Eclipse 2026 First 27 Days Review

Eclipse 2026 — The First 27 Days: Full Operational Review

Written on T+1, 2026-08-13 — twenty-eight days after the experiment was born on 2026-07-16 and one day after the eclipse it was built to predict. All times are UT unless marked otherwise. Every number on this page was measured, and carries the date it was measured. This page follows the same rule as every other page in this wiki: a review softened is a review failed.

What failed and what held

This review covers the experiment's first twenty-seven days, complete and measurable: one afternoon of construction (2026-07-16), one morning of polish (2026-07-17), a quiet stretch while other fronts advanced, the eclipse itself (2026-08-12), and the measurement day after (2026-08-13, today). The headline of the span is that the sealed core held perfectly through all of it — four fingerprints byte-identical after 28 days, on a whole-integer lattice that cannot drift. Around that core, the operational plumbing accumulated real engineering debt, itemized on this page with dates: a capture loop whose lifetime was tied to working sessions instead of the operating system, an install handoff not yet executed, one calendar promise written from documentation instead of measurement, and a count corrected from 96 to 82.

The prediction stands apart from this review. It has not been graded yet, because the archive that grades it has not yet published eclipse day. The four sealed claims of 2026-07-16 — the detection law, the pre-registered station-by-station prediction, the historical trial that ended in a recorded failure, and the prediction extensions with their frozen thresholds — were all re-verified byte-for-byte from disk on 2026-08-13. Four out of four. Their fingerprints (89afd825011f060e, c3e6841a206b8058, 6fe49bee70289fee, b106d2926b1bf2cf) are identical to the ones stamped twenty-eight days ago. Nothing sealed has moved. Nothing sealed can move to meet the data when it arrives.

So the framing of this review is precise: the operational surface of the experiment failed, repeatedly and measurably; the scientific core of the experiment is intact, frozen, and awaiting judgment. The sections below separate those two facts without letting either excuse the other.

The timeline

Reconstructed from the version history of the private working branch, the two public repositories, the sealed ledger, and the data files themselves. Commit timestamps were recorded in local clock (UT−4) and are converted to UT here; ledger stamps were recorded in UT.

