# Rich-Display Overlay Notes

Phase 10 of `docs/SUNDOG_V_GEOMETRY.md` uses these notes to keep the
optional vocabulary layers separate from the calibrated parhelion core.
Image 1 is the label key; images 2, 7, and 13 are the first tuning set.

> **Post-audit state, 2026-05-14.** Required Phase 10 attack passes
> have landed and the re-audit gate is recorded in
> [`PHASE10_OPTICAL_REAUDIT_MEMO.md`](PHASE10_OPTICAL_REAUDIT_MEMO.md).
> This file now carries the post-pass route verdicts; the specialist
> handoff and public-framing surfaces still need their ratchet pass per
> [`../PHASE10_ATTACK_ROADMAP.md`](../PHASE10_ATTACK_ROADMAP.md).
>
> - **Pass B1 landed 2026-05-13** (first technical gate). The
>   parhelion-route residual table now has a per-photo eligibility
>   sub-table with `sec(h) − 1` (geometric lever), `R22-source`
>   (`ring-fit` / `parhelion-derived` / `inferred-other`), and
>   `geometric_validity` columns. Each anchor JSON
>   (`docs/calibration/p*-anchor.json`) carries a top-level
>   `r22_source` + `r22_source_note`; p26's anchor adds a
>   `geometric_validity` block flagging the right-side
>   impossibility (`R22 / right_offset = 1.003 > 1`,
>   `arccos` undefined). Rationale: synthetic optical audit
>   ([`PHASE10_OPTICAL_AUDIT_SYNTHETIC_MEMO.md`](PHASE10_OPTICAL_AUDIT_SYNTHETIC_MEMO.md))
>   §2 items 4–7 and §2 item 6.
> - **Phase 10 Promotion Verdict and Single-handle closeout below
>   are post-pass.** Pass A3 re-derived the CZA and supralateral
>   verdicts, Pass C1 removed p7 from upper-tangent eligibility,
>   and Pass B2 re-derived the parhelion route against the honest
>   eligibility set.
> - **The CZA-route residual table has been repaired.** A1a/A1b
>   replaced the legacy `sun_y - R46` hardcode with the literature
>   CZA expression; p2 now has a +1.3 px y residual, and p27 is no
>   longer CZA evidence.
> - **The tangent verdict remains detector-bounded.** p7 is removed
>   as circumscribed-halo regime; column-peak still fails on p2 /
>   p13 / p27, and optional Pass C2 remains the open tooling
>   question.
>
> Do not export the closeout language below to public-framing
> surfaces (gravity ledger, mesa crossover note, homepage elevator
> pitch) until the post-re-audit ratchet pass has rewritten those
> surfaces from the re-audit memo.

## Artifacts

| image | overlay | role |
| --- | --- | --- |
| `1.Photometeor-jeff-mod_marked_red.jpg` | source annotation | Label key for CZA, supralateral, suncave Parry, upper tangent, Parry supralateral candidate, infralateral, parhelic circle, lower tangent, parhelia, 22° halo, and 46° halo. |
| `2.Photometeor-jeff_mod_red.jpg` | `overlays/p2.rich-vocabulary.overlay.png` | Clean modern version of the annotated display; strongest rich-vocabulary tuning source. |
| `7.625544777_10241089047341957_4435817776770420050_n.jpg` | `overlays/p7.rich-vocabulary.overlay.png` | High-sun / partial-visibility stress case. |
| `13.480859565_17934474635991868_323320248088719839_n.jpg` | `overlays/p13.rich-vocabulary.overlay.png` | Square social-crop / logo-crop tuning source. |

## Anchor Summary

| image | sun px | R22 px | parhelion offset | inferred h | note |
| --- | ---: | ---: | ---: | ---: | --- |
| p2 | `(567, 496)` | 182 | 192 | 18.6° | Reuses the clean multi-photo calibration anchor; dagger residuals are -1 px / -1 px. |
| p7 | `(1033, 946)` | 200 | 393 | 59.4° | Reuses the high-sun anchor; CZA should be treated as not applicable at this altitude even when the vocabulary overlay draws it. |
| p13 | `(543, 372)` | 211 | 213 / 212 | 6.83° | **Re-anchored 2026-05-13 (Task #52 step 1).** Supersedes rough hand-anchor: sun_x corrected −14 px, offsets refined from assumed-symmetric 220 to measured 213/212, h drops from 17.3° to ~6.8°. p13 is a low-altitude photo, not mid-altitude. Anchor file: `p13-anchor.json`. |
| p20 | `(1011, 827)` | 455 | 457 / 457 | ~5° | **Task #55 fallback check.** Plausible R22 anchors put the CZA apex above the frame (`y ≈ -83`), so p20 is not CZA-eligible. Anchor file: `p20-anchor.json`. |
| p27 | `(596, 559)` | 219 | 219 / 219 | ~0.5° | **Re-classified Pass A2 2026-05-13.** Originally anchored as CZA expansion check (Task #55); the visible chromatic arc at `(599, 142)` is now identified as the **46° halo top / supralateral merger** at ~41.89° above sun, not CZA. Per Pass A1a/A1b reverification, the literature CZA at h = 0.5° sits ~57° above sun (px y = −11.5, off-frame above the top), so the visible feature cannot be CZA. CZA fields preserved under `_disputed.cza_apex_legacy_2026_05_13` for reversibility. Now feeds supralateral-route eligibility (candidate, pending Pass A3 coverage check); does NOT feed CZA-route residual evidence. Anchor file: `p27-anchor.json`. |
| p22 | `(668, 453)` | 505 | 508 / 507 | ~5.7 deg | **Phase 10 belt-y FF anchor.** Strong bilateral low-sun parhelia; observed belt y `428` lands on the current `-0.05 * R22` rule. Anchor file: `p22-anchor.json`. |
| p26 | `(464, 204)` | 323 | 332 / 322 | ~9.0 deg | **Phase 10 belt-y FF anchor.** Observed belt y `177` sits ~11 px above the current rule; left/right parhelion y differ by 12 px, so this also carries a parhelic-tilt flag. Anchor file: `p26-anchor.json`. |
| p30 | `(701, 934)` | 650 | 666 / 659 | ~11.1 deg | **Phase 10 belt-y FF anchor.** Rich low-sun halo reference; observed belt y `899` is ~3 px above the current rule. Side parhelia are edge-supported and vertically broad, so keep the tilt/edge flag attached. Anchor file: `p30-anchor.json`. |
| p25 | `(489, 186)` | 300 | 305 / 307 | ~11.4 deg | **Phase 10 belt-y robustness anchor.** Observed belt y `176` is about 5 px below the predicted line in image coordinates (script residual -5 px; FF residual +5 px); right-side reference is cleaner than the foreground/flare-contaminated left, so keep the single-side confidence caveat attached. Anchor file: `p25-anchor.json`. |

All generated overlays now apply `--parhelic-y-offset-r22 = -0.05`, raising
the parhelic belt and dagger markers by 5% of the observed 22° halo radius.
This fixes the shared vertical belt bias without changing the 22°/46° halo
registration. Track that belt-height residual separately from the
parhelion-offset inversion route: p13's committed anchor clears the x-offset
route, but its JSON `parhelion.y = 351` left a single-photo belt-y
watch-list residual against the current `-0.05 * R22` overlay rule
(+10.4 px in the script's predicted-minus-observed convention; -10.4 px
in the FF observed-minus-predicted convention).

## Phase 10-FF Belt-Y Verdict

Result note: [`PHASE10_BELT_Y_RESULTS.md`](PHASE10_BELT_Y_RESULTS.md)

The belt-y watch-list flag is retired. The Phase 10-FF replication set
(p27, p22, p13, p26, p30, p25) falsified FF1: three photos reached
`|residual| >= 5 px`, but the signs disagreed and Spearman
`rho(h, residual)` was +0.086. FF2 was not gated. FF3 passed: no low-h
anchor reached the 12 px primitive threshold. Do not promote a low-h
parhelic-belt correction from p13 alone.

## Phase 10 Measurement Pre-Registration

Geometry resolution, 2026-05-13: Mesa's crossover note makes the residual
table and negative thresholds load-bearing for the next Phase 10 run.

- The `per-inversion-route` residual table lives in this file, alongside the
  per-photo overlay notes. It is a Phase 10 deliverable, not a separate
  document, because the route residuals need to sit next to the photo-level
  visibility classifications that determine whether each route is applicable.
- The quantitative `do not promote` thresholds are a Phase 10 blocker for
  promotion. If the thresholds are not written before an overlay run, that run
  is exploratory only and cannot move a primitive into default/logo language.
- Thresholds are evaluated at the image anchor scale and normalized by each
  photo's measured `R22`.

Pre-registered `do not promote` thresholds:

| candidate class | do not promote if... |
| --- | --- |
| Optional vocabulary primitive | feature residual is `>= 0.06 * R22` or `>= 12 px` on at least two eligible photos |
| Inversion route | route residual is `>= 0.04 * R22` or `>= 8 px` on at least two eligible photos |
| Coverage | fewer than two eligible photos show the primitive/route clearly enough to measure |
| Attribution metric | metric reports a linear "arc X contributes N%" score instead of a route/primitive residual |

`Eligible` means the feature is visible, uncropped enough to measure, and
inside its sun-altitude validity window. Hidden, cropped, or high-sun invalid
features are recorded as `not applicable`, not as failures.

Probation rule, added with the 2026-05-13 verdict: a single eligible photo
touching a cutoff does not fail a route or primitive under the pre-registered
`>= 2 photos` rule, but it does put that item on anchor-probation until the
next anchor-capture pass confirms or clears the residual.

Mesa v3.8 crossover addendum: Phase 11 metric review must keep three
quantities separate: visual/variance salience, route-local residual or
sensitivity, and full-overlay fit. The mesa PC1 result is the warning case:
a component can be variance-heavy and locally sensitive without reproducing
the full multi-component mechanism. For geometry, this means no "arc X
contributes N% of fit" score, no aggregation that hides route-specific
conflicts, and no promotion when a primitive improves one route while
worsening another unless the scope is explicitly narrowed.

Task #54 / Phase 11 measurement schema, inherited from v3.8:

| field | what it records | cannot be used as |
| --- | --- | --- |
| visual salience | how strong / obvious the arc is in the photograph | proof that the arc is load-bearing for inversion |
| route-local residual | residual for the named inversion route at that photo's anchor | total overlay quality |
| route-local sensitivity | whether small anchor / sample changes swing the route residual | primitive importance percentage |
| full-overlay interaction | whether tuning this primitive improves or worsens other routes | an averaged global score |
| partition/conflict flag | shared across photos/routes, photo-specific, or actively conflicting | a reason to hide the conflict in aggregation |

Promotion decisions must name which field they rely on. A primitive that is
visually salient but route-local unstable stays vocabulary; a route that
improves one residual while worsening another is a conflict receipt, not a
mean score to optimize away.

## Per-Inversion-Route Residual Table

This table is the Phase 10 response to the mesa finding that forward and
inverse directions are not symmetric. Do not infer that a good parhelion-offset
fit makes every route good.

