Foreign instrument 01 · seismology meets newspaper history
The Newsquake Has No Locatable Epicentre
A seismic location surface run on nine verified 1868 newspaper issue dates cannot find where a poem started, and the same nine towns and the same whole-day dates locate a planted Chicago origin to 41 km. The geometry decides, not the method. Plant an origin yourself and watch the rule find it or lose it.
THE NEWSQUAKE HAS NO LOCATABLE EPICENTRE: nine dateline towns and whole-day dates leave every cell of the search map admissible, and the same rule locates a planted Chicago origin from the same nine towns.
The question is deliberately unfair to both disciplines: what happens if issue dates are treated as earthquake arrivals? The answer is not a guessed point, and it is not a flat refusal either. It is a map of which origins this particular set of nine towns could have found, and the one the newspapers name is not on it.
Touch the result
Nine dates, one surface, no epicentre
Press the button. Your browser profiles origin day and nonnegative slowness at every half-degree cell from Hawaii to the Atlantic. The ordinary rival reads a printed source line and places its ring at San Francisco. Toggle it off if you want the imported instrument alone.
Ordinary source-reading baseline shown at San Francisco.
Computing boundary contact.
Computing the complete residual field.
The colour field is the result. A lone pin would hide that every candidate cell survives the profile threshold. The same threshold, run on the earthquake below, admits three cells out of 1,517.
I · home calibration
Before the newspaper, the earthquake
The engine has to work where it comes from before it is allowed anywhere else. It gets the first event in Lee and Lahr’s 1975 HYPO71 test deck: twenty station coordinates, twenty station delays, a five-layer velocity model, and nineteen picked arrival seconds. It does not get the answer. Everything below is computed in your browser from those inputs and then compared with what the report printed.
1 · the forward model against the report’s own travel-time column
HYPO71 printed its calculated P travel time for every station. This recomputes all seventeen from the printed hypocentre using the same layered ray tracer the inversion uses. The report asks adopters to reproduce its results to one or two counts of the last figure, which here is 0.02 s.
| Station | Range km | Printed TPCAL | Recomputed | Difference | Printed P-RES | Recomputed |
|---|
Recomputing seventeen travel times.
2 · the same grid, profile and region machinery, run on the earthquake
Latitude and longitude are searched exhaustively; depth and origin second are profiled at every cell, exactly as origin day and slowness are profiled at every cell of the newspaper map. The origin set is built with the same threshold and measured with the same diameter routine.
Profiling 1,517 candidate hypocentres.
Running the iterative fit from five starting points.
The newspaper controls stay locked until this passes.
3 · the digit the scan lost
The printed depth on that page reads 0.41 km. The report’s own columns say it is 8.41 km, and the page recomputes the arithmetic that decides it. This matters because a 0.41 km source and an 8.41 km source predict completely different travel times, and the first would have made the calibration fail for the wrong reason.
| What the scan prints | What the report’s own numbers require | Recomputed here |
|---|
Checking the printed digit.
Original phase cards
| Stn | P sec | Stn | P sec |
|---|
S cards: SR04 62.45 · SR06 62.59
What this does and does not establish
It establishes that this page’s travel-time model, its normal equations and its profile-and-region machinery recover a published hypocentre in the field they came from. It does not port HYPO71: magnitude, the formal error estimates ERH and ERZ, and the program’s reading-quality and residual weighting are not implemented here, which is why the depth and the RMS land close rather than exact. It says nothing at all about issue dates as arrivals, straight-line newspaper travel, unknown speed, day censoring or archive coverage. Those are the collision under test.
II · adapter audit
Every thing lost in translation, and what removing it costs
The loss is larger than the analogy. An earthquake station has a clock and a calibrated velocity model. A newspaper issue has a civil date after unknown receipt, selection, typesetting, cadence, recopying and delay. The fit pushes all of that into one geographic residual.
Each row below can be disabled. Disabling one changes the data the engine consumes and runs the whole pipeline again in your browser: the grid, the nine deletions, the twenty-five starts and the verdict. The number that appears is measured, not stored. Two rows cannot be run at all, and say so instead of printing a figure.
| Source field | Formal object | What this destroys | Ablate | Measured |
|---|
Dropped from the numerical fit
All source attributions, author-place phrases, datelines, transcribed text after membership review, proprietary witnesses, duplicate page fragments, typography, neighbouring content, recurrences outside the window, and every newspaper absent from digitized archives. Postal, rail, telegraph, exchange-list and copying edges are also absent. Those missing network edges are the central model failure.