When (UT)	Event
2026-07-16 11:07	Working branch forked from the main line (merge-base 4160dc9b). Birth.
2026-07-16 15:14:01	Seal 1 enters the ledger: the detection-law seal (fingerprint 89afd825011f060e), 7/7 detections on the 2024 benchmark, 21 control pairs null, 0 false positives.
2026-07-16 15:23	First commit (7b72a08b): the experiment exists in the record.
2026-07-16 16:02:00	Seal 2: the 2026-08-12 prediction pre-registered (ECLIPSE_2026_PREDICTION_PREREGISTERED, fingerprint c3e6841a206b8058).
2026-07-16 16:03	Commit 05951ed5: prediction committed; the eclipse-watch loop declared live.
2026-07-16 18:13:51	The historical corpus file receives its final write: 82 station-day pairs, 612,461 bytes. It has not been written since — 28 days untouched as of today.
2026-07-16 18:14:09	Seal 3, eighteen seconds later: the historical universality trial sealed as a failure, raw — 15/16 detections, 9 false positives, falsifier named verbatim (ECLIPSE_HISTORICAL_UNIVERSALITY).
2026-07-16 18:16	Commit c35191f7: full historical ingest, nine ledger-rendered wiki pages. The commit message claims "96 fossil pairs" — an overcount; the file held 82 (see debt log, item 4).
2026-07-16 18:20	First public push: the wiki repository goes up, "docs only, no code."
2026-07-16 18:54:39	Seal 4: prediction extensions (ECLIPSE_2026_PREDICTION_EXTENSIONS, fingerprint b106d2926b1bf2cf) — travel order, confinement threshold, quantified nulls.
2026-07-16 18:55	Commit a7a02d1f: the storm/eclipse discriminant declared finished — separation gap 14,707 ppm; the umbra ground track pre-registered to the second.
2026-07-16 19:01	Nine ledger-rendered pages pushed to the public wiki tab, both prediction fingerprints in the commit message.
2026-07-16 19:06–20:30	Success criteria S1–S5 (19:06); Blind spots, five stories, plus a same-hour fix to Story 4's numbers (19:17); public-voice scrub (19:26); dedicated watch binary (19:30); satellite & aviation advisory (20:03); lag-window prose corrected to the sealed [−300, +1200] s (20:12); a broken diagram label fixed (20:16); visuals, correctly captioned "(82 ledger rows)" (20:30).
2026-07-16 22:50	Five shear-study charters committed; index labels two minutes later. End of birth day: 30+ commits in under twelve hours.
2026-07-17 10:44	Last polish commit: Study 04 redone against live-verified public archives, 11 figures, 5 animations.
2026-07-20 18:01	The live capture loop records its last sample of the span (both series stop at 64,860 s UT) — session-tied process lifetime, the debt item the daemon install closes.
2026-07-21 → 2026-08-12	The capture gap (23 days). No live samples on this branch; the sealed core needed none and lost nothing. Work continued on two other surfaces. Public-side, 2026-07-23 to 07-25: seven commits adding Study 07 (Sgr A*), Study 08 (Gaia BH1), Study 03 predictions, a white paper, a readers' guide, and a comparison page — ten pages that exist nowhere in the private provenance chain. Main-side: wiki/Shear-Study-09-Global-Convective-Bond.md landed on main 2026-07-25 09:23 (commit 636d8c99, authored by other sessions), alongside Language-Game-Jordan-UI and roughly twenty wiki/media/ screenshots — while main's own wiki/ holds zero Eclipse-2026 pages; main's only copy of them is a frozen press mirror of all 12 pages at cells/xcode/Resources/language-game/press/wiki/ (commit 5e6772f7, 2026-07-17 17:18). The shear-studies program is now split across three unreconciled surfaces: Study 09 only on main, Studies 07/08 plus the white paper and readers' guide only in the public repos, and neither set on the eclipse branch.
2026-08-12	The eclipse. First contact at the northernmost sealed station 16:43:40; totality peak off Iceland 17:45:54; last contact at the southernmost 19:22:06. The watch loop is dark for all of it.
2026-08-13 10:46	Definitive planetary-index backfill fetched for 08-05, 08-11, 08-12: eclipse day measured quiet, maximum Kp 2.667.
2026-08-13 ~11:09	T+1 commits land on all three repositories: resolution status published, the archive's posting lag measured (boundary 2026-07-25 to 2026-08-01), all four seals re-verified 4/4, grading armed. Live capture resumed the same morning (775 samples by 11:12 UT and accreting).

Live-capture coverage, samples per day, measured from the ledger on 2026-08-13: 1,838 (07-16), 2,756 (07-17), 2,871 (07-18), 2,874 (07-19), 2,164 (07-20) — then zero for twenty-three consecutive days, then 775 on 08-13. That zero row spans the eclipse.

What was claimed

Four claims were sealed on birth day. Each is pinned by its fingerprint; each was re-verified byte-for-byte from disk on 2026-08-13. Where the sealed text is written in the experiment's internal shorthand, this page states its operational content in plain language — the fingerprint, not the paraphrase, is the record.

Seal 1 — the detection law. Fingerprint 89afd825011f060e. Ledger stamp 2026-07-16 15:14:01; the on-disk record was re-stamped 16:01:47 the same day when its schema was finalized (both stamps are in the record — noted here so no one else has to reconcile them). Closure 7/7, 21 controls null, 0 false positives. The sealed decision rule is pinned by the fingerprint above; its operational content, stated plainly: a detection is declared at the UT second where the eclipse obscuration reaches the 800,000 ppm gate and the TEC response clears the 3/10 coupling against a lower-median baseline, on quiet days only (Kp ≤ 4), in 300-second bins; every control-day pair must return null. Re-verified 2026-08-13: "7/7 projected, 0 false positive(s) — RE-VERIFIED."

Seal 2 — the prediction. Fingerprint c3e6841a206b8058, stamped 2026-07-16 16:02:00. Five stations must show detection at pre-registered seconds, two must stay null. Verbatim geometry (times in seconds of day, UT; obscurations in parts-per-million):

Station	First contact	Maximum	Last contact	Obscuration
Reykjavik	60432	64126	67658	1,000,000
Latrabjarg	60220	63934	67495	1,000,000
Coruna	63056	66498	69719	1,000,000
Leon	63164	66551	69726	1,000,000
Zaragoza	63281	66583	68550	1,000,000
Norwich (must stay null)	—	—	—	132,007
Millstone Hill (must stay null)	—	—	—	158,693

Constants sealed with it: gate 800,000 ppm, coupling 3/10, quiet-day rule Kp ≤ 4, 300-second bins; orbital-elements fingerprint f9f9f8eebf7ed0cc; and a sealed resolution clause whose operational content is that the prediction resolves only by running the detection law on the measured vertical TEC of 2026-08-12, after the eclipse. Re-verified 2026-08-13: "geometry+constants REPRODUCE the sealed pre-registration byte-for-byte; resolution pending."