> **Pass B1 update, 2026-05-13:** the parhelion-route row below is the
> rolled-up summary; the per-photo eligibility breakdown — including
> the new `sec(h) − 1` (geometric lever), `R22-source`, and
> `geometric_validity` columns required by the audit memo §2 items 4–7
> — lives in the **Parhelion-Route Per-Photo Eligibility** sub-table
> immediately below this table. Pass B2 has now re-derived the row
> against that honest eligibility set; read the rolled-up row together
> with the per-photo table rather than as a free-standing "all photos"
> claim.

| route | p2 residual | p7 / p13 residual | expansion residuals | promotion status | note |
| --- | ---: | ---: | ---: | --- | --- |
| Parhelion offset → sun altitude | −1 / −1 px (0.5% R₂₂) | p7: 0 / n/a px; p13: 0 / 0 px (new anchor; x-offset only) | p20 / p27 are provisional low-sun supports, not promotion measurements | **promoted (post-audit hedged; Pass B2 2026-05-14)** | **Pass B2 verdict, 2026-05-14.** Route survives promotion against the post-B1 eligibility set, with materially weaker language per audit memo §4.1 / §6. **Audit-survived wording:** *passes residual gate at ~0 px on three photos (p2 h = 18.6°, p7 h = 59.4°, p13 h = 6.83°) with both unambiguous bilateral peaks and an independently fittable 22° halo.* On five additional low-h anchored photos the geometric lever `sec(h) − 1` is below 2 % of R22 (p13 0.71 %, p20 0.38 %, p22 0.50 %, p25 2.01 %, p26 1.25 %, p27 ≈ 0 %, p30 1.90 %) — clean residuals on those photos are anchor-noise-bounded, *not* route-validating. On three of those (p20, p25, p26) the 22° halo arc is not visible in the photograph; on those photos the residual measures offset-vs-R22 against an R22 that is not independent of the parhelion pick (tautological). p27 records `offset := R22` explicitly in its anchor file and is a self-consistency check, not a measurement. p26 right side encodes a literal geometric impossibility (`R22 / offset = 1.003 > 1`); flagged `invalid` in the Pass B1 schema. **The pre-audit "~0 px on every eligible photo" framing is retired.** The route stays in calibrated core; verdict reads as 'route holds on the strict eligible subset, anchor-noise-dominated below 2 % lever.' See Parhelion-Route Per-Photo Eligibility below for the full per-photo breakdown. |

### Pass A1a Spec Results — CZA Literature Formula vs. Legacy Hardcode

Pass A1a deliverable from
[`../PHASE10_ATTACK_ROADMAP.md`](../PHASE10_ATTACK_ROADMAP.md), landed
2026-05-13. Module at `scripts/cza_formula.py`; regression test at
`scripts/test_cza_formula.py`. Reproduce with `python3
scripts/test_cza_formula.py` from the repo root.

**Verified literature formula:**

```
CZA_above_sun_deg(h) = arcsin(sqrt(n^2 - cos^2 h)) - h     for h <= 32.196 deg
                     = (CZA disappears)                     for h >  32.196 deg
where n = 1.31 (refractive index of ice for the CZA wavelength).
```

**Memo formula correction.** The audit memo §2 item 2 *states* the
formula as `90 - h - arcsin(sqrt(n^2 - cos^2 h))`, but that expression
gives 0.27° at h = 22° (memo claims ~46°) and 31.69° at h = 0.5° (memo
claims ~57°). The expression that matches the memo's *own numerical
predictions* is `arcsin(sqrt(n^2 - cos^2 h)) - h`. Pass A1a's
verify-gate caught the memo's stated formula was a transcription error
before it leaked into the atlas patch (§7 of the attack roadmap rule).
The qualitative finding (formula bug exists, p27 visible "CZA" is
mis-identified) survives unchanged; only the *expression* used to land
A1b is corrected. This is the second time in this campaign the verify
gate has caught a load-bearing error in upstream prose (the first was
the Persona 3 p22 over-statement caught by the synthetic memo's own
verification gate).

**Sanity check at h = 22° (the legacy operating point):**

| quantity | value |
| --- | --- |
| literature CZA-above-sun (deg) | 45.734° |
| legacy CZA-above-sun, from `WB_R46 / (WB_R22 / 22)` | 44.000° |
| delta | +1.734° |

The legacy `WB_R46 = 440` encodes 44° (because WB_R22 = 220 px at
10 px/deg), not 46°. The audit memo §2 item 1 flags this as the
second compounding bug (`WB_R46` in pixels is ~9.1 % too small). Even
at h = 22° the literature formula and the legacy disagree by ~17 px
in workbench coords; below h = 22° the gap widens fast.

**Per-photo regression — observed apex y minus predicted apex y:**

| photo | h (°) | r22 (px) | sun_y | observed apex y | literature predicted y | literature residual | legacy predicted y | legacy residual |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| p2  | 18.6 | 182 | 496 | 113 | 114.3 | **−1.3 px** | 132.0 | −19.0 px |
| p27 | 0.5  | 219 | 559 | 142 | **−11.5** (off-frame) | +153.5 px (meaningless — visible feature is not CZA) | 121.0 | +21.0 px |

**Qualitative-direction verdict:**

- **p2:** literature residual (−1.3 px) is closer to zero than legacy
  (−19.0 px) by 17.7 px. **CONFIRMED** the audit memo's prediction
  ("residual collapses to <5 px"; memo also predicts "~0.7 px" but
  flags it as illustrative). Memo §2 items 1–2 stand.
- **p27:** literature CZA apex y = −11.5 sits above the top of the
  frame (off-screen). The visible chromatic arc anchored at y = 142
  is therefore not the CZA at all. **CONFIRMED** the audit memo's
  prediction (memo §2 item 3: "the chromatic arc visible at y = 142
  in p27 sits at ~42° above the sun and is the 46° halo top /
  supralateral merger"). Pass A2 will re-anchor it.

**A1b disposition:** **CLEARED to proceed** with the verified literature
formula (`arcsin(sqrt(n^2 - cos^2 h)) - h`), not the memo's stated
formula. A1b touches `scripts/overlay_calibrate.py:381–384` (replace
`anchored = WB_SUN[1] - WB_R46` with a call into
`cza_formula.cza_apex_y_above_sun_px(h, WB_R22)` subtracted from
`WB_SUN[1]`) and `scripts/overlay_calibrate.py:66` (replace
`WB_R46 = 440` with `WB_R46 = round(2.091 * WB_R22)` ≈ 460). Per the
attack roadmap A1b touch block, the per-overlay `R46_px = r22_obs ×
(46/22)` derivation alternative remains explicitly out-of-scope
(would touch supralateral and other workbench-space consumers).

### Pass A1b Smoke Results — Atlas Formula Patch

Pass A1b deliverable from
[`../PHASE10_ATTACK_ROADMAP.md`](../PHASE10_ATTACK_ROADMAP.md), landed
2026-05-13. Two surgical edits to `scripts/overlay_calibrate.py` (line
66: `WB_R46` constant; lines 381–384: `cza_apex` inner function), plus
the new `cza_formula` import.

Reproduce the p2 smoke test from repo root with:

```bash
python3 scripts/overlay_calibrate.py docs/calibration/2.Photometeor-jeff_mod_red.jpg \
    --anchors docs/calibration/p2-anchor.json \
    --sun-altitude 18.6 --supralateral 0.40 --out /tmp/p2_overlay_a1b.png
```

**p2 smoke (h = 18.6°):**

| quantity | value |
| --- | --- |
| sun pixel | (567, 496) |
| R22 observed | 182 px |
| **R46 predicted** | **380.5 px** *(was 364.0 px before A1b; WB_R46 went 440 → 460)* |
| **CZA apex predicted** | **(567.0, 114.3)** *(was (560.0, 132.0) before A1b)* |
| CZA apex observed | (560, 113) |
| **CZA apex residual (predicted − observed)** | **(+7.0, +1.3) px** *(was (+0.0, −19.3) under the legacy comparison; sign-convention swap from A1a noted)* |
| left/right dagger residual | −0.0 / +0.0 px (unchanged; parhelion route not touched by A1b) |

The y-residual collapsed from −19.3 px (legacy, observed-minus-predicted
convention from p2-anchor.json `cza_apex_residual.y_px`) to +1.3 px
(literature, predicted-minus-observed convention from
overlay_calibrate's residual report). The magnitudes match A1a's
measured target ("literature residual = −1.3 px" under the
observed-minus-predicted convention).

**p7 smoke (h = 59.4°, beyond CZA disappearance):**

The CZA disappears geometrically at h ≈ 32.2° (per
`scripts/cza_formula.py`). The patched `cza_apex` falls back to the
corrected `WB_R46 = 460` anchor when the literature formula returns
`None`. Smoke run on p7 (sun = (1033, 946), R22 = 200, h = 59.4°)
renders without crash; CZA apex predicted at (1033.0, 527.8)
= sun_y − (WB_R46 × scale) = 946 − 460 × 0.9091 = 946 − 418.2.
Downstream visibility classification still marks CZA as "not
applicable" at this altitude; the fallback is only there to keep the
curve renderable rather than crash the script.

**Supralateral spot-check on p2 (roadmap exit criterion):**

The supralateral apex base is `WB_SUN[1] - WB_R46` (line 412), so the
WB_R46 = 440 → 460 change shifts it from 44° to 46° above sun
(+2°, or +4.55% relative). In p2 photo coords (scale 0.8273) this is
a 16.5 px upward shift of the supralateral apex. The roadmap exit
criterion expected "~4% outward from sun-above; if much more, the
workbench-space change has a knock-on effect that needs a follow-up
pass." Observed 4.55% is within tolerance — **no follow-up pass
needed.**

**Knock-on: 46° halo radius.** The 46° halo is drawn at
`WB_R46 * scale` (line 344). Before A1b: 440 * scale; after A1b:
460 * scale. On p2 this is 380.5 px instead of 364.0 px — the 46° halo
is now drawn ~4.5% angularly larger, which is the *correct* direction
(literature R46 = 2.091 × R22, not 2 × R22).

**Out of scope (deferred to Phase 10 backlog):** the per-overlay
`R46_px = r22_obs × (46/22)` derivation alternative. That option would
remove `WB_R46` from the workbench-space load-bearing path entirely
and let each photo derive its own R46 from its measured R22, but it
would require touching every workbench-to-photo consumer of `WB_R46`
(46° halo, supralateral, and any future primitive that uses the
constant). Filed; not blocking Pass A2/A3 or the re-audit gate.

### Pass A2 Results — p27 Primitive Re-Classification

Pass A2 deliverable from
[`../PHASE10_ATTACK_ROADMAP.md`](../PHASE10_ATTACK_ROADMAP.md), landed
2026-05-13. One anchor edit + four references walked back across this
file's CZA-route table, supralateral-route table, anchor summary, and
v3.8 tangent receipt.

**Geometric verification (`scripts/cza_formula.py` against
`p27-anchor.json`):**

| primitive | predicted apex y above sun (deg / px) | predicted pixel y | observed pixel y (visible feature) | delta |
| --- | ---: | ---: | ---: | ---: |
| Literature CZA at h = 0.5° | 57.31° / 570.5 px | **−11.5 (off-frame)** | 142 | n/a (off-frame; visible feature is not CZA) |
| 46° halo top (always 46°) | 46.00° / 458.0 px | 101.1 | 142 | +40.9 px (~4.11° below) |
| Supralateral at h = 0.5° (~46° per literature merger) | ~46.00° / 458.0 px | 101.1 | 142 | +40.9 px (~4.11° below) |

The visible chromatic arc at y = 142 is at 41.89° above sun. The
literature CZA at h = 0.5° is off-frame (above the top of the photo),
so the visible feature cannot be CZA. The 46° halo top and the
supralateral arc both predict ~46° above sun and theoretically coincide
at h = 0.5° (per audit memo §2 item 12: supralateral angular distance
changes only ~0.5° across h = 0–22°, so the two are observationally
indistinguishable on this single low-h photo). Visible feature sits
~4° below the theoretical line — most likely chromatic broadening of
the diffuse halo lower edge or visual-edge measurement bias (consistent
with the existing `_meta.note` on the original p27 anchor: "the flat
apex makes the fitted math apex sensitive").