Conventions supplied by hand
Each issue date is a local civil-day interval represented at noon. The onset window is 180 days from 26 February 1868. One canonical issue page is one observation. Coordinates are reviewed modern representative points. Distance is spherical great-circle distance on the WGS84 coordinate frame. Depth is fixed to zero. Slowness cannot be negative. The map is 18 to 50 N and 161 to 65 W at 0.5 degree spacing. Seed 400732 fixes 2,000 title-block bootstrap replicates. Each of the nine fitted titles appears exactly once, so a title block and a row are the same object here and the block bootstrap does nothing a row bootstrap would not.
| Paper | Issue date | Reviewed place | Role | Page fragments |
|---|
III · the positive control
Plant an origin and see whether the rule can find it
A verdict that no data could overturn is not a verdict. So: choose a city, choose how precisely the dates record the arrival, and the page generates an arrival date at each of the same nine towns from that origin under its own forward model, censors each one the way a printed issue date censors a real arrival, and hands the result to the same untouched pipeline. Then it reports whether the rule found what you planted.
Waiting for the native calibration.
Choose an origin and press the button.
The ladder, precomputed by the same code path
Generated by node research/the-newsquake/precompute.mjs and recomputed by the page’s own check. Every row here is reachable from the controls above.
| Planted origin | Date precision | Recovered | Error | Origin set | Contains the plant | Verdict |
|---|
What the control shows, including against this page
Three things, and the third is not flattering. First, whole-day dates at these nine towns can locate an origin: a planted Chicago source is recovered to 41 km with an origin set 187 km across, and the rule returns HELD. Second, the same nine towns and the same dates cannot locate a planted San Francisco source at all, because eight of the nine sit east of the Rockies and the ninth is Gold Hill: the distance pattern westward is nearly flat, so the surface has nothing to bite on. Third, when the dates are made precise enough for the set to become compact, the compact set stops containing the planted origin. The profile threshold is a diagnostic and not a calibrated confidence region, and this is what that costs: at one-hour precision the San Francisco set is 176 km across and misses the truth by 175 km.
IV · fault bench
Plant what it claims to notice
One fault changes an arrival by thirty days. A residual locator ought to flag Gold Hill as the largest externally studentized residual above 3.0. The other changes a name inside the transcribed text that travels with every fitted row, so if the adapter is honest about dropping text, no number should move.
Choose a planted fault.
The predeclared date-fault rule is not relaxed after inspection. A miss is a miss, and it is reported as one in the transfer ledger.
V · the sophisticated dismissal
Maybe the point failed, but the uncertainty still knows something
That is the stronger version of the method, so this layer tests it. The full profile set, title deletion, onset-window sensitivity, the bootstrap boundary rate and the local-start audit all have to agree before a geographic source is identifiable.
Remove one newspaper title
The gold rule marks 500 km. A stable locator keeps every bar left of it.
Change the onset window
| Window | Rows | Minimum | RMS days | Origin set | Diameter |
|---|
The 30-day window admits three issue pages and the fit has four free quantities, so there is no residual scale and no region to report. Printing an inferential set there would be a number with nothing behind it.
Checking the compactness threshold.
Frozen title-block bootstrap, recomputed by the page’s check
Loading the 2,000-replicate summary.
Each title is sampled as a block and each date is jittered uniformly inside its civil day. With one row per title, that is a row bootstrap. This is the only figure on the page that is not recomputed in your browser: it is generated by node research/the-newsquake/precompute.mjs and recomputed from seed 400732 by the independent check.
Twenty-five iterative starts
Computing the local-start audit.
The exhaustive profile adjudicates, and it needs no starting point, so this audit is reported beside the verdict and not folded into it. Folding it in was how the verdict became unable to say anything else: with dates generated from a real origin by this page’s own model, everything else passes and only this check fails. The same optimizer reaches the earthquake minimum from all five of its starting points, so what is start-dependent here is the newspaper objective and not the solver.