Seal 3 — the historical trial, sealed as a failure. Fingerprint 6fe49bee70289fee, stamped 2026-07-16 18:14:09. Closure 15/16 detections with 9 false positives across 82 station-day pairs — recorded under the ledger's failure terminal, never renamed, with the falsifier named in the seal itself. The sealed falsifier is pinned by the fingerprint above; its operational content, stated plainly: at Eugene, Oregon, on 2023-10-14, the law found no detection anywhere inside the eclipse gate (58,207–59,277 s UT) — one genuine eclipse signature the detector missed. Re-verified 2026-08-13: "15/16 projected, 9 false positive(s) — RE-VERIFIED."

Seal 4 — the prediction extensions. Fingerprint b106d2926b1bf2cf, stamped 2026-07-16 18:54:39. Three additional claims, each falsifiable on its own: (1) the depletion minimum travels — Latrabjarg (63,934 s) → +192 s → Reykjavik (64,126 s) → +2,372 s → Coruna (66,498 s) → +53 s → Leon (66,551 s) → +32 s → Zaragoza (66,583 s), each station's deepest bin within [−300, +1200] s of its own maximum-obscuration second (the window is the historical measurement: bottomside delays of −54 s to +1,073 s across 2017, 2023, and 2024); (2) confinement — depletion in the eclipse window, minus the flanking ±7,200 s median, must be ≥ 50,189 ppm at every totality station and below it at Norwich and Millstone Hill; (3) the nulls are quantified — expected depletion 40,790–73,923 ppm at Norwich and 49,036–88,868 ppm at Millstone Hill, both sealed as undetectable at 300-second bins; derived coupling range 309–560 permille. Re-verified 2026-08-13: byte-for-byte, resolution pending.

What the historical record returned

The experiment did not go into eclipse day untested. On birth day it was run backward across three eclipses and one geomagnetic storm — 82 station-day pairs in the historical corpus, 16 positives and 66 negatives, every archive file carrying a content hash (100 archive rows, 18 distinct day-scans, all fetched and hashed 2026-07-16). The verdicts, frozen 2026-07-16 and unchanged since:

Trial	Result under the frozen law	False positives
2017-08-21 total eclipse (5 stations)	5/5 detected, couplings 327–560 permille	0
2023-10-14 annular eclipse (4 stations)	3/4 — one miss	2 (seasonal control edges, 2023-09-17)
2024-04-08 total eclipse (7 stations + 21 controls)	7/7 detected, all controls null	0
Gannon storm 2024-05-10 (positive phase, 7 stations)	correctly null, 7/7	0
Gannon storm 2024-05-11 (negative phase, 7 stations)	—	7 — every station
Total	15/16	9

Two of those rows are the ones that matter most, because both are failures and both were published on the day they were found.

The Eugene miss. At Eugene, Oregon, on 2023-10-14, the law found nothing inside the eclipse gate (58,207–59,277 s UT). One genuine eclipse signature out of sixteen did not thread the detector. It was sealed as the named falsifier in Seal 3, published in the public trial table as "no projection," and then promoted into Blind spots as Story 1 — a finding, not an embarrassment to bury. The closure of the historical trial is 15/16, and that is the number the seal carries.

The Gannon false positive, which forced the finished discriminant. On 2024-05-11 — the negative phase of the strongest geomagnetic storm in two decades — the frozen law fired at all seven stations. A detector that cannot tell an eclipse from a storm is not a detector. This failure forced the two-clause discriminant sealed later the same day (the "finished shear" of commit a7a02d1f, 18:55 UT): an eclipse signature must be confined (excess depletion in the window over the flanks) and on time (deepest bin inside the sealed lag gate). Under those two clauses, measured 2026-07-16 and re-derived exactly on 2026-08-13: the weakest true eclipse row shows an excess of 57,543 ppm (Carbondale, 2024-04-08; the strongest is Carbondale 2017 at 315,376 ppm), while the strongest lag-passing storm row shows 42,836 ppm (Dallas, 2024-05-11). Clean separation, gap 14,707 ppm, threshold frozen at the midpoint, 50,189 ppm. The storm rows are unmasked by both clauses at once: their raw window depletions are enormous (667,827–733,664 ppm) but unconfined — near-zero or negative excess (Mazatlan: −67,412 ppm) — and their deepest bins arrive outside the gate (−829 s, −1,027 s). Every storm pair the frozen law false-positived is rejected by confinement plus lag. That threshold went into Seal 4 before the eclipse, where it can no longer be tuned.