**Anchor JSON re-shape (`docs/calibration/p27-anchor.json`):** top-level
`cza_apex` and `cza_apex_residual` fields moved under
`_disputed.cza_apex_legacy_2026_05_13` (preserved verbatim with
rationale, reversible). New top-level `supralateral_46halo_merger`
block with apex `(599, 142)`, `deg_above_sun = 41.89`,
`candidate_primitives = [halo_46_top, supralateral_arc]`. New
top-level `supralateral_route_eligibility` block flags the photo for
Pass A3 consideration. `_meta.method`, `_meta.status`, `_meta.note`
rewritten to record the re-classification.

**A3 disposition:** **CLEARED.** Pass A3 re-derives the CZA-route and
supralateral-route verdicts against the post-A2 eligibility set.

### Pass A3 Results — CZA + Supralateral Re-Verdict

Pass A3 deliverable from
[`../PHASE10_ATTACK_ROADMAP.md`](../PHASE10_ATTACK_ROADMAP.md), landed
2026-05-13. No code edits, no anchor edits — A3 is purely a verdict
+ doc-update pass against the now-honest atlas (A1b) and the
re-classified p27 (A2).

**CZA-route eligibility (all 8 anchored photos + p7 from the residual
table):**

| photo | h (°) | literature CZA above sun | predicted apex y | status |
| --- | ---: | ---: | ---: | --- |
| p2 | 18.6 | 46.14° / 381.7 px | **114.3** | **IN-WINDOW** (residual: +1.3 px) |
| p7 | 59.4 | n/a (CZA disappeared) | n/a | NOT APPLICABLE (h > 32.2°) |
| p13 | 6.83 | 51.88° / 497.6 px | −125.6 | OFF-FRAME (above top) |
| p20 | ~5.0 | 53.28° / 1102.1 px | −275.1 | OFF-FRAME (above top) |
| p22 | ~5.7 | 52.69° / 1210.7 px | −757.7 | OFF-FRAME (above top) |
| p25 | ~11.4 | 48.96° / 667.7 px | −481.7 | OFF-FRAME (above top) |
| p26 | ~9.0 | 50.36° / 739.9 px | −535.9 | OFF-FRAME (above top) |
| p27 | 0.5 | 57.31° / 570.5 px | −11.5 | OFF-FRAME *(also REMOVED Pass A2: visible feature reclassified)* |
| p30 | ~11.1 | 49.10° / 1450.1 px | −516.1 | OFF-FRAME (above top) |

**Only p2 is eligible.** Every low-h photo has the literature CZA apex
predicted above the top of the photo because CZA sits ~50–58° above
sun at low altitudes — typical sun-and-halo photos don't have that
much vertical headroom above the sun. p7 is past the disappearance
threshold. p27's visible feature was reclassified by Pass A2. The
single in-window photo (p2) has a sub-px residual (+1.3 px) under the
literature formula.

**CZA-route verdict: fails coverage gate.** This is a structurally
different failure from the pre-audit verdict ("fails residual gate on
expanded set; opposite-sign residuals on p2 and p27"). The pre-audit
verdict was an artifact of (a) the WB_R46 formula bug creating the p2
−19.3 px residual, and (b) the p27 chromatic arc being miscategorized
as CZA. With both corrections landed, the route has no residual-gate
failure on its own merits — it simply has only one in-window anchored
photo. Reopening coverage requires new anchors in 5° < h < 32°.

**Supralateral-route eligibility (post-A2):**

- p2: route-eligible per existing notes; no supralateral apex measured
  in `p2-anchor.json`.
- p27: route-candidate post-A2; apex measured at (599, 142) = ~41.89°
  above sun.

Strict reading of the coverage gate: 1 measured + 1 eligible-but-
unmeasured = below the two-photo threshold. Permissive reading: 2
candidates at threshold.

**Either reading is dominated by structural h-discrimination** (audit
memo §2 item 12). Supralateral angular distance from sun varies only
~0.5° across the h = 0–22° eligibility range:

- p2 (h = 18.6°) → ~45.7° above sun
- p27 (h = 0.5°) → ~46.0° above sun
- delta: **~0.3°** (~2.5 px at p2's R22)

Visual-edge measurement noise on a chromatic-broadened halo arc is
easily 5–10 px, which is the regime p27's own `_meta.note` flags
("flat apex makes the fitted math apex sensitive"). The route's
h-discrimination is below the noise floor on either photo.

**Supralateral-route verdict: fails coverage gate + structural-
discrimination rider.** Distinct from pre-audit ("coverage failure;
p27 might reopen it") because the structural-discrimination finding
says coverage *can't* fix it: even ten new low-h anchored photos
would not restore the route as a useful inverse handle, because
atmospheric physics constrains h-sensitivity below the noise floor.
This is a *route-physics* limitation, not a *dataset* limitation.

**Three-failure-layer framing update.** Pre-audit, the geometry-side
closeout headlined: *"the other three routes fail at three structurally
different layers of the measurement stack"* (residual / coverage /
detection). Post-A3, the count of *gates* is unchanged (CZA, supralateral
both coverage; tangent detection — that's 2 + 1, not 3 independent),
but the *failure modes* remain three structurally different things:

1. **CZA — coverage on physics grounds.** Route physically exists,
   forward formula is correct, p2 residual is sub-px. The single-photo
   coverage limit is an artifact of the photo set's altitude
   distribution and aspect ratio, not of the route or the atlas.
2. **Supralateral — coverage on discrimination grounds.** Route exists
   and is in-window across most of h, but h-sensitivity is structurally
   below the noise floor. Coverage cannot fix this; new physics could.
3. **Tangent — detection on tooling grounds.** Route remains blocked
   under column-peak detection on the post-C1 sampled set (p2 / p13 /
   p27). p7 is removed as circumscribed-halo regime, not upper-tangent
   route evidence. Pass C2 (optional) tests whether a wing-based
   detector recovers the route.

The mesa↔geometry crossover note's "two-of-three have non-atlas
explanations" hedge (per §6 step 2) now lands its substantive
content: of the three failures, the CZA finding is dataset-shape, the
supralateral finding is physics-shape, and the tangent finding is
tooling-shape. None of the three is a "the atlas inversion math is
broken" finding.

**Pass C1 cleared to proceed.** Drop p7 from the tangent eligibility
set per the circumscribed-halo literature regime. **Pass B2 cleared
to proceed** as soon as C1 lands, since B2 reads the now-finalized
post-A3 + post-C1 eligibility tables.

**Pass C2 disposition (optional).** Without C2, the tangent verdict
stays "detection gate failure under column-peak; protocol-conditional."
*(Post-C2 update 2026-05-14: the wing-radial Lab b\* ridge detector
with 22°-halo-radial-profile subtraction has now been built and run;
result is not-recovered on p2 / p13 / p27. The specialist handoff and
re-audit memo carry the post-C2 verdict. See ### Pass C2 Update later
in this file for the full receipt.)*

### Pass B2 Results — Parhelion-Route Re-Verdict

Pass B2 deliverable from
[`../PHASE10_ATTACK_ROADMAP.md`](../PHASE10_ATTACK_ROADMAP.md), landed
2026-05-14. No code edits, no anchor edits — B2 is the verdict pass
that closes Phase 10's technical work before the re-audit gate.

**Verdict: route survives promotion, language tightens.**

Audit memo §8 predicted this outcome: *"The single-handle verdict will
likely survive a re-audit but with the hedged language from §4 and §6
instead of the current closeout-level language."* B2 lands that exact
trade.

**Audit-survived wording (replaces "passes residual gate at ~0 px on
every eligible photo"):**

> *Parhelion-offset → h passes the residual gate at ~0 px on **three
> photos (p2 h = 18.6°, p7 h = 59.4°, p13 h = 6.83°)** with both
> unambiguous bilateral peaks and an independently fittable 22° halo.
> On five additional low-h anchored photos the geometric lever
> `sec(h) − 1` is below 2 % of R22, so clean residuals there are
> anchor-noise-bounded and informational rather than route-validating.
> On three of those (p20, p25, p26) the 22° halo arc is not visible
> in the photograph, so R22 was parhelion-derived and the residual
> measures `offset` against itself (tautological). p27 records
> `offset := R22` explicitly in its anchor JSON and is a
> self-consistency check, not a measurement. p26 right side encodes a
> literal geometric impossibility (`R22 / offset = 1.003 > 1`,
> `arccos` undefined) and is flagged `invalid` in the Pass B1 schema.*

**Per-photo verdict roll-up (against the Pass B1 eligibility sub-table):**

| photo | h (°) | lever (`sec(h)−1`) | R22-source | geom | parhelion residual | B2 reading |
| --- | ---: | ---: | --- | --- | ---: | --- |
| p2 | 18.6 | **5.52 %** | ring-fit | valid | −1 / −1 px | **eligible, route-validating** |
| p7 | 59.4 | **96.5 %** | ring-fit | valid | 0 / n/a px | **eligible, route-validating** |
| p13 | 6.83 | 0.71 % | ring-fit | valid | 0 / 0 px | **eligible** per audit memo §4.1 (unambiguous bilateral peaks + fittable halo); residual is below the noise floor, so informational not route-validating |
| p22 | ~5.7 | 0.50 % | ring-fit | valid | 0 / 0 px | informational (low-lever, anchor-noise-bounded) |
| p30 | ~11.1 | 1.90 % | ring-fit | valid | 0 / 0 px | informational (low-lever, just under 2 %) |
| p20 | ~5.0 | 0.38 % | parhelion-derived | valid | 0 / 0 px | **ineligible** (R22 tautological) |
| p25 | ~11.4 | 2.01 % | parhelion-derived | valid | 0 / 0 px | **ineligible** (R22 tautological) |
| p26 | ~9.0 | 1.25 % | parhelion-derived | **invalid right** | left: 0; right: undefined | **ineligible** (R22 tautological + geometric impossibility right) |
| p27 | ~0.5 | ≈ 0 % | ring-fit (parhelion-derived in opposite dir) | valid | 0 / 0 px | **ineligible** (parhelion offsets stipulated as `offset := R22`, tautological per memo §2 item 5) |

**Three-photo eligibility set** matches the audit memo §4.1
recommendation. Of those three: two (p2, p7) have meaningful geometric
lever; one (p13) has very low lever (0.71 %) and contributes via
"unambiguous bilateral peaks + ring-fit R22" rather than discrimination
per se.

**The pre-audit "every eligible photo" framing is retired** from this
file and from the Phase 10 closeout in `SUNDOG_V_GEOMETRY.md`.
Downstream public-framing surfaces (`SUNDOG_V_GRAVITY.md`,
`MESA_CROSSOVER_NOTE.md`, the homepage `#elevator-pitch` v1.1 audit
hedge) already landed walk-back hedges in attack roadmap §6 steps
2–4; B2's substantive content matches what those hedges anticipated.

