US topic pages, an Overview tab, and four papers corrected against their own artifacts - #178
Merged
Conversation
…ibration inputs a replication The boe-svar card had grown to three sentences and 30 words against a 10-16 word grid, because the FEVD correction was written into it as prose rather than summarised; it is the card people read first and it read like a footnote. Same claim, half the words, link to the page that carries the detail. The us-hank card said it 'hits every Auclert et al. (2021) calibration target'. Those targets are calibration *inputs* — Y=1, K=10, r=1.25% are imposed and the parameters are backed out to make them consistent, so 'target 10, achieved 10' is a constant equalling itself. The replication is that the solved parameters reproduce the published Table B.III to four decimals, which is what the card says now. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…erving it as evidence The FRB/US review established that the 5.6e-17 baseline tracking residual carries no economic information: `init_trac` sets each add-factor to minus that equation's residual at the input data, so the solve is algebraically the identity on whatever it was tracked to, for any input at all — and the model repo now gates the demonstration, passing the same check at 6.7e-9 on a baseline scrambled by random factors with every accounting identity destroyed. The validation page body was rewritten to say so; eight other surfaces were not, including the section heading directly above the corrected body, four meta descriptions, the methodology diagram label, and the results-table cell that still graded it "machine precision". Worse, `frbus_summary` served it as a validation `value` with a `gate` and no qualifier, and the registry counted it as implementation-fidelity evidence — so a consumer reading validation off the API got the retired headline as the strongest-looking number in the payload. The value stays (it is a real software check and the gate is worth keeping) but now ships with `is_evidence: false` and the reason, and the registry's fidelity judgement rests on the pyfrbus shock comparison alone. Also from the catalog audit: the DEFINE registry entry still said "some thirty-five defects in the manual", the last place on the project carrying that count — thirteen are pinned as machine-readable gap records, and an audit re-characterised four of the wider set as first-period jumps and withdrew one claim. And us-hank/validation said the suite "gates the steady-state calibration targets" four lines above a figcaption correctly calling the top five of those rows imposed inputs; the prose now leads with the solved parameters, which are the actual replication. Separately, on copy the user flagged: the home hero caption, the /connect limits block, and the data-store as-of recipe are each about half their previous length, with no caveat dropped — the ones that mattered moved to fewer words rather than disappearing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…r own artifacts Five reviews ran in parallel over the papers, the model catalog and the economy section. Everything below is a claim the site was making that its artifacts do not support, or a structure the two countries did not share. **The economy section had no Overview tab.** Both hubs showed a topic strip with nothing marked current, above four sections the nav never named — you landed on a tab set with no selected tab. `Overview` now leads the strip on both countries and carries `aria-current` on the hub. **The US has topic pages now.** Five of them, from one generator: `UK_TOPICS` and `US_TOPICS` are two tables feeding one renderer, so the template, the four-layer order and every layer builder are shared rather than forked. `/economy/us` is an index rather than a second dashboard, matching the UK hub band for band, and every reading it used to print is now on the topic page that owns it, with its period and vintage. Three gaps are stated on the page rather than left silent, and two of them are *checked*: there is no US public-finances page because no fiscal series exists in the store and no US model answers a fiscal question, and the generator raises if such a series ever lands; there is no US release calendar because FRED stores neither `release_updated` nor `next_release`, and the generator raises if either is ever populated. No US page shows a forecast with a range — `no_us_forecaster()` re-derives that from the registry on every render and raises if a US forecaster appears, so the refusal cannot outlive its reason. **A live data bug, found while building it.