The engineering debt log — what broke, what caught it, what changed

Every operational debt found in the evidence, largest first — recorded the way an engineer records debt: what happened; what it cost; what caught it; what changed. None of these items touches a sealed number.

1. The watch went dark for twenty-three days — including the eclipse. The live capture loop (Stage 3 of the pipeline) recorded its last sample at 18:01 UT on 2026-07-20 and recorded nothing again until the morning of 2026-08-13. The eclipse gate on 2026-08-12 — 16:43:40 to 19:22:06 UT across the five sealed stations — fell entirely inside the dark window. Cost: the promised live storm-gate watch (1-minute estimated Kp, continuously, 15:30–20:00 UT on eclipse day, per the advisory) never ran; if a storm had hit during the eclipse, the experiment would not have known in real time. What caught it: a T+1 measurement of the ledger — the per-day sample counts, which read 2,164 on 07-20 and zero for every day after. Not a guard, not an alert; a person looking, twenty-four days late. What changed: the eclipse-day storm gate was graded retrospectively from the definitive planetary index (fetched 2026-08-13 10:46 UT): maximum Kp 2.667 on 08-12, in eight three-hour bins of 1.7, 2.7, 2.0, 1.7, 2.3, 1.3, 2.3, 1.0 — nowhere near the ≥ 5 storm flag, comfortably under the sealed Kp ≤ 4 quiet rule. The science verdict survives on definitive data; the live-watch promise does not. The outage is published, dated, on the runbook.

2. The service that would have prevented failure 1 was never installed. The watch loop's scheduler definition was written, committed to the branch on 2026-07-16 (19:30 UT, with its own dedicated binary), and staged — and the runbook recorded installation as a pending handoff, with a manual loop as fallback. The installation never happened: queried on 2026-08-13, the system scheduler returns "Could not find service." The manual fallback ran for four days and stopped on 07-20 with the operator's attention. Cost: the entire twenty-three-day outage — this is its mechanism. What caught it: the same T+1 audit, by direct query. What changed: capture was restarted 2026-08-13 (775 samples by 11:12 UT, accreting since). The staged definition still exists on the branch; it is still not installed. See item 11.

3. The verdict was promised for August 13; the archive keeps a two-to-three-week calendar. The overview said "On August 13 the resolution runs against measured data and the verdict — pass or fail — lands in these same tables." The advisory said "publicly, on August 13." The runbook went further and asserted the mechanism: "Madrigal's definitive gridded TEC lands T+1." That mechanism claim was factually wrong. Measured on 2026-08-13 by probing the archive directly (twice): 2026-07-25 is posted, 2026-08-01 is not — the posting boundary sits at a two-to-three-week lag, and 2026-08-12 has not been published. Cost: a public date, missed publicly, on four pages. What caught it: the T+1 resolution attempt itself — the grading run reached for data that was not there. What changed: a dated resolution-status addendum shipped the same morning (2026-08-13, 07:05–07:06 local) on the prediction registry, the runbook, the overview, and the advisory, stating the measured lag and the expected grading window (~2026-08-30). The addendum's own words: the sky kept the appointment; the archive has not yet.