**B2 closes the technical-pass wave.** Re-audit gate (§5) is now open:
all required passes (§6 hedges, B1, A1a, A1b, A2, A3, C1, B2) are
landed. C2 was the only optional pass and was skipped per the
recommended execution order. *(Post-C2 update 2026-05-14: Pass C2 has
now landed; the wing-radial Lab b\* detector with halo-radial
subtraction returned **not-recovered** on p2 / p13 / p27. The
"Unresolved Open Question" framing narrows to non-literature-standard
detector designs. See ### Pass C2 Update below for the full receipt.)*

### Parhelion-Route Per-Photo Eligibility

Pass B1 schema rollout, 2026-05-13. Reads
[`r22_source`](#) and [`geometric_validity`](#) from each photo's
anchor JSON; computes `sec(h) − 1` from the inferred altitude.
*Eligible* means **R22-source = ring-fit AND geometric_validity = valid
AND sec(h) − 1 ≥ 2 %**. The 2 % lever threshold comes from the audit
memo §2 item 4 — below it, residuals are at or below typical anchor
noise. Photos with ring-fit R22 but lever &lt; 2 % are *informationally*
useful but cannot route-validate the parhelion verdict on their own.

| photo | h (°) | sec(h) − 1 | R22-source | geom validity | parhelion residual | eligibility |
| --- | ---: | ---: | --- | --- | ---: | --- |
| p2  | 18.6 | **5.52 %** | ring-fit | valid | −1 / −1 px | **eligible** |
| p7  | 59.4 | **96.5 %** | ring-fit (anchor table; no JSON) | valid | 0 / n/a px | **eligible** |
| p13 | 6.83 | 0.71 % | ring-fit | valid | 0 / 0 px | low-lever caveat (anchor-noise-bounded; informational) |
| p20 | ~5.0 | 0.38 % | parhelion-derived | valid | 0 / 0 px | **ineligible** (R22 from parhelion; tautological) |
| p22 | ~5.7 | 0.50 % | ring-fit | valid | 0 / 0 px (presumed) | low-lever caveat (anchor-noise-bounded; informational) |
| p25 | ~11.4 | 2.01 % | parhelion-derived | valid | 0 / 0 px (presumed) | **ineligible** (R22 from parhelion; tautological) |
| p26 | ~9.0 | 1.25 % | parhelion-derived | **invalid (right)** | left: 0 px (presumed); right: undefined (`R22 / offset = 1.003 > 1`) | **ineligible** (R22 from parhelion AND right-side geometric impossibility per memo §2 item 6) |
| p27 | ~0.5 | ≈ 0.00 % | ring-fit (parhelion-derived in opposite direction) | valid | 0 / 0 px (tautological) | **ineligible** (parhelion offsets stipulated as `offset := R22`; memo §2 item 5: "the 0 px residual is tautological") |
| p30 | ~11.1 | 1.90 % | ring-fit | valid | 0 / 0 px (presumed) | low-lever caveat (anchor-noise-bounded; informational; just below 2 %) |

**Pass B1 conclusion against the audit memo's three-photo finding:**
the audit memo §2 item 4 + §4 item 1 says parhelion-offset passes "on
three photos (p2, p7, p13) where the test has meaningful discrimination
and an independently fittable 22° halo." This sub-table's *eligible*
column matches that finding for p2 and p7, and lists p13 as
*low-lever-caveat* rather than fully eligible because its lever is
0.71 % &lt; 2 %. The audit memo's looser 3-photo set hinges on
"unambiguous bilateral peaks AND independently fittable 22° halo"
without a hard lever threshold; this sub-table is the stricter
schema-level reading. **Pass B2 will rule on whether the parhelion
verdict survives on the strict (2-photo) eligible set, the audit's
permissive (3-photo) set, or somewhere in between with the low-lever
photos contributing as informational evidence.** Either way, the
"every eligible photo" framing is gone; the audit-survived
language (memo §6) is "passes residual gate at ~0 px on photos where
sun altitude provides a non-trivial geometric lever and an
independent 22° halo ring is fittable."

`r22_source` is now mirrored in each anchor JSON's top-level field of
the same name plus a `r22_source_note` rationale. p26's anchor adds a
`geometric_validity` block with per-side flags and the audit memo
citation.
| CZA apex → sun altitude / visibility | **+7.0 px / y: +1.3 px (literature; Pass A1b)** | p7 NOT APPLICABLE (h = 59.4° > 32.2° cutoff; CZA disappears); p13 (h = 6.83°) OFF-FRAME, predicted apex y = −125.6 (above top of photo) | p20 (h ≈ 5°), p22 (h ≈ 5.7°), p25 (h ≈ 11.4°), p26 (h ≈ 9°), p30 (h ≈ 11.1°): all OFF-FRAME (predicted apex y < 0 at every low-h photo because CZA sits ~50–58° above sun in the disappearance approach); **p27 REMOVED Pass A2** (visible feature is 46° halo top / supralateral merger, not CZA) | **fails coverage gate** *(Pass A3 verdict 2026-05-13; was: "fails residual gate on expanded set")* | **Pass A3 verdict, 2026-05-13.** Re-derived against the post-A1b atlas and post-A2 eligibility set. **Only p2 is in-window with an independent CZA-apex measurement**, and on that single photo the residual is **+1.3 px** — well below the 8 px / 0.04*R22 route threshold. The route does *not* fail the residual gate on its own merits; the pre-audit "opposite-sign residuals on p2 and p27" framing dissolved when (a) A1b corrected the atlas formula bug that was creating the p2 −19 px artifact, and (b) A2 re-classified the p27 chromatic arc as a different primitive. **The route fails the coverage gate**: the pre-registered "fewer than two eligible photos" rule binds because all other anchored photos are out-of-window (p7) or off-frame at the low altitudes where the CZA sits ~50–58° above sun. This is a structurally different failure mode from the pre-audit verdict — *not* a "route-reliability" finding but a "single in-window photo in the calibration set" finding. Reopening coverage would require either anchoring new photos in the 5° < h < 32° range where CZA stays on-frame, or accepting p2's sub-px residual as the single eligible result. Anchors: `p2-anchor.json` (in-window, sub-px); `p27-anchor.json` (re-classified per Pass A2); other six anchored photos all off-frame at literature CZA position. |
| Tangent-arc curvature → sun altitude | **detection-degenerate** (fuses with halo top at h=18.6°) | **p7 REMOVED Pass C1** (h=59.4° is circumscribed-halo regime, not upper-tangent-route eligibility); p13 (h=6.83°): chromatic-haze contamination, no clean smile to fit | p27 (h=0.5°): sun-bloom flare contaminates the sun-meridian column; p20 not measured | **partially recovered on p2 under manual sample selection; coverage gate fails; C5↔C6 substrate tension flagged** *(Pass C2 + C4 + C5 + C6 landed 2026-05-14)* | **Pass C2 + C4 + C5 + C6 verdict, 2026-05-14.** Pass C1 dropped p7. Three automated detectors fail on every photo: Pass C2 wing-radial Lab b\* ridge (`scripts/tangent_detector.py`, 0-2 / ~24 coherent samples); Pass C4 wing-slope luminance-gradient curvature (`scripts/tangent_curvature.py`, 88-100% weak-gradient rejection); **Pass C6 matched-filter against parameterized arc on halo-subtracted b\* (`scripts/tangent_matched_filter.py`), the natural follow-up the C5 receipt named** — negative correlation across the entire R\_uta scan on p2 (peak −0.117 at R\_uta = 263, outside tolerance). **Pass C5 (manual sample selection) recovered the route on p2** (R\_uta / R22 = 0.824, RMS = 1.23 px) with a methodology hedge (possible symmetry bias). On p13 only the apex was marginally anchorable; on p27 the sun-bloom column prevents any anchoring. The C5↔C6 contradiction on the same b\* substrate flags **two open interpretations**: (a) gestalt signal in a different substrate (untested: absolute b\*, L\* magnitude, chromaticity magnitude); (b) C5's tight fit is hand-anchoring symmetry-bias artifact. Recommended verify gate: specialist re-anchoring of p2. Coverage gate fails (1 / 3 photos). See ### Pass C6 Update and ### Pass C5 Update below for the full receipts. |
| Supralateral position → sun altitude | _measurement pending_ (route-eligible but no apex anchored) | p7 cropped; p13 not visible | **p27 measured Pass A2** with apex at (599, 142) = ~41.89° above sun (~4° below the ~46° literature prediction at h = 0.5°, likely chromatic-broadening); p20 weak/cropped | **fails on structural discrimination (Pass A3 verdict 2026-05-13)** *(coverage-gate hedged: 1 measured + 1 eligible-unmeasured)* | **Pass A3 verdict, 2026-05-13.** Post-A2 the eligibility set is p2 (eligible per existing notes, apex unmeasured) + p27 (measured at 41.89° above sun). On a strict reading the route fails the coverage gate (1 measured photo); on a permissive reading it's at the two-photo threshold. **Either reading is dominated by the structural-discrimination finding from audit memo §2 item 12.** Supralateral angular distance from sun varies only ~0.5° across the entire h = 0–22° eligibility range, an order of magnitude less h-sensitivity than parhelion-offset. The predicted h-spread between p2 (h = 18.6° → ~45.7°) and p27 (h = 0.5° → ~46.0°) is **~0.3°**, which at p2's R22 = 182 px corresponds to **~2.5 px**. Visual-edge apex measurement noise on a chromatic-broadened halo arc is easily 5–10 px (see p27's existing "flat apex makes the fitted math apex sensitive" flag in `_meta.note`). The h-spread the route can resolve is below the measurement noise floor on either photo, so **the route would not be a useful inverse handle even with perfect coverage**. Distinct from the pre-audit framing in two ways: (a) it is *not* "p27 might reopen coverage" — coverage may reopen but the route is still dead; (b) the structural limit is in the *atmospheric physics*, not in the dataset or the detection protocol. Reopening coverage on a third low-h photo will not restore the route. Anchors: `p2-anchor.json` (no supralateral apex), `p27-anchor.json` (`supralateral_46halo_merger` block). |
| Parhelic-belt y residual (primitive, not route) | not separately measured (Phase 2 set only had x-residuals) | p7 not applicable; p13 -10.4 px under the FF observed-minus-predicted convention | p27 -0.0 px; p22 +0.2 px; p26 -10.8 px; p30 -2.5 px; p25 +5.0 px (caveated) | **watch-list retired 2026-05-13** (FF1 falsified; FF3 passed) | Phase 10-FF tested the shared `--parhelic-y-offset-r22 = -0.05` rule across six low-h anchors. FF1 falsified: the >=5 px residuals do not share a sign (p13 / p26 negative, p25 positive) and Spearman rho is +0.086. FF3 passed: no photo breaches the 12 px primitive threshold. Record p13 as photo-specific / anchor-local, not a promoted low-h correction. See `PHASE10_BELT_Y_RESULTS.md`. |

## Tangent-Curvature v3.8 Receipt *(updated Pass C1 + Pass C2 + Pass C4 + Pass C5 + Pass C6 2026-05-14)*

Task #54 originally sampled p2 (h=18.6°), p7 (h=59.4°), p13
(h=6.83°), and p27 (h=0.5°). Pass C1 drops p7 from the
upper-tangent-route eligibility set: at h=59.4° the display is in the
circumscribed-halo regime, so testing it as an "upper tangent arc" was a
primitive misclassification. The post-C1 receipt is therefore narrower:
**column-peak detection fails on the sampled in-window / low-sun tangent
set (p2, p13, p27), while p7 becomes a high-sun circumscribed-halo
backlog reference.**

This is still a useful clean-negative for the current detector, but it is
not a class-level negative for tangent curvature. C2 is the optional
detector spike that decides whether a wing-based / Lab b* method recovers
the route.