** `us_cpi` and `us_unemployment_rate` are each missing 2025-10, and the year-on-year helper compared by list *position* — so `observations[-13]` on the CPI index landed on 2025-05 and reported 3.7% where the true 2026-06 figure is 3.5%. Now matched on period. The UK pages are byte-identical under the new code because every UK series is gapless, which is exactly why it survived. **FRB/US: the retired claim was still the headline of its paper.** The 0.99 pegged-funds-rate multiplier was corrected in `results.json` last week and in no `.tex` at all; the figure generator hardcoded `0.72 / 0.99 / 0.32` as literals, which is how a corrected artifact and a wrong chart coexisted. The year-two multiplier is 0.904 and sits *below* the 1.1–1.2 accommodation range. Two labels were also exactly backwards: the shocked pyfrbus comparison — the one real piece of evidence — was marked "exact by construction" while the add-factor identity was presented as validation. **OBR: the anchored fit is one test, not two.** The anchored GDP error *equals* the anchored consumption error, in £m, in all twelve quarters (max discrepancy 3e-6). Every other demand-identity term is exogenous or pinned, so two gates that read as independent evidence are one test of one equation. Separately, "consumption crowded out by £1bn cumulative" was −£0.96**m** in the final quarter and −£4.27m cumulative — a £m figure read as £bn and called cumulative, overstating a null result by more than 200x. And the free-running scorecard had been stale since the OSHH anchor, understating the model on every line. **boe-svar: a claim that reversed on inspection.** The paper said importance weights "move both shares substantially towards the published values"; on the corrected statistic they move GDP away, and the fast-vs-production gap is Monte-Carlo noise in 51 accepted draws. Figure 4 was still the superseded sum-of-medians statistic; it is now the posterior mean of per-draw shares of total variance, at four quarters rather than five. **psl-og: the figure showed a fit the model never produced.** It plotted the labour share 60.0 vs 59.5 where the paper records a ~5pp miss, plotted the saving ratio by copying the official value into the model column, and used the OBR *forecast* as the model's debt value. Rebuilt as a difference chart: imposed rows grey and tagged "cannot disagree", the saving ratio with no bar at all. Also: the us-hank landscape report credited a **fork** to Bayer–Born–Lütticke — 4 stars against the authors' 70, understating it seventeenfold in the column its verdict rests on; define's "27 identity tests" is 62 collected, 44 of them identities; the emissions divergence was computed from rounded levels; and the one test that recomputes the £6.46bn costing asserted `4.0 < x < 9.0`, a band admitting both HMRC figures, so it could not distinguish our costing from the benchmark. It pins ±5% now, and names every surface that has to move with it. Two more hardcoded developer home directories fixed — four this round. Each one is why a bad artifact could survive: nobody else could regenerate it. Papers rebuilt: frb-us 36→38, obr-macro 36→40, boe-svar 30→33, pe-microsim 23→25 pages. The page-count gate had been missing pe-microsim entirely. 1332 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.
Five parallel reviews over the papers, the model catalog and the economy section.
Structure
UK_TOPICS/US_TOPICSfeeding one renderer, not a fork)./economy/usis now an index rather than a second dashboard, matching the UK hub band for band.no_us_forecaster()re-derives this from the registry every render).Data bug
us_cpiandus_unemployment_rateare each missing 2025-10, and the year-on-year helper compared by list position — reporting 3.7% where the true figure is 3.5%. Now matched on period. UK pages are byte-identical because every UK series is gapless, which is why it survived.Papers
.tex; the figure generator hardcoded it as a literal. It is 0.904 and sits below the 1.1–1.2 range. Two labels were backwards: the real evidence marked "exact by construction", the identity presented as validation.Also
The us-hank landscape report credited a fork to Bayer–Born–Lütticke (4 stars vs the authors' 70). Define's "27 identity tests" is 62 collected, 44 identities. The one test that recomputes the £6.46bn costing asserted
4.0 < x < 9.0— a band admitting both HMRC figures — and now pins ±5%. Two more hardcoded home directories fixed, four this round.Papers rebuilt: frb-us 36→38, obr-macro 36→40, boe-svar 30→33, pe-microsim 23→25. The page-count gate had been missing pe-microsim entirely.
1332 site tests, 276 integration tests, every drift gate clean.
🤖 Generated with Claude Code