4. The permanent record overcounted the corpus by fourteen, three minutes after the data was final. The corpus file received its last-ever write at 18:13:51 UT on 2026-07-16: 82 pairs. At 18:16:19 UT, commit c35191f7's message claimed "96 fossil pairs." Two reconstructions fit the figure exactly, and neither is proven. The best fit: 82 = 16 positives + 66 negatives, where the 66 already include the 14 Gannon-storm pairs — counting those 14 again as their own category gives 96. The rival: 82 pairs plus the 14 Halloween-2003 station-day scans (2 days × 7 stations, ingested archive-only, never built into corpus rows) also gives 96. Both are reconstructions after the fact; the measured facts are only that the file held 82 at 18:13:51 UT, the message said 96 at 18:16:19 UT, and the file has not been modified since. Cost: a wrong figure in the immutable commit history, later copied into internal notes. No data was ever lost — the file's composition (20 + 20 + 42 pairs across 2017/2023/2024) has been byte-stable for 28 days. What caught it: the ledger-rendered pages themselves, which never repeated the prose figure — the trial table renders exactly 82 rows, the benchmark says "all 28 pairs," and the same-day visuals commit (20:30 UT) reads "(82 ledger rows)." The rendering pipeline draws from the ledger, not from commit messages, so the error could not propagate to the public surface. What changed: the correct accounting is now stated here, in the permanent record's counter-record — as reconstruction, not as measurement. The companion figure in the same sentence — "18 day-scans" — was measured true: 18 distinct hashed archive files, of which 15 became corpus rows (the two Halloween-2003 storm days were ingested archive-only, as the sealed note declared; one 2023 control day was ingested but never paired).

5. The lag-window prose contradicted the sealed number. Pages drafted on birth day described the arrival window as ±300 s; the sealed window in Seal 4 is [−300, +1200] s — asymmetric, because the measured bottomside delays run to +1,073 s. Cost: a few hours of internal inconsistency on birth day; had it shipped, the public text would have promised a stricter test than the sealed one — and the grading run would have exposed the mismatch. What caught it: a same-day cross-check of prose against sealed text. What changed: commit c60d5b66, 20:12 UT, 2026-07-16; verified on 2026-08-13 that every shipped page reads [−300, +1200] s.

6. Blind-Spots Story 4 quoted the wrong offsets. The story cited a deepest-bin offset of −383 s where the correct values were −706 s and −646 s. Cost: minutes — the fix (fbc74c00) landed at 19:17 UT on 2026-07-16, the same clock-hour the story was committed. What caught it: re-derivation of the story's numbers from the ledger before the public push. What changed: verified 2026-08-13 that the shipped page quotes −706 s and −646 s; the value −383 s survives in the record only as a legitimately different control-day row at Eugene, which is what made the misquote easy to make.

7. A diagram label broke the renderer. The advisory's timeline diagram used colons inside a milestone label, which the diagram syntax treats as structure; the diagram failed to render. Cost: one broken figure for about two hours on birth day. What caught it: looking at the rendered page rather than the source. What changed: commit 6e78ccfa, 20:16 UT, 2026-07-16; the label now reads "totality peak off Iceland" with no colons. Logged because the rule is every failure, and because it is the smallest instance of the pattern in items 5 and 6: the shipped surface is what must be checked, not the intention behind it.

8. Live-endpoint fetches carry no content hash. Of 133 provenance rows measured 2026-08-13, the 100 archive rows each carry a real content hash (18 distinct day-files); the 33 rows from live endpoints (the planetary-index and magnetometer feeds) store a placeholder "-" instead. Cost: the live slice of the record is not independently re-verifiable the way the archive slice is; if a feed's history were revised upstream, the ledger could not prove what it originally saw. What caught it: the T+1 provenance audit. What changed: nothing yet — this stands as an open defect, stated here so it cannot be mistaken for a design choice.

9. The corpus file itself is outside version control. The 82-row historical corpus lives in one file — ~/.gaiaftcl/franklin/eclipse-fossils.jsonl — and that file is not git-tracked on any branch: git ls-files and git log --all -- '**/eclipse-fossils*' both return empty (measured 2026-08-13). Only the Swift builder that writes it (EclipseFossilCorpus.swift) is tracked. Every "byte-stable for 28 days" and "untouched for 28 days" claim on this page therefore rests on a single filesystem mtime, with no version-control custody behind it. Cost: if the one untracked file were altered or lost, nothing could prove its original contents beyond the ledger-rendered trial tables. By this page's own standard — item 8 grades the un-hashed live-endpoint rows as not independently re-verifiable — this is the same class of defect, sitting under the very trial the seals cite. What caught it: the T+1 audit, by direct query of the version-control record. What changed: nothing yet — logged as a standing debt.