Per-photo v3.8 schema:

| photo | h | salience | route-local residual | route-local sensitivity | full-overlay interaction | partition/conflict |
| --- | ---: | --- | --- | --- | --- | --- |
| p2 | 18.6° | high (b-contrast 35.7) | unmeasurable: samples hit halo top within 5 px | n/a — fit degenerate | depends on sun_x as expected | tangent fuses with 22° halo top at this h (compression regime) |
| p7 | 59.4° | high-sun circumscribed-halo display | **removed from upper-tangent-route eligibility by Pass C1** | n/a | n/a | At h>29° the upper/lower tangent arcs have joined into the circumscribed halo. Historical column-peak result is retained only as a cautionary sample, not as a tangent-route residual. |
| p13 | 6.83° | medium (b-contrast 15.8 = 23% sky) | RMS=41 px on saturation fit; radius 196 vs predicted 1774 | ±5 px wiggle → ±1.2 px Δr (misleadingly low; fit is already stuck) | n/a | chromatic-haze contamination; no clean smile to fit |
| p27 | 0.5° | tangent essentially absent between ~~CZA~~ *(retracted Pass A2)* and halo top | both fits have RMS=22 px with non-physical centers (brightness fit center *below* sun) | n/a | n/a | **Pass A2 2026-05-13 retracts "tangent merges with CZA":** at h = 0.5° the literature CZA apex is ~57° above sun while the upper tangent arc is ~22° above sun — separated by ~35°, they cannot merge optically (audit memo §4.4). The actual failure mode is sun-bloom flare contaminating the sun-meridian column. Substantively: tangent-route detection still fails at p27 (RMS=22 px, non-physical centers), but the failure cause is sun-bloom contamination, not feature merger. |

The Mesa #5 reading survives C1, but in a narrower form: visual salience,
route-local residual, sensitivity, and full-overlay interaction are **not
the same property**. For p2 / p13 / p27, the tangent primitive may be
atmospherically real and geometrically meaningful while remaining
route-local-unmeasurable under the column-peak detector. For p7, the
mistake was earlier in the stack: the photo belongs to the
circumscribed-halo regime, not the upper-tangent-route eligibility set.
The v3.8 schema's job was to record those distinctions; Pass C1 sharpens
that record instead of treating all four failures as one class.

Atmospheric-optics atlas-model implication: the upper tangent arc's
brightness/chromatic peak does NOT generally coincide with its
geometrically-predicted spine. The arc's *visual identity* lives in
combined brightness+chromaticity, not in a single peak that column-wise
detection can grab. Real measurement needs either gradient-based edge
detection (the spine is at a brightness or chromaticity *transition*, not a
peak) or manual sample selection from visual crops.

This receipt also weakens the earlier framing that p2 alone was the
exception. The post-C1 + post-C2 state is: CZA-apex fails coverage
after formula repair and p27 reclassification; supralateral fails on
coverage plus structural h-discrimination; tangent-curvature fails
detection under **two** literature-standard detectors (column-peak
intensity and wing-radial Lab b\* with halo-radial subtraction); the
tooling-conditional framing narrows to non-literature-standard
designs. Only parhelion-offset remains as a currently promoted
inverse route (Pass B2 verdict landed 2026-05-14).

### Pass C2 Update — Wing-Radial Lab b* Detector with 22°-Halo-Radial Subtraction *(landed 2026-05-14)*

Pass C2 was the optional detector spike that audit memo §4.8 left
open: build a literature-standard wing-based / Lab b\* ridge detector,
re-test the post-C1 sampled set (p2 / p13 / p27), and decide whether
the column-peak negative was *tooling-conditional* or *class-level for
the implemented detector family*.

**Implementation:** [`scripts/tangent_detector.py`](../../scripts/tangent_detector.py)
with regression runner
[`scripts/test_tangent_detector.py`](../../scripts/test_tangent_detector.py).
Captured run output: [`PASS_C2_DETECTOR_OUTPUT.txt`](PASS_C2_DETECTOR_OUTPUT.txt).

**Algorithm.**

1. RGB → CIE Lab via sRGB-gamma and D65 white-point matrix (vendored
   in `tangent_detector.py`; no skimage / cv2 dependency, only `numpy`
   and `PIL`).
2. Compute the 22° halo's azimuthally-symmetric chromatic profile as
   the per-integer-radius median of Lab b\* centered on the sun.
   Subtract this radial profile from the image — the result is the
   azimuthally-asymmetric Lab b\* residual that audit memo §4.8 /
   Persona 1 §4 item 6 explicitly recommends as the detector substrate.
3. For each wing (left, right), sample the residual b\* along the
   sun-radial direction at 24 azimuth-offset wing points,
   azimuth ∈ [15°, 75°] from the sun meridian. Locally-circular
   tangent approximation with R\_uta = R22 (matches
   `overlay_calibrate.py:462`). Radial profile window: ±18 px,
   37 samples per profile.
4. For each wing sample, the ridge candidate is the maximum residual
   b\* in the radial profile, with the offset reported in pixels from
   the predicted tangent locus.
5. **Pre-registered coherence gate:** a wing is *ridge-detected* if
   ≥ 8 / 24 samples carry amplitude ≥ 3.0 in residual b\* AND radial
   offset ≤ ±10 px. The amplitude threshold matches a "noticeably
   yellow above local baseline" chromatic ridge; the offset tolerance
   matches the route-residual threshold (0.04 × R22 ≈ 8 px for
   typical R22 values in this set).

**Per-photo results.**

| photo | h | left coherent / in-frame | right coherent / in-frame | left max b\* | right max b\* | photo verdict |
| --- | ---: | ---: | ---: | ---: | ---: | --- |
| p2  | 18.6° | 2 / 24 | 2 / 24 | +7.83 | +7.78 | **not-recovered** |
| p13 | 6.83° | 1 / 22 | 1 / 22 | +13.42 | +12.77 | **not-recovered** |
| p27 | 0.5°  | 0 / 24 | 2 / 24 | +3.64 | +3.91 | **not-recovered** |

(p13 has 22 in-frame samples per wing because outer-azimuth wing
points exit the photo frame at this geometry; the 22 / 22 in-frame
count is honest, not a count restriction.) Median radial offsets
across the six wings are −7.5 to +10.5 px — well inside the ±18 px
search window, not pinned at the boundary as they were in the
pre-halo-subtraction pilot run. The detector is finding ridge
candidates inside the window; the candidates simply do not cohere as
a continuous arc at the predicted locus.

**Verdict (audit memo §4.8 receipt template, resolved):**

> *Column-peak-on-sun-meridian-in-intensity-image fails on p2 / p13 /
> p27; a literature-standard wing-based Lab b\* ridge detector with
> 22°-halo-radial-profile subtraction does not recover the route on
> p2 / p13 / p27 either.*

This is a **stronger** negative than the pre-C2 "tooling-conditional
pending C2" framing in two directions:

1. The *second* literature-standard detector family also fails — the
   "column-peak is the wrong instrument" reading does not generalize
   to "any wing-based or chromatic-ridge detector would work."
2. The halo-radial-profile subtraction step explicitly addresses
   audit memo §4.8's methodological recommendation, so the failure
   cannot be attributed to the 22° halo's own b\* ridge contaminating
   the detection window. (The pilot run without subtraction *did*
   exhibit boundary-pinned offsets at ±18 px; halo-subtraction
   un-pinned them and the negative survived.)

It is **not** a class-level claim that the upper tangent arc lacks a
chromatic spine in principle. Significant chromatic features ARE
present in the wing-sample regions (max residual b\* amplitude 3.6 to
13.7 across the three photos), but they do not cohere at the
locally-circular R\_uta = R22 predicted locus with the pre-registered
amplitude × offset gate. The implementation falsifies Persona 1's
specific wing-based Lab b\* prediction (audit memo Appendix A §4 item
6) on this calibration set; it does not falsify all wing-based
detectors as a class.

**Open questions narrowed (filed as Phase 10 backlog, not Phase 10
deliverable):**

- Wing-slope **geometric** curvature detector (curvature of the
  apex-flanked smile, measured against the parameterized arc shape
  rather than against a chromatic ridge). Persona 1 §3 p2 entry
  explicitly recommended this for p2; C2 tested chromatic, not
  geometric.
- Matched-filter detection against a parameterized tangent-arc model
  (template-correlation), which would use the same geometry but
  exploit the full arc shape rather than per-sample ridge picks.
- Polarization-channel filtering — the upper tangent arc is
  partially polarized; a polarizer-equipped follow-up photo at
  controlled altitude (cf. `SUNDOG_V_PERCEPTION.md` Phase 1) would
  be a different *substrate* test, not a different detector on the
  same substrate.

**Implications for the failure-mode taxonomy (post-A3, post-C1):**

The "three structurally different failure modes" framing survives:
CZA = coverage (dataset / aspect-ratio); supralateral =
structural-discrimination (atmospheric physics); tangent = detection
tooling. C2 narrows the tangent mode: the failure persists across
*two* literature-standard detectors, so the residual open question is
about *non-literature-standard* detector designs (wing-curvature,
matched-filter, polarization), not "any wing-based method would
work." The route remains a candidate for recovery, but the
candidate-detector list is now more specific.

**Visibility-based promotion unchanged.** The upper tangent arc
remains a promoted **logo / animation vocabulary primitive** based on
visibility; only the inversion route remains blocked. See the
Vocabulary Classification table below.

### Pass C6 Update — Matched-Filter Detection Against Parameterized Arc Model *(landed 2026-05-14)*

Pass C6 was the natural follow-up to Pass C5: C5's manual sample
selection recovered the upper-tangent route on p2 (R\_uta\_obs / R22 =
0.824, RMS = 1.23 px), and the C5 receipt named matched-filter
detection as the highest-leverage next detector candidate. The
hypothesis was: **C2 (wing-radial Lab b\* ridge) and C4 (wing-slope
luminance-gradient) sampled per-point features and missed p2's signal;
C5 manual recovered at "whole-arc-shape" resolution; a matched-filter
that correlates a synthetic whole-arc template with the b\* residual
image should recover what C2 / C4 missed.**

**That hypothesis is falsified by Pass C6.**

**Implementation.**
[`scripts/tangent_matched_filter.py`](../../scripts/tangent_matched_filter.py)
with regression runner
[`scripts/test_tangent_matched_filter.py`](../../scripts/test_tangent_matched_filter.py).
Captured run output: [`PASS_C6_DETECTOR_OUTPUT.txt`](PASS_C6_DETECTOR_OUTPUT.txt).
The detector reuses Pass C2's halo-subtracted Lab b\* substrate
(via `tangent_detector.compute_halo_b_radial_profile` and
`halo_subtracted_b`).

