Corrections log
The Uneven Month — two malformed tags in the page’s markup
The disclaimer box under the markets panel was closed with the wrong kind of tag,
and a second, unrelated stray tag sat further down the page. Browsers silently
repair both, so nothing a reader sees was ever wrong — but the page did not
pass our own pre-publication structure check, and the two errors cancelled each
other in a way that made a simple count of tags look correct. Found by parsing the
page properly instead of counting. Both are fixed; no report text, number, chart or
script changed, and the checksum of the study’s calculation script is
unchanged.
Caught by the Roundtable, while checking why a routine audit was
failing on an already-published page.
Both studies and the Method page — public GitHub mirror of the reproducibility files linked
Each study’s file list gains one closing row linking a public repository that
mirrors the same files byte for byte, with a checksum per file, instructions for
re-running each engine, and a script that compares a re-run with the published
results. The Method page says the same under “What is published, always.”
This site remains the source of record. No report text, number, chart or script
changed.
The Keys to Orbit — reproducibility materials: two scripts corrected
race_model_d3b.py shipped with file paths from the machine it was
built on and would not run anywhere else; multishell_run.py wrote an
earlier wording of two annotation fields than the published results file carries.
Both scripts are corrected. Every output they produce is byte-identical to the
published files, so no number changed. One script,
kessler_stakes_build.py, reads a satellite-catalog export we do not
redistribute, so it documents a derivation rather than re-running from these
files; its table was independently re-derived in review.
Same day: our first corrected multishell_run.py, live for under an hour,
left a file handle unclosed, so a reader’s re-run wrote 1,968 of the 2,000
raw-sample rows. The published sample itself was never affected. The clean-copy
re-run caught it; fixed.
Caught by the Roundtable, re-running every published engine from a
clean copy.
The Uneven Month — reproducibility materials: two labels swapped, one file unlisted
In the file list, the one-line descriptions of coupled_mc.py and
coupled_tail_mc.py were attached to the wrong files, and
coupled_results.json was served but not listed. The provenance note
beneath the list was already correct: coupled_tail_mc.py produced the
published numbers, and coupled_mc.py is the independent
reimplementation that checked it. Labels fixed, file listed. No file and no
number changed.
Caught by the Roundtable, in the same re-run.
The Uneven Month — curiosity-quiz layer (opt-in) added
The same opt-in prediction layer added to the first study: five calls resolve
inline at Figures 1, 4 and 9, in both the paper and energy-realist themes, with a
calibration read before the updates log. The refereed report body and its script
are byte-untouched — additive chrome rendered only for opted-in readers,
plus three id= anchors on figure headings. The homepage card’s
markup changed from a whole-card link to an equivalent stretched-link card to
admit the quiz door; its content is unchanged.
The Keys to Orbit — curiosity-quiz layer (opt-in) added
An opt-in prediction quiz now rides the study: five calls made before reading
resolve inline at the sections and tour beats that earn them, with a calibration
read at the end — no scores, no verdicts. The refereed report body and its
scripts are byte-untouched: the layer is additive chrome, rendered only when a
reader has opted in, plus three id= anchors on section headings.
A reader arriving by any direct link sees the study exactly as published.
The Uneven Month & site chrome — body typeface
A leftover generation-2 @font-face block declared the body family ‘Lora’
but pointed its src at the old Source Serif 4 files; declared after the correct Lora
block, it won the cascade — so every page rendered Source Serif 4 while the design system
specified Lora. Removed the stale block across all four pages; the body now renders Lora as intended.
Chrome-only and additive — no number, chart, caption, or word of the study changed, and the
report’s script is byte-identical.
Caught ourselves during a design-consistency audit.
The Uneven Month — mobile navigation
On narrow screens the study’s chrome navigation links were hidden with no replacement,
leaving phone readers unable to reach Studies, Method, Avenues, Corrections, or About. A
hamburger menu now reveals the same links: an editorial dark-ink dropdown extending from the
(fixed-dark) nav bar, keyboard-operable and reduced-motion aware. Part of a site-wide mobile-nav
fix. Chrome-only and additive — no number, chart, caption, or word of the study changed,
and the report’s script is byte-identical.
A gap we found ourselves while making the site legible on phones.
The Uneven Month — lens toggle legibility
A reader asked, reasonably, whether the energy-realist lens toggle was “just dark mode.”
It is not — it recomputes every figure on the page under fuller energy-cost
pass-through — but the page never said so, and the recolor drowned the recompute. The toggle
now states its action, and while the lens is active a persistent marker labels every view as the
stress case, roughly 2–3× the consensus burden, an adaptation and not published
figures — so a screenshot taken mid-page can no longer travel without its provenance.
No number or model behavior changed; the report’s script is byte-identical.
Prompted by a reader’s question. They were right that it was unclear.
The Uneven Month — invisible text after the redesign
The 14 July restyle stripped dark card backgrounds site-wide; two components — the scenario
control panel and the chart tooltips — were deliberately dark islands whose text was
light, and the pass removed their ground but not their ground-dependent text color, leaving
“Probability of each path (Jun–Dec 2026)” and tooltip text invisible on the paper
page. Both components are restored as dark islands (corners squared only). Root cause: a styling
override applied to a container class without auditing the text colors that depended on its
ground.
Reported by a reader within hours of the redesign shipping. The corrections form
exists for exactly this.
Site-wide — typography and palette
The site and the study page shipped in a typographic costume now widely identified as the
default output of AI design tools — the cream-and-rust palette, letterspaced capital labels,
and a recognizable font pairing served from a third-party CDN. For a publication whose method
page documents its AI architecture, wearing the default undermined the claim that
every choice here is a decision. Replaced with a deliberate system: libre faces from the
Velvetyne foundry and Adobe/IBM (self-hosted, licenses shipped), and a bone-and-bronze palette
that shares a family with the site’s dark ground. No number, chart, caption, or word of
any study changed; the study’s script is byte-identical. Chart chrome recolors via CSS
variables; the data-series colors inside the engine-drawn charts are untouched.
Raised by readers; the criticism was correct and is documented on the
method page.
The Uneven Month — masthead label
The study’s masthead kicker reads “Household Economics · Forecast ·
June 2026.” The word “Forecast” is wrong there. The report contains a
probabilistic forecast in the narrow statistical sense — a predictive distribution for the
magnitude, reported as two bands — but the study is an exposure-and-incidence
analysis with explicitly conditional magnitudes, a distinction the report itself draws and our
standing rules treat as non-negotiable (“not a forecast unless labeled one”). Labeling
the whole study a forecast in its first line overclaims exactly the way we tell others not to.
Per our post-publication protocol, the study body is preserved as published — the kicker
stands uncorrected on the page, and this entry is the correction of record. Any future edition
will carry the accurate label: incidence study.
Caught by the site seat during publication integration; reported against ourselves.
The Uneven Month — reproducibility materials
The published raw-draw sample (coupled_mc_sample.csv) was generated from a
superseded run (crisis probability 0.15) and showed a 99th-percentile burden of
$71.5B/month — disagreeing with the $68.7B carried by the report, the paper, the
workbook, and the results file, all of which use the adopted 0.12. The report was never wrong;
the sample file was the wrong vintage. The file has been regenerated at the adopted knobs and
replaced, and the full 200,000 raw draws are now published alongside it.
Caught in a post-publication sweep by the Economics & Statistics Referee seat,
which reported its own error.