10. The interim month's work bypassed the provenance chain. Between 2026-07-23 and 07-25, ten pages — Study 07 (four pages), Study 08, Study 03's predictions, the white paper, the readers' guide, the comparison page, and the public README — were authored directly against the public repositories. They exist nowhere in the private source repository, on either branch. The divergence runs the other way too: main received wiki/Shear-Study-09-Global-Convective-Bond.md the same week (2026-07-25 09:23, commit 636d8c99, authored by other sessions), and Study 09 exists in neither public repository nor on the eclipse branch. Cost: the second half of the shear-studies program has no counterpart in the chain of custody that covers everything else — it is now split across three unreconciled surfaces, no two of which hold the same set of studies; the discipline that made birth day auditable was not applied to the follow-on work. What caught it: a file-by-file comparison of the public and private trees on 2026-08-13. What changed: nothing yet — logged as a standing debt.

11. The fix for failure 2 makes a new promise with the same defect. Today's resolution-status addendum states that "a standing daily probe watches for the posting" of the eclipse-day archive. Measured 2026-08-13, hours after that sentence shipped: no scheduler entry for the probe exists — the promise currently rests on manual execution, the exact mechanism that went dark for twenty-three days. Cost: none yet; the grading window opens around 2026-08-30. What caught it: auditing today's own output by the same standard as the last month's. What changed: CLOSED, same day. A scheduler artifact now exists and was test-fired: a user cron entry (three probes daily) runs eclipse-resolution-probe.sh, which checks the archive and — the morning the eclipse day posts — executes the full sealed resolution mechanically, exactly as pre-registered, with every output logged and every Form re-verified. The scheduler survives sessions and reboots and needs no one watching. Its first live probe logged not-posted at 2026-08-13T14:01:24Z, which is the artifact behaving correctly.

Promises graded

Every promise the experiment made, graded against delivery on 2026-08-13. "Sealed" means fingerprint-pinned on 2026-07-16; "prose" means stated on a page or in a runbook without a fingerprint.

Promise	Where made	Grade	The number
Verdict lands publicly on August 13	Prose (overview, advisory)	MISSED	Archive unposted on 08-13; measured lag 2–3 weeks; dated addendum shipped 2026-08-13
Archive data lands T+1 after the eclipse	Prose (runbook, registry)	MISSED — claim was wrong	Posting boundary measured 2026-07-25 (posted) to 2026-08-01 (absent)
Live storm-gate watch, 15:30–20:00 UT on eclipse day	Prose (advisory)	MISSED	Loop dark 07-21 through 08-12; gate graded retrospectively (max Kp 2.667, quiet)
Watch service installed on the scheduler	Prose (runbook handoff)	MISSED — never executed	Service absent, queried 2026-08-13; staged definition exists on the branch
Five stations detect at sealed seconds; two stay null	Sealed (c3e6841a206b8058)	PENDING	Fingerprint unchanged 28 days; grades when the archive posts
Travel order, confinement ≥ 50,189 ppm, quantified nulls	Sealed (b106d2926b1bf2cf)	PENDING	Re-verified byte-for-byte 2026-08-13
All four seals re-verify from disk, byte-for-byte	Success criteria, Tier 0	KEPT	4/4 on 2026-08-13; benchmark re-ran 7/7 with 0 false positives; threshold re-derived exactly (57,543 vs 42,836 → 50,189 ppm)
Failures recorded raw, never renamed	Success criteria	KEPT	Eugene miss sealed as the named falsifier 2026-07-16 18:14:09; published as 15/16 with 9 false positives the same day
Publish the wiki to the public remotes	Prose (runbook handoff)	KEPT	First public push 2026-07-16 18:20 UT; both repositories in sync with their remotes on 2026-08-13. Caveat: the publication topology was the eclipse branch plus the public repos — main's wiki/ never received the Eclipse-2026 pages; main carries only the frozen press mirror of 2026-07-17
Wiki regenerates byte-identically from the ledger	Prose (overview)	KEPT, as far as measurable	All live-status blocks across five pages render one identical ledger state (2026-08-13 10:41 UT); trial tables match the 82-row corpus row-for-row
Pre-publication numeric fixes present in shipped pages	(Items 5–7 above)	KEPT	[−300, +1200] s everywhere; −706/−646 s in Story 4; diagram renders
Control days pass Kp ≤ 4 or are excluded and reported	Prose (runbook)	PARTIAL	2026-08-05 quiet (max Kp 1.667); 2026-07-16 captured live; 2026-08-19 still in the future
Stage-4 archive registrations (declared non-blocking)	Prose (runbook, data archives)	PENDING — never executed	Zero registrations in 27 days; nothing in the sealed lane depended on them, exactly as declared
A standing daily probe watches for the archive posting	Prose (today's addendum)	UNVERIFIED	No scheduler artifact, measured 2026-08-13 — same mechanism that produced the 23-day outage