**Algorithm.**

1. Compute the halo-subtracted Lab b\* residual image (same substrate
   as C2).
2. For each candidate R\_uta ∈ [0.5 × R22, 1.5 × R22] in 4 px steps
   (~45 candidates per photo), generate a 2D synthetic template: a
   Gaussian-thickened arc trace (σ = 4 px) along the predicted
   locally-circular upper-tangent locus, restricted to azimuth
   ∈ [0°, 75°] from the apex direction. Templates include the apex
   (different from C2 / C4 which excluded the inner strip).
3. Compute Pearson correlation between template and b\* residual over
   the template's support (pixels with template weight > 0.05).
4. Find the R\_uta with peak correlation. Report peak score, ratio
   R\_uta\_obs / R22, baseline mean (correlation over the other
   R\_uta values), and prominence ratio (peak / baseline mean).

**Pre-registered gates.**

- R\_uta\_obs / R22 ∈ [0.7, 1.3]
- Peak correlation score ≥ 0.10 (positive, above-noise correlation)
- Peak prominence ≥ 1.5× baseline mean

**Per-photo results.**

| photo | h | peak R\_uta | ratio | peak corr | baseline mean | prominence | verdict |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | --- |
| p2  | 18.6° | 263 px | 1.445 | **−0.117** | −0.185 | 0.00× (peak still negative) | **not-recovered** |
| p13 | 6.83° | 106 px | **0.502** | +0.191 | +0.093 | 2.05× | not-recovered *(prominence passes but R\_uta way below tolerance)* |
| p27 | 0.5°  | 258 px | 1.178 | −0.001 | −0.021 | 0.00× | **not-recovered** *(score essentially zero)* |

**The headline finding: the matched-filter hypothesis is falsified on
the halo-subtracted b\* substrate.**

- On **p2**, the correlation is **negative across the entire R\_uta
  scan** (from −0.244 at R\_uta = 91 to −0.117 at R\_uta = 267). There
  is no positive correlation anywhere — the template-vs-residual
  signal-coupling reverses sign in the wing region. The geometry C5
  hand-anchored on p2 (R\_uta\_obs = 149.9, ratio = 0.824) shows a
  matched-filter score of approximately −0.21 — the *opposite* sign
  from what a successful match would produce.
- On **p13**, matched-filter found a positive correlation peak (0.191)
  with 2.05× prominence, but at **R\_uta = 106 px = 0.502 × R22** —
  way below the predicted curvature. This is not the upper tangent
  arc; the matched-filter is picking up some other chromatic feature
  (possibly haze-edge structure or a parhelic-circle artifact at half
  the predicted radius).
- On **p27**, the score is effectively zero everywhere. The sun-bloom
  column dominates the upper region; no coherent chromatic structure
  matches the predicted upper-tangent template at any R\_uta.

**Two interpretations of the C5↔C6 contradiction:**

(a) **Substrate mismatch.** C5's manual identification picked up the
arc by *visual gestalt* (combining luminance, chromaticity, spatial
coherence, and likely some prior expectation of arc shape). C6
correlates only Lab b\* halo-subtracted residual. The arc's gestalt
signal may not be in the b\* channel alone — it could be in absolute
b\* (no subtraction), in L\* magnitude, in chromaticity vector
magnitude (`sqrt(a*² + b*²)`), or in some derived combination. C5
positive + C6 negative narrows the substrate space the route's signal
lives in; it does not rule out automated detection on a different
substrate.

(b) **Symmetry bias in C5.** The C5 receipt explicitly hedged: *"RMS
= 1.23 px is suspiciously tight given ±15 px per-point uncertainty,
possibly reflecting symmetry bias."* If the manual hand-anchoring
unconsciously favored points consistent with circular shape, C5's
clean fit would be partly artifact, and the route's "geometric
realness on p2" claim would weaken. C6's null result on the same
substrate is consistent with this reading — the matched-filter
doesn't share the symmetry bias and reports honestly.

Both interpretations are testable. (a) tests by running matched-filter
on different substrates (no halo subtraction; L\* magnitude;
chromaticity magnitude). (b) tests by specialist re-anchoring. Both
are filed as **Phase 10 backlog**.

**Implication for the failure-mode taxonomy.** The "hybrid coverage +
detection-tooling" classification from Post-C5 stands, but its
detection-tooling sub-claim now has a key qualifier:

> *Three automated detectors (C2 wing-radial Lab b\*; C4 wing-slope
> luminance-gradient; C6 matched-filter on halo-subtracted b\*) all
> miss p2's signal. Manual sample selection (C5) recovers on p2 but
> with a methodology hedge. Whether the recovery is a real positive
> awaiting the right automated substrate, or a hand-anchoring artifact
> from symmetry bias, is the remaining question.*

The single-handle verdict survives unchanged. Parhelion-offset is
still the sole *promoted* inverse handle. Tangent's verdict stays at
"partially recovered on p2 under manual sample selection; coverage
gate fails (1 / 3 photos); the manual positive is now in tension with
C6's automated negative on the same b\* substrate — recommend
specialist re-anchoring as the verify gate before propagating C5's
positive into public framing."

**Methodology note.** The C6 gates were pre-registered (R\_uta /
R22 ∈ [0.7, 1.3], peak corr ≥ 0.10, prominence ≥ 1.5×). All three
photos fail the gates. The negative is a clean pre-registered
result, not a tuning artifact. Future matched-filter passes on
alternative substrates would need their own pre-registration.

### Pass C5 Update — Manual Sample Selection from Visual Crops *(landed 2026-05-14)*

Pass C5 tests the **other** literature-standard alternative Persona 1
§5 explicitly named alongside gradient-based edge detection:

> *"Real measurement needs either gradient-based edge detection (the
> spine is at a brightness or chromaticity transition, not a peak) or
> manual sample selection from visual crops."*

C4 tested gradient-based edge detection (not-recovered). C5 tests
manual sample selection. The two together exhaust Persona 1's literal
list of literature-standard alternatives to peak-based ridge
detection.

**Implementation.** Hand-anchored (x, y) point lists added to the
`upper_tangent_manual_samples` block of each of the post-C1 eligibility
photos' anchor JSONs (`p2-anchor.json`, `p13-anchor.json`,
`p27-anchor.json`). Runner at
[`scripts/test_tangent_manual.py`](../../scripts/test_tangent_manual.py)
loads the manual points and applies `tangent_curvature.fit_circle` (the
same linearized least-squares circle fit Pass C4 used). Captured run
output: [`PASS_C5_DETECTOR_OUTPUT.txt`](PASS_C5_DETECTOR_OUTPUT.txt).

**Pre-registered gates** (different from C4 — fewer manual points are
expected than per-azimuth automated candidates, and each manual point
is selected with visual care, so the gates relax candidate count and
tighten RMS):

- ≥ 5 surviving manual points per photo
- circle-fit RMS ≤ 10 px
- R\_uta\_obs / R22 ∈ [0.7, 1.3]

**Hand-anchoring procedure.** Each photo was read at original display
resolution and the visible upper-tangent chromatic feature was
identified by eye. Apex was anchored at the point where the chromatic
dome touches the 22° halo crown; wing points were anchored at the
visible spine of the chromatic dome at approximately ±25° and ±50°
from the apex. Per-point uncertainty estimated at ±15 px. The
anchor-JSON `upper_tangent_manual_samples.method` field records the
methodology per photo and explicitly flags photos where anchoring was
not possible.

**Per-photo results.**

| photo | h | n manual points | fit center | R\_uta\_obs | R\_uta / R22 | RMS | verdict |
| --- | ---: | ---: | --- | ---: | ---: | ---: | --- |
| p2  | 18.6° | 5 | (567.5, 160.3) | 149.9 px | **0.824** | 1.23 px | **manual_recovered** |
| p13 | 6.83° | 1 (apex only) | n/a | n/a | n/a | n/a | not-recovered (insufficient samples) |
| p27 | 0.5°  | 0 | n/a | n/a | n/a | n/a | not-recovered (sun-bloom contamination; no anchorable points) |

**Two structurally distinct findings:**

1. **On p2, the arc is geometrically real and circular.** Five
   hand-anchored points trace a circle with **RMS = 1.23 px** — far
   below the 10 px tolerance, and below the typical noise floor for
   visual edge measurement on a chromatic-broadened halo arc. The
   fitted radius R\_uta\_obs = 149.9 px is **82.4% of R22**, well
   inside the [0.7, 1.3] tolerance band. This is a clean *positive*
   result on a single photo: the upper tangent arc on p2 sits where
   the locally-circular literature approximation predicts, with
   measurable circular curvature.

2. **The signal that p2 carries is not what C2 / C4 measured.** Pass
   C2 (wing-radial Lab b\* with halo-radial subtraction) found 0-2/24
   coherent ridge candidates per wing on p2; Pass C4 (wing-slope
   luminance-gradient circle-fit) rejected 42/48 wing samples as
   weak-gradient. Both detectors operated on the same wing region the
   manual anchoring used and missed the arc. The conclusion is that
   **the upper tangent arc's signature on p2 is at the resolution of
   visual chromatic-dome identification rather than at the resolution
   of per-sample radial-profile peak picks** — a substantive update
   to the C2 / C4 detector-design space, not a substrate failure.

3. **On p13 and p27, even manual anchoring fails.** p13 is washed
   into a luminous chromatic gradient at the halo crown with no
   localizable wing spine (Persona 1 §3 p13 entry); only the apex
   could be marginally anchored. p27 has the sun-bloom column
   dominating the halo-crown region; no upper-tangent feature is
   separable. The "tooling-protocol" framing collapses on these
   photos to a **substrate-signal-availability** finding: the arc
   carries insufficient image information for *any* detector
   (automated or manual) to localize.

**Coverage gate.** Pre-registered ≥ 2 eligible photos rule for route
promotion. Manual sample selection recovers on **1 / 3** photos.
**Coverage gate fails.** Route promotion remains blocked.

**Failure-mode taxonomy update.** The pure "detection-protocol
tooling" classification from the C2 + C4 receipts is now refined into
a **hybrid coverage + detection-tooling failure**:

- **On p2:** detection-tooling failure (signal present, three
  automated detectors miss it; manual anchoring recovers).
- **On p13 / p27:** substrate-signal-availability failure (signal too
  weak / contaminated for any detector — automated or manual).

This is structurally different from the pre-C5 framing in two ways:
(a) the route is no longer presumed to fail uniformly across the
sampled set; p2's signal is real and circular, (b) future detector
improvements (matched-filter, smarter ridge picks) could plausibly
recover the route on p2-like photos, but **coverage expansion is
needed regardless** because p13 / p27 don't carry enough substrate
signal even for manual measurement.

**Implications for the open-questions list.** The remaining tangent
recovery candidates are now narrower:

- **Matched-filter detection against a parameterized arc model.**
  Still open. The p2 finding (signal at visual-dome resolution, not at
  per-sample-peak resolution) makes matched-filter MORE promising
  than C2 / C4: a template-correlation detector would test the
  whole-arc-shape hypothesis the manual anchoring confirmed.
- **Polarization-channel filtering.** Different substrate; still
  open.
- **New calibration photos with stronger tangent display.** Most
  load-bearing for coverage expansion: even with a perfect detector,
  the route needs ≥ 2 photos with anchorable signal, and the current
  set has 1.
- **Manual sample selection (now tested).** Confirmed as
  literature-standard recovery on p2; insufficient coverage on
  current calibration set.

