Cut the validation pages, rebuild the charts that could not fail, and give define-uk a paper - #180
Merged
Merged
Conversation
… give define-uk a paper
Seven parallel reviews over the model validation pages, the paper
landings, the chart generator and the one model without a working paper.
**Text.** Every model validation page is shorter: svar 2981 -> 1892 words
(-36%), frb-us 1527 -> 945 (-38%), obr 1827 -> 1347 (-26%), define -23%,
pe -24%, us-hank -16%. The pattern everywhere was one caveat stated five
or six times — on frb-us, "the tracking residual is an identity, not
evidence" appeared in the opening paragraph, the metric note, a
figcaption, both cells of one table and a row of another. Each now lives
in full in exactly one place with pointers to it. Correction narratives
came out; git holds the history and a reader wants the current state.
**A sentence that was flatly wrong.** svar/validation opened its evidence
section with "the one-year global-shock shares land within about a point
of the paper's benchmarks". CPI is 8pp short, and the corrected table sat
directly beneath it.
**The retracted HMRC claim, four more times.** "Within the range of the
ready reckoner" survived in the obr-macro paper's own abstract and
conclusion — contradicting its own comparison table — in the registry
entry served over MCP, in a figure-generator docstring, and on the
catalog page. £6.46bn is 6.4% *below* £6.9bn. A sweep now finds zero
instances in any .html, .tex or .py.
**Two gates that are one gate.** The obr page presented anchored GDP
(0.15%) and anchored consumption (0.25%) as independent CI checks. The
GDP error *equals* the consumption error in £m in all twelve quarters,
max discrepancy 3.35e-06: every other demand-identity term is exogenous
or pinned. Stated once, at the point of claim.
**Charts.** 11 of 14 rebuilt. The flagged win-rate chart drew its count
label six pixels above the wins stack — inside the grey "does not beat
it" segment — and every column was the same height because the stack
always totalled 8, so height encoded nothing. It is a count chart now
with the ceiling as a reference line. `hank-targets` tabulated eight rows
of which six were calibration inputs held fixed through the solve: eight
passes no implementation could fail. It now shows the seven solved
parameters against Table B.III, *including* the vphi row that misses by
17.3%. `frbus-residuals` painted the tracking identity — the row the page
calls not evidence — in the accent colour, so it read as the headline.
`svar-coverage`'s mean line concealed CPI energy at 29% coverage and its
axis floor hid it. Two charts used colour to encode what position already
encoded.
**A CSS bug that silenced every warning in every chart.** `.vchart
.vc-warn` set `color`, which SVG text does not read, so `.vchart text {
fill: var(--paper-dim) }` won and every warning line rendered as muted
grey at the browser default size. The HTML cells sharing that selector
were always fine, which is why it survived.
**define-uk has a working paper**, 28 pages — it was the only model
without one. Its Table 1 splits the gates under a header reading "cannot
fail as evidence of agreement", and argues each at the point of claim:
the ±0.31pp growth gate was set after the run and clears by 0.01pp; the
FMM bands are ours and cannot discriminate; the emissions gate was set at
the observed divergence. It also found that the GPI anchor compares a
policy *input* ("invest an additional 1% of GDP") against a GDP
*response* (+0.92%), that one of the thirteen gap records is our own
modelling choice rather than a manual defect, and that the two SSRN
papers the report conflated are different papers.
The five existing paper landings now emit an identical structure — the
spread in length fell from 161–393 words to 321–339 — and both "Corrected
August 2026" changelogs are gone. The us-hank landscape report is inside
the site chrome instead of being a standalone A4 sheet with its own
palette, which is what made it feel like leaving the site.
96 pages, 1341 site tests, 276 integration tests, every drift gate clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Seven parallel reviews: the model validation pages, the paper landings, the chart generator, and the one model without a working paper.
Text
svar −36% (2981→1892 words), frb-us −38%, obr −26%, pe −24%, define −23%, us-hank −16%. The pattern everywhere was one caveat stated five or six times; each now lives in full in one place with pointers to it. Correction narratives removed — git holds the history.
Errors found while cutting
svar/validationopened with a false summary: "the shares land within about a point of the paper's benchmarks". CPI is 8pp short, and the corrected table was directly beneath it.obr/index.htmlsaid the second-round GDP effect deepens to 0.058%; the artifact says −0.057%.Charts
11 of 14 rebuilt. The win-rate chart drew its count label inside the "does not beat it" segment, and every column was the same height because the stack always totalled 8.
hank-targetstabulated eight rows of which six were calibration inputs held fixed through the solve — eight passes no implementation could fail; it now shows the seven solved parameters against Table B.III including thevphirow that misses by 17.3%.frbus-residualspainted the tracking identity in the accent colour.svar-coverage's mean concealed CPI energy at 29% coverage.A CSS bug silenced every warning in every chart:
.vchart .vc-warnsetcolor, which SVG text does not read.define-uk paper
28 pages. Table 1 splits the gates under "cannot fail as evidence of agreement" and argues each at the point of claim. It also found that the GPI anchor compares a policy input against a GDP response, that one of the thirteen gap records is our own modelling choice, and that the two SSRN papers the HTML report conflated are different papers.
96 pages, 1341 site tests, 276 integration tests, every drift gate clean.
🤖 Generated with Claude Code