### BINI78 — Nelson Island
**CRD page:** [Blenheim, Nelson Island & Victory Bank Coral Surveys (Apr 2015)](https://coralreefdiagnostics.com/new-page-4)
**Date:** 2015-04-20 (dive window 10:20:18–11:07:00, 46 min 42 s, from `DateTimeOriginal` EXIF on all 16 files, Olympus TG-2. The counter runs P4200892→P4200930, so **the edited set keeps 16 of 39 exposures** — a frame selection upstream of every photo-derived number below, and at this site the selection is unusually explicit: see flag 12.)
**GPS:** 5.68049°S, 72.38146°E · **Depth strata:** D2 (n=2 transects) and D3 (n=7 transects), 9 transects total
**Frames reviewed:** 16 of 16 (100% of the site folder, `BiNi78-no good pics-finished`; 0 evicted, 0 EPERM, and **all 16 carry CV data — complete coverage, nothing lost to the filename collision**)
**Coral cover:** 19.11% (formal transect surveys; algal cover 47.68%, bare substrate 7.79%, invertebrates 25.43%) · **% bleached:** 11.57% (`% bleaching (2nd training)`, Mayfield & Dempsey 2026 supplement)
**Reef type:** submerged lagoonal bank reef, inside the Nelson Island lagoon, protected · **Temperature:** 29.4 °C · **Salinity:** 33.7
**Live hard coral cover across methods** (n=16 photos of 16): transect 19.107% · CoralNet points 17.708% · Coralscapes mask 18.428%. Deviation vs transect **−1.40 and −0.68 pp — the 6th smallest gap of 98 sites**, and a point/transect ratio of **0.927 against a corpus median of 0.650** (12th highest of the 90 sites with a transect value and ≥5 frames). **This is the closest three-method agreement in the Chagos set so far, and flag 1 is the reason it should not be read as accuracy.**
**Coral genera identified:** the transect recorded **14 scleractinian genera** across 9 transects — *Porites* 4.36%, *Acropora* 3.93%, ***Stylophora* 3.75%** (unusually high, third-ranked), *Pocillopora* 2.25%, *Galaxea* 1.00%, *Favia* 0.93%, *Platygyra* 0.86%, *Montipora* 0.79%, then a tail through *Gardineroseris* 0.36%, *Psammocora* 0.32%, *Echinopora* 0.21%, *Goniastrea* 0.14%, *Leptoseris* 0.14% to *Symphyllia* 0.07%. **Notable zeros: *Millepora*, *Fungia*, *Seriatopora*, *Pavona*, *Turbinaria*, *Acanthastrea*.** From the photographs: ***Acropora*** in three habits (large tabular plates forming near-continuous fields in P4200929/930, corymbose-digitate white-tipped colonies, and low encrusting plates), ***Porites*** massive/domal and encrusting, ***Pocillopora***- and ***Stylophora***-type verrucose clumps, ruffled foliose sheets compatible with *Echinopora*/*Pachyseris*, pale green massive discs compatible with *Porites*/*Goniastrea*, plus pink CCA and a conspicuous **blue-grey encrusting sponge** (flag 3). Growth-form calls from photographs are given as "compatible with"; the genus percentages above are transect IDs.
**Representative photos:** P4200928.jpg, P4200927.jpg, P4200926.jpg, P4200902.jpg, P4200930.jpg

**Nelson Island is a low-relief lagoonal bank plain — an old dead tabular *Acropora* framework carrying a thin live veneer.** The structural signature is consistent across the folder: **P4200920** and **P4200926** are wide views over pale carbonate pavement and sand channels with small, individually resolvable colonies scattered on it, nothing touching; **P4200929** and **P4200930** are the counter-view, broad fields of large overlapping tabular *Acropora* plates, dull olive-brown and low-contrast in hazy blue water, with a diver in-frame in both. The transect agrees with the eye that this is not a high-cover reef — 19.11% coral against 47.68% algal — but it also says bare substrate is only **7.79%**, which is low, and that is because the plain is not bare: it is covered. The single transect line D3_T6 reads dead coral 41%, CCA 43%, *Acropora* 16% — a dead framework with a crustose veneer and a little living plate, which is precisely what the photographs show.

**The invertebrate fraction is the site's most distinctive number, and it explains two false bleaching calls.** Total invertebrate cover is **25.43% — the 93rd percentile of the Chagos set**, more than three times BITB33's 8.27%, and it is the single largest non-algal component after coral. The imagery identifies the likely driver: a pale blue-grey **encrusting sponge** that sheets over dead pavement, conspicuous in **P4200904** and **P4200905**. In both frames it reads at a glance as coral tissue that has lost pigment, and at P4200905 it sits under a bleached point call. It is not bleaching. This is a site where the substrate's dominant secondary occupant is an organism that mimics paling, and any automated or rapid-visual bleaching assessment here will over-call for that reason alone.