**Single-handle verdict unchanged.** Parhelion-offset remains the
sole *promoted* inverse handle. Tangent-arc curvature → h is now
classified as **"partially recovered (1 / 3 photos under manual
sample selection); coverage gate fails"** rather than "fails under
three literature-standard detectors." The route's status moves from
"blocked by detector tooling" to "blocked by hybrid
coverage-plus-tooling; signal confirmed present on the cleanest
display, absent on the two contaminated ones." This is a meaningful
weakening of the route's verdict in a constructive direction — the
arc is now known to be measurable in principle, just not on enough of
the current set.

**Methodology hedge.** The p2 manual anchoring was performed by
visual identification at display resolution by claude assistant, with
per-point uncertainty estimated at ±15 px. The RMS = 1.23 px fit
result is suspiciously tight given that uncertainty and could
partially reflect symmetry bias (the visual identification favored
points consistent with circular shape). A specialist hand-anchoring
the same photo could confirm or weaken the result; recording the
points in the anchor JSON makes that re-verification a one-script
operation.

### Pass C4 Update — Wing-Slope Geometric Curvature Detector *(landed 2026-05-14)*

Pass C4 was the immediate follow-up that the Pass C2 receipt above named
as the first non-literature-standard detector candidate: a wing-slope
geometric curvature detector, picking up Persona 1 §3 p2 entry's specific
recommendation ("a curvature detector applied to the wings would work")
and §5's broader call ("real measurement needs either gradient-based
edge detection ... or manual sample selection from visual crops"). C2
tested chromatic ridge; C4 tests geometric edge.

**Implementation:** [`scripts/tangent_curvature.py`](../../scripts/tangent_curvature.py)
with regression runner
[`scripts/test_tangent_curvature.py`](../../scripts/test_tangent_curvature.py).
Captured run output: [`PASS_C4_DETECTOR_OUTPUT.txt`](PASS_C4_DETECTOR_OUTPUT.txt).

**Algorithm.**

1. RGB → CIE Lab (vendored, reuses `tangent_detector.rgb_to_lab`).
2. For each wing (left, right) and each azimuth in [15°, 75°] from the
   sun meridian (24 samples per wing), sample CIE L\* along the
   sun-radial direction at the predicted wing point. Profile window:
   ±18 px, 37 samples.
3. Compute the radial gradient ∂L\*/∂r via central difference and
   record the peak gradient location per profile — Persona 1 §5's
   gradient-based edge detection.
4. Filter to candidates with (a) gradient magnitude ≥ 1.5 L\*/px (a
   visually clear radial brightness transition) and (b) radial offset
   ≤ ±12 px from the predicted tangent locus.
5. Fit a circle to the surviving (x, y) edge candidates via linearized
   least-squares (x² + y² + Dx + Ey + F = 0). Fitted radius is
   R\_uta\_obs.
6. Pre-registered "curvature-recovered" gates: ≥ 12 surviving candidates
   AND circle-fit RMS ≤ 8 px AND R\_uta\_obs / R22 ∈ [0.7, 1.3].

**Per-photo results.**

| photo | h | candidates kept | weak (< 1.5 L\*/px) | far (> ±12 px) | out-of-frame | fit | verdict |
| --- | ---: | ---: | ---: | ---: | ---: | --- | --- |
| p2  | 18.6° | 2 / 48 | 42 | 4 | 0 | too few points | **not-recovered** |
| p13 | 6.83° | 4 / 48 | 38 | 2 | 4 | too few points | **not-recovered** |
| p27 | 0.5°  | 0 / 48 | 48 | 0 | 0 | singular | **not-recovered** |

The "weak gradient" rejection rate (88-100% of wing samples) is the
load-bearing observation. The wing region of every photo simply does
not carry a radial luminance transition strong enough to outline the
tangent arc at the L\* ≥ 1.5/px scale. p27's 48/48 weak result is
consistent with Persona 1 §3 p27 entry: at h = 0.5° the wing region is
washed by sun-bloom and forward-scatter haze; no clean L\* edge survives.
p2 and p13 carry a handful of strong-gradient candidates but not enough
to fit a circle, and the offsets-vs-azimuth pattern doesn't trace a
coherent arc.

**Verdict.** The wing-slope geometric curvature detector also returns
**not-recovered** on the post-C1 sampled set. Combined with C2:

> **Three literature-standard detectors have now been tested on
> p2 / p13 / p27 and all return not-recovered:** column-peak intensity
> at sun meridian (the original Phase 10 detector, Pass C1 framing);
> wing-radial Lab b\* with 22°-halo-radial-profile subtraction
> (Pass C2); wing-slope luminance-gradient edge detection with circle
> fit (Pass C4 — the gradient-based edge detector Persona 1 §5
> explicitly named as the literature-standard alternative).

This is materially stronger than the C2 negative: C2 left the
gradient-based-edge-detection family untested; C4 closes that
sub-question. The remaining tangent-route open questions are
substantially narrower:

- **Manual sample selection from visual crops** (Persona 1 §5
  alternative). This is hand-anchoring the tangent on each photo, not
  an automated detector; would supply curvature data but not an
  inverse-route detector.
- **Matched-filter detection against a parameterized arc model.**
  Template correlation across the wing ROI; different signal
  (whole-shape matching) than per-sample edge or ridge picks.
- **Polarization-channel filtering.** Different *substrate* — requires
  polarizer-equipped follow-up photos at controlled altitude, lining
  up with [`../SUNDOG_V_PERCEPTION.md`](../SUNDOG_V_PERCEPTION.md)
  Phase 1.
- **New calibration photos with stronger tangent display** (e.g.,
  h ∈ [10°, 25°] with high b\* contrast). Coverage expansion, not
  detector design.

All four remain Phase 10 backlog. None is gated for Phase 11 work.

**Methodology hedge.** The 1.5 L\*/px gradient threshold was
pre-registered without prior empirical calibration of the wing-region
noise floor. The "weak" rejection rate (88-100%) suggests the threshold
sits well above whatever radial-luminance noise is present; the
sensitivity is real, not a calibration artifact. If a future detector
characterizes the wing-region L\*-gradient noise floor and re-runs at
a noise-relative threshold, the answer may still be the same — but the
methodology will be cleaner.

**Implications for the failure-mode taxonomy.** The "three structurally
different failure modes" framing (CZA: coverage; supralateral:
physics; tangent: tooling) survives. C4 further narrows the tangent
mode: three literature-standard detectors fail across two distinct
signal modalities (chromatic ridge in C2; luminance edge in C4). The
remaining open question is no longer "have we tried the right
detector?" — it is "is the upper tangent arc's image signature
inverse-route-measurable in principle on calibration sets like this
one, or does this require either denser hand-anchoring or new
substrate data?"

**Single-handle verdict unchanged.** Parhelion-offset remains the sole
promoted inverse handle on the strict 3-photo subset. Tangent route
stays blocked. Visibility-based upper-tangent promotion as logo /
animation vocabulary is unchanged.

## Vocabulary Classification

| primitive | p2 | p7 | p13 | promote? |
| --- | --- | --- | --- | --- |
| CZA | visible | not applicable / high sun | cropped or not visible | Conditional core: render only inside the CZA sun-altitude validity window; CZA-apex inversion route fails coverage after A1b/A2/A3, not residual. |
| Supralateral arc | visible | candidate only | not visible / cropped | Keep as optional vocabulary; p27 is now a measured 46°-halo-top / supralateral-merger candidate, but the inversion route fails structural h-discrimination. |
| Upper tangent arc | visible | high-sun circumscribed-halo regime, not upper-tangent-route evidence | candidate | Promote as stable logo/animation shape language from visibility; curvature-as-inversion-route partially recovered on p2 via Pass C5 manual sample selection (R\_uta\_obs / R22 = 0.824, RMS = 1.23 px), coverage gate fails at 1 / 3 photos; three automated detectors (C2/C4 plus original column-peak) miss the signal that manual recovers; remaining candidates are matched-filter detection (now more promising), polarization filtering, or new calibration photos. |
| Lower tangent arc | visible near lower 22° contact | not clear | not clear | Annotation only until another clean low-sun source supports it. |
| Suncave Parry arc | visible/candidate using image 1 key | weak candidate | weak candidate | Candidate label only; useful for education, not logo default. |
| Parry supralateral arc | candidate in image 1 | not visible | not visible | Do not promote beyond optional annotation. |
| Infralateral arcs | visible on p2 periphery | not visible / cropped | not visible / cropped | Good vocabulary label; not a core logo layer unless the design brief wants a rich-display variant. |

## Phase 10 Promotion Verdict *(landed 2026-05-13)*

Verdicts apply the pre-registered thresholds above. Status is one of:
**promoted** (moves into calibrated core / logo defaults), **promoted as
optional vocabulary** (annotation layer, not core), **fails gate** (held
back; reason recorded), or **pending** (cannot be promoted until measured).