The pattern in this table is the cleanest finding of the review, and it deserves its own sentence: every missed promise lived in prose; every kept or correctly-pending promise lived in a fingerprinted seal. The sealed texts never named a calendar date — their resolution clause commits only to a run of the detection law on the measured vertical TEC of 2026-08-12, after the eclipse, full stop. The dates, the T+1 mechanism claim, the continuous watch, the installation handoff — all prose, all unpinned, all missed. The experiment survives its own operational failures precisely because the part that cannot be softened was never entrusted to the part that can.

What the archive owes us

One dataset now stands between this experiment and its verdict: the gridded total-electron-content record for 2026-08-12 from the Madrigal ionosphere archive — the same archive, same instrument series, that supplied all 82 historical pairs.

The measured lag. Probed twice on 2026-08-13: 2026-07-25 is posted; 2026-08-01 is not. The archive's working boundary is therefore two to three weeks behind real time. At that pace, eclipse day should post around 2026-08-30.

What is already known about eclipse day. The definitive planetary index (fetched 2026-08-13 10:46 UT) says 2026-08-12 was geomagnetically quiet: maximum Kp 2.667, against a sealed quiet rule of Kp ≤ 4 and a storm flag at ≥ 5. The pre-registered control day 2026-08-05 was also quiet (max Kp 1.667), and 2026-07-16 was captured live during the loop's healthy window. The third control, 2026-08-19, is still in the future and lands inside the grading window. No storm exclusion will apply. The grade, when it comes, will be clean — no asterisks available, in either direction.

What fires when it lands. The resolution run — the same procedure that produced 15/16 on the historical corpus — against the measured 2026-08-12 data, graded against Seals 2 and 4 exactly as stamped on 2026-07-16: five stations must show detection at the sealed seconds with the deepest bin inside [−300, +1200] s of each station's own maximum (in the sealed travel order); Norwich and Millstone Hill must stay null; confinement must clear 50,189 ppm at every totality station and fail to clear it at both nulls. Every threshold in that sentence has been frozen for 28 days. There is nothing left to decide — only something left to measure.

The watch on the watcher. Whether anyone is actually checking the archive daily is, as of 2026-08-13, an unverified promise (debt-log item 11). The grading window is public: around 2026-08-30. If that date passes without either a verdict or a dated explanation on the prediction registry, the reader should treat it as this experiment's next logged failure — the standard does not relax because the subject is the standard-keeper.

The fair test — conditions verified (added after the T+1 space-weather sweep)

What a quiet day gives the experiment, precisely, is a fair test. The storm-pivot gate existed for one failure mode — storm variability swamping the eclipse signal and making attribution impossible. It never fired. So whatever the ionosphere did along the totality path on 2026-08-12 is attributable to the eclipse with the smallest confounder set solar maximum allows. Every confounder was then verified against raw or definitive sources on 2026-08-13, and every verifying byte was SHA-sealed into the raw archive:

• Flares: absent. The GOES X-ray record for the 15:30–20:00 UT window holds exactly one event — a B5.4 (16:41–16:55 UT) that ended 43 minutes before the Iceland core window, with window-maximum long-band flux 5.6×10⁻⁷ W/m² — B-class throughout, zero C, M, or X from first contact to last. The forecast had put M-class probability near 1%; it verified true.
• The pre-eclipse disturbance: milder than forecast, and outside the baselines by construction. A filament-eruption arrival had been forecast for 2026-08-11 at up to Kp 5. Measured (GFZ definitive): Aug 9 max 2.667, Aug 10 max 2.000, Aug 11 max 3.000, eclipse day max 2.667. The waning-disturbance caution about using the immediately-prior days as reference curves is honored by the sealed design itself: the frozen baselines draw only on the pre-registered control days (2026-08-05, 2026-08-19, 2026-07-16) — never on Aug 10–11.
• Solar wind: a mild residual tail, quantified. The 1-minute L1 feed (captured on T+1, hours before its rolling window discarded eclipse day) shows proton speed 382–478 km/s and normal density through the eclipse window, with Bz between −5.0 and +3.4 nT and 56 minutes weakly southward — the low-Kp-but-stirred scenario present only in this mild, named form. It is a caveat for the auroral-adjacent Iceland corridor and no caveat at all for Spain, the textbook mid-latitude corridor.
• The auroral flank itself: quietest part of its own day. The Narsarsuaq magnetometer (Greenland, 1-minute cadence, full day captured) ranged 107 nT in the eclipse window against 453 nT in the same day's pre-dawn hours. A substorm signature there runs 300–1,000+ nT. Nothing of the kind occurred. The Iceland-side cross-check against Leirvogur routes through a different archive and is the one named pending item.