**The published 11.57% bleaching figure is carried by six frames, and it is decoupled from where white coral actually appears in the imagery.** The per-frame mean of bleached/(live+bleached) reproduces the published value **exactly** (11.574), and its composition is extremely top-heavy: **P4200899 alone contributes 4.861 pp — 42.0% of the entire site signal** — followed by P4200901 and P4200913 at 1.786 pp (15.4% each), P4200905 at 1.562 (13.5%), **P4200929 at 1.250 (10.8%)** and P4200902 at 0.329 (2.8%). **Ten of the sixteen frames contribute exactly zero.** Reviewed against the pixels, the ranking inverts:

- **P4200899**, the largest contributor, is a **pigmented** *Acropora* colony with pale **growing tips**. It draws 13 bleached-*Pocillopora* points and one bleached-*Acropora* point — the genus and the condition are both wrong — while the Coralscapes mask reads it at 0.95% bleached. Pale apical tips are normal *Acropora* growth, not pigment loss.
- **P4200913** is the same error on a different colony: a smooth digitate colony (*Millepora*- or branching-*Porites*-type) with pale tips over pigmented bases, 4 bleached points.
- **P4200905** and **P4200904** are the sponge (above).
- **P4200929** is the frame this pass added, and it is the clearest single-point artefact at the site. It is a broad field of uniformly dull olive-brown tabular *Acropora* with a diver upslope; nothing in it is white. It carries **one** bleached point out of 30. But it also carries only **four** live-coral points, so the frame ratio is 1/(4+1) = **20.0%**, nearly double the site's published value, and that one point becomes **10.8% of the site's entire bleaching signal**. The Coralscapes mask puts the frame at **0.0022% bleached**. This is the leverage problem stated plainly: the published metric is a per-frame mean of a ratio whose denominator is the frame's own coral points, so on a coral-poor frame a single point moves the site.
- **P4200902 is the only defensible paling at BINI78** — a massive *Porites*-type colony, cream over most of its surface with one retained tan pigmented sector, the classic partial-paling morphology. It receives **one** bleached point and contributes **2.8%** of the site signal, the smallest non-zero share.

**And the two frames that do contain unambiguous bone-white coral score zero bleached points.** **P4200927** carries a distinct cluster of stark white colonies just right of centre — white corymbose *Acropora*, white encrusting patches, white massive fragments — and **P4200926** carries a bright white branching colony left of centre plus several very pale plates on the right. Both are scored **0.0% bleached** by the point set. The Coralscapes mask does see something at P4200927 (1.13%, the highest model bleaching value of the six frames reviewed in this pass), which is the correct direction. **Whether those white colonies are bleached or recently dead cannot be settled at this resolution** — a clean white skeleton and a fully bleached colony are not separable in a wide-angle frame without a tissue check — but that ambiguity does not rescue the point set, which called neither.

**The net effect is that the bleaching signal at BINI78 is anti-correlated with the imagery.** The frames driving the published number show pale growing tips, an encrusting sponge, and a coral-poor tabular field; the frames showing white coral contribute nothing. **This site should not be cited as evidence of April bleaching in Chagos.**

**BLEACHING TIMING — the discrepancy applies here and is not resolved.** BINI78 was surveyed **20 April 2015**, inside the window in which Anderson recalls **no bleaching at all — only May**. The published 11.57% sits well above the March–April half-month means (6.1 / 8.2 / 8.4 / 10.0%) and far below the early-May figure of 31.8%. The frame-level audit above is consistent with the recollection rather than with the number: of the six contributing frames, four are demonstrable misreads (pale tips ×2, sponge ×1, single-point leverage on an unbleached field ×1), one is a genuine partial paling contributing 2.8%, and one (P4200901) is ambiguous. **Corrected for the four misreads, this site's April bleaching is on the order of a few per cent, not 11.6%.** This is the fourth April/March site to resolve this way and the open question in `RESUME_CHAGOS.md` §6.1 stands.

**THE HEADLINE METHODOLOGICAL FINDING: BINI78's excellent site-level agreement is cancellation, not per-frame accuracy.** The three methods land within 1.4 pp of each other at site level — 19.107 transect, 17.708 points, 18.428 mask — which is the 6th tightest agreement of 98 sites. Per frame, the point set and the mask disagree by a **mean absolute 8.04 pp**, with a range of **−31.5 to +24.5 pp**, and **5 of 16 frames disagree by more than 8 pp** (Pearson r = 0.701). The two largest errors point in opposite directions and very nearly annihilate: **P4200899 is −31.5 pp** (points 13.3, mask 44.8 — the pale-tipped colony the point set scored as bleached *Pocillopora* rather than live *Acropora*) and **P4200930 is +24.5 pp** (points 30.0, the site's highest live reading, against a mask value of 5.5 on a hazy field of dull tabular plates the model largely declined to call coral). **A site-level number this good, built from per-frame errors this large, is a warning about how the corpus-wide multimethod comparison should be read** — agreement at n=16 aggregation says little about whether either method is right on any given frame, and `CHAGOS_MULTIMETHOD_README.md`'s site-level statistics will contain an unknown amount of this cancellation.

