at the meter: the walking phone leaves the meter at the window's start, the second phone stays there, and the range runs from that start to the window's end. Sliding the start along a walk samples many meter placements; the error depends on the path walked from the meter, not on where in the dataset the meter sits.
 
 ## Pass criteria, fixed before the run
 
@@ -20,13 +20,13 @@ The server allows 0.16 ft of error per foot along the wall. That allowance domin
 
 ## Result
 
-All four fail. Tables: [results/drift_anatomy.md](results/drift_anatomy.md).
+A1, A2 and A4 fail their numeric criteria. A3 fails the 60% criterion for a travel-axis proxy; this is not evidence that a wall-plane experiment failed. Wall-plane benefit remains untested without wall annotations. Tables: [results/drift_anatomy.md](results/drift_anatomy.md).
 
 - **A1.** 30 of 56 MARViN loops (54%). ADVIO has no revisit under 0.3 m while ARKit tracks. Scale and heading error cancel on return, so r misses them.
 - **A2.** 5.7% SD on MARViN, 19% on ADVIO: drop. A few outlier windows inflate MARViN's SD; its robust SD is 1.8%.
-- **A3.** At 20 ft, along is 0.92 of the p90 on MARViN and 0.76 on ADVIO.
+- **A3.** At 20 ft, along is 0.92 of the p90 on MARViN and 0.76 on ADVIO. The straight-path check selects windows by reference-path deviation from their start-to-end line: within 0.25 m, along is 0.98 of the p90 on MARViN and 0.75 on ADVIO; within 0.5 m, 0.95 and 0.75. These untuned, non-pre-registered filters select straightness only. With no wall annotations, they cannot identify wall-parallel travel or establish a wall-plane constraint's benefit.
 - **A4.** MARViN goes from 10.4 and 13.6 in to 7.7 and 8.5 in. At σ 10 cm the range's own p90 is 6.5 in, so the bar was nearly out of reach. At 5 cm: 4.9 and 6.1 in.
 
 **Run:** `uv run python run.py`. Needs the evals checkout at 190ed33 (`EVALS_HARNESS`) and MARViN and ADVIO in `~/house-scanning-data`.
 
-**What it changes:** the error worth attacking runs along travel. A range to the meter reduces it, a wall plane wouldn't, and r can't bound a session.
+**What it changes:** the error worth attacking runs along travel. A range to the meter reduces it, and r cannot bound a session. These unannotated walks do not establish whether a wall-plane constraint helps.
diff --git a/experiments/drift-anatomy/README.md b/experiments/drift-anatomy/README.md
@@ -2,7 +2,7 @@
 
 ## Question
 
-The server allows 0.16 ft of error per foot along the wall, which dominates a check's error bar beyond a few feet. Which part of ARKit's error is worth attacking? A1 asks whether returning to the meter bounds it, A2 whether scale is one number per walk, A3 how much runs across travel, where a wall plane would help, and A4 what a UWB range to the meter leaves. The data is MARViN (iPhone 14 Pro Max) and ADVIO (iPhone 6s); [METHODS.md](METHODS.md) has the details.
+The server allows 0.16 ft of error per foot along the wall, which dominates a check's error bar beyond a few feet. Which part of ARKit's error is worth attacking? A1 asks whether returning to the meter bounds it, A2 whether scale is one number per walk, A3 how much runs across travel, and A4 what a UWB range to the meter leaves. Each window stands for a session that starts at the meter, so in A4 the second phone sits at the window's start. The data is MARViN (iPhone 14 Pro Max) and ADVIO (iPhone 6s); [METHODS.md](METHODS.md) has the details.
 
 ## Pass criteria, fixed before the run (397507f)
 
@@ -13,13 +13,13 @@ The server allows 0.16 ft of error per foot along the wall, which dominates a ch
 
 ## Result
 
-All four fail ([tables](results/drift_anatomy.md)).
+A1, A2 and A4 fail their numeric criteria. A3 fails the 60% criterion for a travel-axis proxy; the unannotated walks cannot test wall-plane benefit ([tables](results/drift_anatomy.md)).
 
 - **A1.** The closure residual bounds the peak error in 30 of 56 loops. Scale and heading error cancel on return.
 - **A2.** Scale spread within a walk is 1.8% robust SD (5.7% SD) on MARViN.
-- **A3.** At 20 ft, the along-travel p90 is 9.6 of 10.4 in.
+- **A3.** At 20 ft, the along-travel p90 is 9.6 of 10.4 in. In the straight-path check, along is 0.95 to 0.98 of the total on MARViN (added after review). The walks have no wall annotations, so this does not establish wall-parallel travel or the benefit of a wall-plane constraint.
 - **A4.** UWB cuts 20 ft from 10.4 to 7.7 in at σ 10 cm, and to 4.9 in at σ 5 cm.
 
 **Run:** `uv run python run.py`: 20 s. Needs the evals checkout at 190ed33 and MARViN and ADVIO in `~/house-scanning-data`.
 
-**What it changes:** the error worth attacking runs along travel. Measuring a span inside one photo avoids it. A wall plane, a return to the meter or one ruler at the meter don't.
+**What it changes:** the error worth attacking runs along travel. Measuring a span inside one photo avoids it. The return-to-meter and one-ruler criteria fail. These unannotated walks do not establish whether a wall-plane constraint helps.