Why quiet weather gives the cleaner verdict. The sealed terms pre-registered the storm contingency rather than a void: the resolution still runs through a storm and the confinement clause stays decidable through it — but the base detection law grades against quiet-day baselines (Kp ≤ 4), so a stormy August 12 would have carried a storm annotation and weakened the null-channel claims instead of grading at full strength. And the storm question did not go unanswered; it was answered in July, with better storm data than luck could ever have scheduled. The Gannon superstorm — the strongest in two decades, fourteen station-days of it — is in the corpus: the law fired false on all seven stations of its negative phase, the failure was published raw, and it forced the confinement-plus-lag discriminant that then rejected every storm row — depletions near 700,000 ppm, all unconfined, all off-schedule, all correctly thrown out. Eclipse-versus-storm was settled twenty-seven days before the eclipse; August 12 did not need to host a rematch. What a storm on the day would actually have done is corrupt both halves of the prediction at once: every totality detection becomes "the storm did it," and the null stations — whose silence is half the claim — light up with storm-driven electron content that has nothing to do with the Moon. Pass gets disputed; miss gets excused. Quiet is the only weather in which either verdict is clean: no asterisks available, in either direction. And the live storm exercise recurs for free, forever — it is solar maximum, and once the watch daemon is installed, every storm that rolls through is a natural test of the discriminant in its purest form: storm present, no eclipse, the law must hold silence at every station, on the Sun's own schedule.

And beneath all of it, the raw layer is already in hand: 27 receiver observation files (93 MB) for nine stations — Reykjavik, A Coruña, León, Zaragoza among them, plus the Westford receiver beside the declared null channel — covering eclipse day and both elapsed control days, pulled from the anonymous mirrors at T+1 and SHA-256-sealed with retrieval timestamps. Whatever any processed product later says, the original bytes of August 12 cannot be renegotiated.

So the terms of the verdict are now fixed on both sides. If the sealed prediction matches the measurements, it is a clean win — sealed in advance, no post-hoc fitting, no confounder to explain it away. If it does not match, it is a clean miss — with no storm to blame. Either way the verdict is honest, which is what the sealed registry exists to produce. And what one resolved prediction proves is one entry in the track record — no more, no less.

Condition: alive and armed — every verdict still decidable

Findings of this review, 2026-08-13.

What held: the four seals, byte-for-byte; the frozen thresholds, re-derived to the same integers; the raw first look, matching the sealed windows at every custody station; the fair test, verified from captured feeds. What accumulated debt: the capture loop (session-tied, now to be daemonized), the install handoff, one calendar promise made from documentation instead of measurement, one corrected count. The bookkeeping caught every item — which is the design working: on this architecture the record audits itself, and the plumbing is repaired without a single sealed number moving.

The vital signs, by contrast, are strong. Four seals, stamped in one afternoon on 2026-07-16, re-verified byte-for-byte 4/4 on 2026-08-13. A historical trial that returned a failure — 15/16, 9 false positives — and was published as one the same day, falsifier named in the seal. A discriminant forced into existence by its own worst false positive and frozen, with a 14,707 ppm gap and a 50,189 ppm threshold, before the event it would judge. An 82-row corpus untouched for 28 days. A prediction registered to the second, twenty-seven days in advance, for an eclipse that then arrived on a geomagnetically quiet day (max Kp 2.667) — the cleanest possible grading conditions.

The work is alive. This review covers its first twenty-seven days because that span is complete and can be measured; the prediction stands apart from it, and no finding on this page moves it by one part per million or one second. Condition: alive and armed — the capture loop is running again, the seals re-verify 4/4, and the one open item before the grading window is the daemon install that makes the watch self-sustaining (debt-log item 11). Prognosis: decidable, which is the only prognosis a pre-registered experiment is entitled to. The archive posts around 2026-08-30. The thresholds are frozen. The verdict — pass or fail — will land in these same tables, and if it is late, that too will be logged here, as a failure, with its number.