**The best-agreeing frame tells you what drives the divergence: blue water.** **P4200928** is the only close-range, downward-facing frame in the folder with no water column at all — and it is the tightest human/model pair at the site (points 23.3%, mask 27.1%, −3.7 pp) on the lowest Unknown fraction (10.0%). Its Coralscapes background-dark share is **2.6%**; every wide frame in the folder runs 46–54%. Across the 16 frames, the frames with large blue-water fractions are the frames where the two methods part company. **P4200928** is also the site's best structural portrait: a large dark purple-brown digitate *Acropora* with white growing tips ringing its margin (the same morphology the point set misread at P4200899 — here it draws **zero** bleached points, which is the correct call and shows the error at 899 is inconsistent rather than systematic), an olive-yellow ruffled foliose colony compatible with *Echinopora*/*Pachyseris* upper right, white-tipped corymbose *Acropora* at left and bottom, pale green massive discs, and blue chromis over the central colony.

**The point set does not use its SAND label anywhere at this site.** Across all 480 points in 16 frames, **zero** are labelled SAND, on a folder in which pale carbonate sand and rubble channels are a dominant substrate — **P4200900** (a large intact tabular *Acropora* over white sand: 13 Turf points on bare sand), **P4200920** (23.3% Turf, 0% sand, over obvious sand channels) and **P4200930** (bright sand patches bottom-right, 0% sand) are the clearest cases. The sand is being absorbed into Turf and Unknown. Site-wide Unknown is **23.13%** and peaks at **50.0%** (P4200927) and 46.7% (P4200904, P4200926).

**The Unknown-denominator artefact behaves exactly as predicted for a high-reading site — this is the 10th confirming site.** Dropping Unknown from the denominator lifts the point set from 17.708% to **22.365%** against a 19.107% transect: an **overshoot of +3.26 pp**. Per `RESUME_CHAGOS.md` §3.5 the correction lands *on* the transect at low-reading sites (BIEA26, BITB27, BITB28, BITB29) and *overshoots* at high-reading ones (BIEA20, BIEA22, BIEA23, BIEA24, BIEA25). BINI78 reads high (ratio 0.927 vs corpus median 0.650) and overshoots. **It is the sixth high-reading site to do so and the tenth site overall to fit the pattern, with no exceptions yet recorded.**