VI · independent target check
Read the line the fit never saw
The ordinary baseline uses the earliest reviewed source line, in the New Orleans Crescent on 8 March. Three further newspapers are held out of every numerical fit and only appear once the date-surface state is locked. All three name the San Francisco Times, in the same sentence, with one word different.
How independent that really is
Two things weaken it, and both are in the table below. The three held-out pages print one sentence, not three: word for word the same, except that Bellows Falls prints Mr. H. A. Deming where Rutland and St. Johnsbury print Mrs. That is one attribution carried by three physical objects, not three independent testimonies. And the same claim is printed on the pages the fit does use: seven of the nine fitted issues name San Francisco outright in their frozen crop, in text the adapter throws away before the first arithmetic happens, and the two that do not are the two whose crop begins below the naming clause. Holding three pages out of a fit that never reads text is a formality, not an experiment.
What survives is narrower and still worth having. There are two distinct wordings, not one. Six pages say the poem is a contribution to the San Francisco Times, which names a publication. Four say Mrs. H. A. Deming, of San Francisco, which names a person’s city. Two crops begin below the naming clause and show only the tail of the second sentence. Two wordings agreeing on a city is better evidence than one wording repeated twelve times, and it is still not a San Francisco Times issue, which nobody here has seen.
| Paper | Date | Role | Wording | What it names |
|---|
VII · the check
What can actually go red
Uncertainties that remain
Issue date is interval-censored and is not receipt time. The archive omits unknown newspapers and excludes proprietary witnesses; admitting the eleven in-window rows the eligibility rules throw away moves the minimum 6,471 km, and admitting only the earliest of them, the Hawaiian Gazette of 26 February, moves it 9,952 km. Modern representative coordinates are not 1868 press offices. The three held-out attributions share one wording. Great-circle distance ignores transport and editorial networks. The profile threshold is a diagnostic and not a calibrated confidence region, and the synthetic control measures how badly: its compact sets can exclude the origin they were generated from. The 0.5 degree grid limits spatial resolution, and the search box excludes the Atlantic, so an English origin could never have been found. No issue of the San Francisco Times has been seen.
Free choices fixed before the result
Cluster 400732 · 180-day primary window · 0.5 degree grid · 18 to 50 N · 161 to 65 W · nonnegative slowness · one issue page per observation · title-block bootstrap · 2,000 replicates · seed 400732 · 500 km compactness · 500 km deletion stability · residual fault threshold 3.0 · home tolerances 1.0 km, 0.02 s and one percent of the box.
Sources
- W. H. K. Lee and J. C. Lahr, HYPO71 (Revised), USGS Open-File Report 75-311, June 1975. The calibration uses the Appendix 2 card deck and the Appendix 3 printed results, file page 95.
- Viral Texts cluster 400732, used for discovery and for checking witness identifiers, page fragment counts and the Coverage coordinate proposals. The unlicensed cluster CSV is not shipped.
- Library of Congress, Chronicling America, twelve public-domain newspaper issue images. Exact page URLs and image-service URLs are recorded in the frozen extract.
- 2025 United States Census places gazetteer, representative coordinates for reviewed publication places.
- GNIS populated places, public-domain coordinate for Gold Hill, Nevada, feature 858000.
- Ryan Cordell, Two (of Three) Ways of Looking at C19 Newspaper Exchange Networks, the closest target-side precedent.
- Pedro C. Pinto, Patrick Thiran and Martin Vetterli, Locating the source of diffusion in large-scale networks, 2012, a stronger network-aware source-localization precedent.
We searched the general web, Crossref and OpenAlex on 10 September 2026 for combinations of “hypocenter inversion,” “earthquake location,” “arrival time residuals,” “newspaper reprinting,” and “information diffusion.” We found extensive work on geographic reprinting and network source localization, but no direct application of iterative seismic hypocentre least squares to nineteenth-century newspaper reprint dates. A Semantic Scholar API attempt returned HTTP 429.
Built from frozen public-domain newspaper crops. Nothing is uploaded, tracked or fetched from a third party at runtime. The check that stands behind this page recomputes the visible figures and the six-cell transfer ledger.