| candidate | verdict | reason |
| --- | --- | --- |
| Parhelion offset to h inversion route | **promoted (post-audit hedged; Pass B2 2026-05-14)** | **Pass B2 verdict 2026-05-14:** Route survives the post-B1 eligibility re-derivation, with the language hedged per audit memo §4.1 and §6. Audit-survived wording: *passes residual gate at ~0 px on three photos (p2, p7, p13) with both unambiguous bilateral peaks and an independently fittable 22° halo.* On five additional low-h anchored photos the geometric lever is below 2 % of R22; on three of those (p20, p25, p26) the 22° halo arc is not visible so R22 was parhelion-derived (tautological); p27 explicitly stipulates `offset := R22`; p26 right encodes a geometric impossibility (Pass B1 flagged `invalid`). The pre-audit "~0 px on every eligible photo" framing is retired. The route stays in calibrated core; verdict now reads as 'route holds on the strict eligible subset, anchor-noise-dominated below the 2 % lever.' Original Task #52 step 1 finding ('bad anchor, route OK' after re-anchoring p13 from (557, 372) / R22=210 / offset=220 to (543, 372) / R22=211 / offsets=213/212; x-offset residual dropped to ~0/0 px) survives unchanged. |
| 22 deg halo, parhelia (L/R), parhelic-circle | **promoted** (calibrated core) | Already in core via Phase 2. No regression observed in Phase 10 inspection. |
| Upper tangent arc | **promoted as logo / animation vocabulary** | Visible shape-language support remains; p7 is now treated as high-sun circumscribed-halo vocabulary rather than upper-tangent-route evidence. Curvature-as-inversion-route partially recovered on p2 via Pass C5 manual sample selection (R\_uta\_obs / R22 = 0.824, RMS = 1.23 px); three automated detectors (C2/C4 plus original column-peak) miss the signal manual recovers; coverage gate fails at 1 / 3 photos; remaining candidates are matched-filter detection (now more promising given p2's circular fit), polarization filtering, or new calibration photos. |
| CZA primitive (rendered) | **promoted as conditional core** | Render only when `h < 32.2 deg` (atmospheric cutoff). Visible and measured on p2; p27's former CZA mark is reclassified as 46° halo top / supralateral merger. The primitive remains core while the CZA-apex inversion route fails coverage. |
| CZA apex to h inversion route | **fails coverage gate** *(Pass A3 2026-05-13)* | A1b fixed the CZA formula and p2 now has a +1.3 px y residual, below threshold. A2 removed p27 from CZA evidence. The remaining eligible set has only p2; all other anchored photos are out-of-window or off-frame at the literature CZA position. |
| Tangent-arc curvature to h inversion route | **partially recovered on p2 under manual sample selection; coverage gate fails (1 / 3 photos); C5↔C6 substrate tension on b\* residual** *(Pass C1 + C2 + C4 + C5 + C6 landed 2026-05-14)* | Pass C1 removed p7. Passes C2 (wing-radial Lab b\* ridge), C4 (wing-slope luminance-gradient), and **C6 (matched-filter on halo-subtracted b\* — the natural follow-up the C5 receipt named)** all return not-recovered on every photo. **Pass C5 (manual sample selection)** recovered the route on **p2** (R\_uta / R22 = 0.824, RMS = 1.23 px) — with methodology hedge (possible symmetry bias). C6 specifically falsified the C5→matched-filter natural-extension: on p2 the correlation is negative across the entire R\_uta scan, including at R\_uta values consistent with C5's manual fit. The route's recovery on p2 is now in **C5↔C6 substrate tension**: either the gestalt signal C5 picked up lives in a different substrate (not halo-subtracted b\*), or C5's tight fit is hand-anchoring bias. Recommended verify gate: specialist re-anchoring of p2 before further propagation. Coverage gate still fails (1 / 3 photos). Remaining open paths: matched-filter on alternative substrates (absolute b\*, L\* magnitude, chromaticity magnitude — untested); polarization filtering; new calibration photos for coverage expansion. Promotion remains blocked. |
| Supralateral position to h inversion route | **fails structural-discrimination gate** *(coverage hedged; Pass A3 2026-05-13)* | Post-A2/A3, p27 is measured as a supralateral / 46°-halo-top candidate and p2 is eligible but unmeasured. Even if coverage is treated permissively, the predicted p2-vs-p27 h-spread is ~0.3° (~2.5 px at p2 R22), below visual-edge noise. The route is not a useful inverse handle on atmospheric-physics grounds. |
| Lower tangent arc | **promoted as low-sun annotation only** | Visible at p2's lower 22 deg contact; not clear on p7 or p13. One photo of evidence -- fails the two-photo bar for core promotion. Held as low-sun-display annotation; compatible with `low-altitude.json` named-pose examples, but the pose is not additional evidence. |
| Suncave Parry arc | **fails gate** (weak evidence) | Weak/candidate on all three Phase 10 photos. Below the visibility bar for inversion measurement; insufficient for shape-language promotion. Stays in atlas vocabulary as an educational label only. |
| Parry supralateral arc | **fails gate** (weak evidence + coverage) | Candidate on image 1 key only; not visible on p7 or p13. Promotion blocked on coverage. Vocabulary-label-only status confirmed. |
| Infralateral arcs | **promoted as optional vocabulary** (not core) | Visible on p2 periphery; not clear on p7 / p13. Single-photo evidence on the Phase 10 set, but image 1 key labels it. Held as a rich-display vocabulary option; compatible with `forty-six-halo.json` named-pose examples, but the pose is not additional evidence. |
| Any linear arc-importance score (e.g. "arc X = 23% of fit") | **fails gate** (per Mesa crossover rule 5) | Linear additive attribution of fit is mesa's documented failure mode for field-shaped objects. Reject any Phase 11 metric that produces this. |

### Coverage and threshold summary

Pre-registered gate (from the Phase 10 Measurement Pre-Registration section):

- **Visibility coverage**: a primitive or route needs >= 2 eligible photos
  (visible, uncropped, in sun-altitude validity window) to be a candidate.
- **Primitive residual**: >= 0.06 * R22 **or** >= 12 px on >= 2 eligible
  photos -> do not promote.
- **Inversion-route residual**: >= 0.04 * R22 **or** >= 8 px on >= 2
  eligible photos -> do not promote.
- **Attribution form**: any linear "arc X contributes N%" metric -> reject
  outright.

### Do-not-promote list

Concrete output of the Phase 10 gate, in the form the roadmap asks for:

1. **Suncave Parry as logo / animation shape language.** Weak evidence
   across the entire Phase 10 set. Stays in atlas vocabulary as an
   educational label.
2. **Parry supralateral as anything beyond an optional annotation.**
   Coverage fails -- candidate on image 1 only.
3. **Supralateral inversion route on the current committed evidence.**
   Pass A3 records a coverage hedge plus structural-discrimination
   failure. p27 is now a measured supralateral / 46°-halo-top candidate,
   but the route's h-signal is below visual-edge noise even if coverage
   is read permissively.
4. **Tangent-arc curvature inversion route under the column-peak detection
   protocol.** Detection-degenerate on the post-C1 sampled set
   (p2 / p13 / p27). p7 is removed from route eligibility because it is
   circumscribed-halo regime. Reopening the route requires C2's
   wing-based / Lab b* detector, not simply more photos.
5. **CZA-apex inversion route on the current committed evidence.** Pass
   A3 converts the verdict from residual-gate failure to coverage-gate
   failure: p2 has a good +1.3 px residual under the repaired formula,
   p27 was a misidentified primitive, and all other anchors are
   out-of-window or off-frame.
6. **Lower tangent as core logo geometry.** Single-photo evidence; usable
   only as labeled low-sun annotation or in named-pose examples, not in core
   marks.
7. **Any linear arc-importance attribution metric** at Phase 11 review.
   Per Mesa crossover rule 5.

### Single-handle closeout

The Phase 10 gate closed 2026-05-13 and was post-audit hedged by Passes
B1, A1a/A1b, A2, A3, C1, and B2. **One inversion route remains promoted;
the other candidate inversion routes fail or remain blocked for distinct,
now better-named reasons.**

| route | gate outcome | failure layer |
| --- | --- | --- |
| Parhelion offset → h | **promoted (post-audit hedged; Pass B2 2026-05-14)** | Eligibility restricted to photos with ring-fit R22, valid geometry, and non-trivial lever; ~0 px residual on the three audit-survived eligible photos (p2, p7, p13). Low-lever / parhelion-derived / geometrically-invalid photos contribute as informational evidence only. |
| CZA apex → h | fails | **coverage gate** after A1b/A2/A3: p2 residual is good (+1.3 px), but it is the only in-window measured CZA anchor. |
| Supralateral → h | fails | **coverage + structural-discrimination gate** after A2/A3: p27 is measured and p2 is eligible, but h-sensitivity is below measurement noise. |
| Tangent-arc curvature → h | fails / unresolved | **detection gate under column-peak** on p2 / p13 / p27 after C1 removes p7 as circumscribed-halo regime; C2 decides whether this is tooling-conditional or recovered. |

The old "three independent failure layers" language is retired. The
post-audit taxonomy is more precise: CZA is dataset/aspect-ratio
coverage-limited, supralateral is physics-discrimination-limited,
tangent is detector-limited under the current column-peak protocol, and
parhelion offset is the only currently promoted inverse handle, with B2's
final wording now landed.

Phase 11 can proceed against the visibility-promoted vocabulary now.
Three open questions are filed but not blocking:

1. **CZA coverage expansion.** Would a controlled photo set in
   5° < h < 32° supply a second in-window CZA apex anchor? The old
   "direction inconsistency" question was dissolved by A1b/A2.
2. **Tangent-curvature tooling.** Does a gradient-based edge-tracking or
   template-matching detector recover the route on existing photos? This
   is a tooling question, not a physics question — the arc is real and
   visible; the column-peak method is the wrong instrument.
3. **Parhelic-belt-y replication.** Closed by Phase 10-FF:
   p13's residual is photo-specific; the belt-y watch-list flag is
   retired.

Each is scoped well enough to file as a separate task when attacked. None
gate Phase 11.

## Anchor-Capture Closeout (Tasks #52, #53, #54, #55)

All four anchor-capture tasks closed 2026-05-13. The execution-order
discipline below is preserved as historical record; outcomes are recorded
inline.

Closeout summary:

- **Task #52 step 1 — re-anchor p13 parhelion offset:** done. p13
  re-anchored to (sun=543,372; R22=211; offsets=213/212), dropping the
  x-offset residual to ~0/0 px and revealing p13 is a low-altitude photo
  at h ~= 6.83°, not the mid-altitude h = 17.3° the rough hand-anchor
  implied. Parhelion-offset route probation cleared with verdict "bad
  anchor, route OK." The p13 belt-y residual (+10.4 px) is tracked
  separately under the Per-Inversion-Route Residual Table.
- **Task #53 / #55 — CZA apex anchors:** done. ~~p2 residual y = -19.3 px;
  p27 residual y = +21 px. Both exceed the `0.04 * R22` / 8 px route
  threshold, opposite signs. CZA-apex inversion route fails the residual
  gate on the expanded set; not a stable systematic.~~
  **Walked back 2026-05-13 (Pass A1b + Pass A2 + Pass A3):** the p2
  residual was formula-bug contaminated (legacy `WB_R46` hardcode);
  literature formula collapses it to y = +1.3 px under predicted-minus-
  observed. The p27 residual was against a misidentified primitive (the
  visible feature at y = 142 is the 46° halo top / supralateral
  merger, not CZA). Pass A3 re-derived the CZA-route verdict: **fails
  coverage gate** (only p2 is in-window with an independent measurement;
  every other anchored photo is out-of-window or off-frame). The p2
  residual is sub-px; the route has no residual-gate failure on its own
  merits.
- **Task #54 — tangent-arc curvature sampling:** done. Column-peak
  detection fails on the post-C1 sampled set (p2 / p13 / p27), with p7
  removed from upper-tangent-route eligibility as a circumscribed-halo
  display. Historical p7 sampling is retained only as a primitive-ID
  caution. p27's failure mode is ~~CZA / sun-bloom merge~~ **sun-bloom
  flare contaminating the sun-meridian column** (Pass A2 retracts the
  "merges with CZA" framing — at h = 0.5° CZA and tangent are ~35°
  apart in altitude, not coincident; per audit memo §4.4). See the
  Tangent-Curvature v3.8 Receipt section above. Tangent-arc curvature
  inversion remains a *protocol-conditional* negative pending Pass C2
  wing-based detector rebuild.
- **Supralateral inversion:** ~~coverage gate failure on the current
  committed set (only p2 eligible). p27 remains a candidate but no
  supralateral residual has been measured~~ **Pass A3 verdict 2026-05-13:
  fails coverage gate + structural-discrimination rider.** Pass A2
  promoted p27 from CZA-misidentified to supralateral candidate (apex
  measured at ~41.89° above sun). p2 is route-eligible but apex
  unmeasured. **Even at the two-photo threshold the route fails on
  structural h-discrimination:** audit memo §2 item 12 documents that
  supralateral angular distance from sun varies only ~0.5° across
  h = 0–22°, an order of magnitude less than parhelion-offset. The
  predicted spread between p2 and p27 is ~0.3° = ~2.5 px at p2's R22,
  below the typical 5–10 px visual-edge measurement noise. The route
  would not be a useful inverse handle even with perfect coverage; this
  is an atmospheric-physics limitation, not a dataset limitation.

Historical execution order (now retired):

1. **Re-anchor p13 parhelion offset first** — done. Replaced rough
   hand-anchor with measured (sun_px, R22_px, parhelion_left_px,
   parhelion_right_px) tuple; new residual ~0/0 px. Anchor Summary and
   Per-Inversion-Route Residual Table updated.
2. **CZA apex anchor check (Task #53, expanded by Task #55).** Historical
   result was p2 y = -19.3 px