**Flags:**
1. **ANALYSIS — THE TIGHTEST SITE-LEVEL AGREEMENT IN THE SET IS CANCELLATION, NOT ACCURACY.** Site-level spread 1.40 pp (6th smallest of 98); mean absolute per-frame human/model divergence 8.04 pp, range −31.5 to +24.5, 5 of 16 frames >8 pp, r = 0.701. P4200899 (−31.5) and P4200930 (+24.5) nearly cancel. Do not cite BINI78 as a site where the photo methods agree with the transect; cite it as the clearest demonstration that site-level agreement is not evidence of frame-level validity. Bears directly on `CHAGOS_MULTIMETHOD_README.md`.
2. **BLEACHING-TIMING-DISCREPANCY.** Surveyed 2015-04-20; published 11.57% bleached; Anderson recalls no bleaching before May. Frame audit supports the recollection: 4 of the 6 contributing frames are misreads. Corrected estimate a few per cent. Unresolved — `RESUME_CHAGOS.md` §6.1.
3. **NOTABLE-OBSERVATION / FALSE-PALING — blue-grey encrusting sponge at P4200904 and P4200905.** Reads as depigmented coral tissue; carries a bleached point at 905. Consistent with the site's 93rd-percentile invertebrate cover (25.43%). Not bleaching.
4. **ANNOTATION-PROVENANCE — pale *Acropora* growing tips scored as bleached *Pocillopora*.** P4200899: 13 B_Pocillo + 1 B_Acro on a pigmented *Acropora* with pale tips; genus and condition both wrong; mask reads 0.95% bleached. Contributes 42.0% of the site's bleaching signal. Same error at P4200913 (4 points). Note the inconsistency: the identical morphology at P4200928 draws zero bleached points. Provenance artefact — the Chagos point set is ~76% unconfirmed machine prediction; do not attribute to the annotator.
5. **METRIC-LEVERAGE — one point at P4200929 supplies 10.8% of the site's bleaching signal.** Frame has 4 live + 1 bleached point → 20.0% frame ratio on a field of uniformly pigmented tabular *Acropora*; Coralscapes reads 0.0022% bleached. The published per-frame-mean-of-a-ratio formula gives unbounded leverage to single points on coral-poor frames. Corpus-wide concern, not site-specific.
6. **ANNOTATION-PROVENANCE — visible bone-white colonies score ZERO bleached points.** P4200927 (white cluster right of centre; 50.0% Unknown; mask 1.13% bleached) and P4200926 (white branching colony left of centre; 46.7% Unknown; mask 0.015%). Bleached vs recently-dead is not separable at this resolution — flagged as ambiguous, not as bleaching. The point set missing them in both directions is the finding.
7. **ANNOTATION-ERROR — SAND label unused across all 480 points.** Zero SAND points in 16 frames over a sand-and-pavement plain. Clearest at P4200900 (13 Turf points on bare sand under a tabular *Acropora*), P4200920, P4200930. Sand is absorbed into Turf (site mean Turf inflated) and Unknown (23.13% site-wide, max 50.0%). Affects any turf- or algal-cover statistic derived from this point set.
8. **UNKNOWN-DENOMINATOR — 10th confirming site, high-reading branch.** 17.708% → 22.365% excl-Unknown vs 19.107% transect = +3.26 pp overshoot. Sixth high-reading site to overshoot; pattern still has no exceptions.
9. **METHODOLOGY-IMAGERY — divergence tracks blue water.** P4200928, the only frame with no water column (mask background 2.6% vs 46–54% for every wide frame), is the tightest human/model pair (−3.7 pp) on the lowest Unknown (10.0%). Supports excluding or down-weighting wide-angle blue-water frames from photo-derived cover.
10. **METHODOLOGY-IMAGERY — sampling macro in the survey set.** P4200901 is a sampling photograph (gloved hand and bone cutters on a *Tridacna*), not a benthic survey frame, and it contributes 15.4% of the site's bleaching signal via an ambiguous white pocilloporid nubbin bottom-left. Third site at which measurement/methodology frames are inside the CV sample — `RESUME_CHAGOS.md` §6.2.
11. **ANNOTATION-ERROR — highest-cover frame by eye scores 6.7% live.** P4200904 reads 30–50% coral visually, scores 6.7% live with 46.7% Unknown. Worst single annotation artefact at the site.
12. **NOTABLE-OBSERVATION — folder is named "no good pics", and this site has one of the smallest photo/transect gaps in the corpus.** Suggestive for the frame-selection hypothesis (a set chosen as leftovers should not over-represent coral), but **it is the only folder of 120 named this way, so n = 1 and the hypothesis cannot be tested from folder naming.** Recorded as a hypothesis, not a result. Testing it needs a deliberate frame-selection experiment.
13. **DATA-QUALITY — genus richness is effort-limited.** 14 genera on 9 transects (D2 n=2, D3 n=7) against a corpus median near 15 transects. Per `RESUME_CHAGOS.md` §3.6 (r = 0.66 between richness and effort), this is a richness floor, not a depauperate assemblage. Do not compare BINI78's 14 genera to BIBL87's 37 (16 transects).
14. **DATA-QUALITY — negative informal-vs-transect difference.** Published informal image cover 24.405% (reproduced exactly as the per-frame mean of (live+bleached)/(30 − water − fish)) exceeds the 19.107% transect. BINI78 is one of only **32 of 105** sites where the image method reads *higher* than the transect.
15. **HOUSEKEEPING — CRD page URL is the `new-page-4` placeholder** and needs repointing with the other 24 links once that page is renamed.
16. **NON-CANONICAL BLEACHING RANK — added 2026-08-26.** Any "of 98" bleaching rank above was computed off the site table's `Mean(%bleaching)` column, which equals the **1st** CoralNet training run. Anderson's decision of 2026-08-26 (HOW_TO rule 4e) makes **`% bleaching (2nd training)` canonical**, The file carries TWO pools for it, and they are not interchangeable: a **110-row pool** ranking every dive (revisit dives included, so BIPB70B/72B/69C compete) and a **99-site pool** ranking one value per site. The rank quoted below is on the 110-row dive pool. For this site the canonical value is **11.57%, rank 32 of 110** (1st training: 11.57%, rank 28 of 110). The two runs agree to <0.05 pp on 100 of 110 rows and diverge only on the bleaching-onset dives, so for most sites the distinction changes the rank but not the number.

**Benthic data:** [Download BINI78 benthic survey data (CSV)](../benthic_data/BINI78_benthic.csv) — scleractinian cover by genus with full names, plus substrate and algal categories (bare substrate, CCA, turf, macroalgae, invertebrates), from the LOF BIOT-1 survey (Mayfield & Dempsey 2026, *Oceans*, supplemental file S1).