adds-up

adds-up

Reads a price list and reports the arithmetic in it that contradicts itself.

Runs entirely on your machine. No account, no API key, no upload, no network call, no dependencies beyond Python itself.

What it found across 29,521 real published tariffs, and the caveat that matters more than the finding, is written up in Thirty correct findings, and only one looked like a mistake.

$ adds-up examples/volume-tiers.csv
adds-up 0.1.3: examples/volume-tiers.csv

6 row(s) under a header read from row 4, the only header-shaped row above the data with data under it; 3 row(s) above it were read as a title block.
5 of 5 column(s) were given a role.
3 of 7 check(s) ran, comparing 4 pair(s) of numbers. 1 finding(s).

VOLUME-TIER
  a larger quantity priced higher than a smaller one, in the price column named

  'W-100' has a Unit price of 11.80 at Quantity 250, and 11.20 at Quantity 50
      row 6: Quantity 50, Unit price 11.20
      row 7: Quantity 250, Unit price 11.80
      at volume-tiers.csv!row 6, Unit price, volume-tiers.csv!row 7, Unit price

CHECKS
  ran      volume-tier          1 finding(s) from 4 comparison(s): read 'Quantity' against 'Unit price'
  DID NOT RUN discount-tier     no column names a discount percentage
  DID NOT RUN bundle-above-parts no column names the items a bundle is made of, so no composition is stated in this document
  DID NOT RUN stated-discount   no column names a list price
  ran      two-prices           0 finding(s) from 0 comparison(s): read 'SKU' against 'Unit price', holding 3 other column(s) equal
  DID NOT RUN inverted-range    no column names the bottom of a range
  ran      unit-mismatch        0 finding(s): read 'Currency' against 'SKU', holding 2 other column(s) equal (compares text, not numbers)

  A check that did not run has found nothing because it was not made.

COLUMNS
  'SKU'                              item               'sku' is a known name for item
  'Description'                      item               'description' is a known name for item
  'Quantity'                         quantity           'quantity' is a known name for quantity
                                                        numbers: no cell in this column carries a separator
  'Unit price'                       unit_price         'unit price' is a known name for unit_price
                                                        numbers: 6 cell(s) in this column can only be read as 1,234.56
  'Currency'                         currency           'currency' is a known name for currency

The thing to understand before anything else

A price list does not state a cost.

That one sentence decides everything this tool is. Almost every question a person wants to ask about a price list (is this profitable, is it competitive, is the discount too generous, should tier 3 be cheaper) needs at least one number the document does not contain.

So this tool answers none of them, and it is built so that it cannot look as though it might. There is no margin here, no benchmark, no score, no severity, no recommendation. It will never tell you a price is right, wrong, high, low, fair or profitable.

What is left is the set of statements a document makes that cannot all be true at once. "Tier 3 costs more per unit than tier 2" is a finding. "Your pricing is wrong" is not, and never will be.

The reasoning, and every check that was considered and rejected, is in docs/superpowers/specs/2026-09-09-adds-up-design.md.

The second thing: a finding is not a defect

The tool reports that two numbers in a document cannot both describe one consistent price. It does not know which is wrong, or whether either is.

The corpus proved this the hard way. Inclining block pricing, charging more per unit as you use more, is deliberate and everywhere in utility tariffs. Of thirty volume-tier findings read by hand against 29,521 published tariffs, all thirty were arithmetically correct and twenty-nine were not mistakes: they were pricing somebody meant to set. The check says a larger band is priced higher, in those words, and stops.

Install

pip install adds-up

Python 3.11 or newer. Nothing else. To install from a clone instead, pip install . from the repository root.

Use

adds-up prices.csv
adds-up prices.xlsx --sheet "Trade rates"
adds-up prices.csv --format markdown       # to paste into an email
adds-up prices.csv --format json           # for a pipeline
adds-up prices.csv --skip volume-tier
adds-up prices.csv --header-row 3          # the header is not where it looks
adds-up prices.csv --map 3=unit_price --map 'Band C'=quantity
adds-up --list-checks
adds-up --list-columns                     # every column name it recognises

Exit codes: 0 nothing found, 1 at least one finding, 2 could not run.

Exit 0 requires that two numbers were actually compared. A file that did not parse exits 2, and so does a run in which nothing was compared: every check switched off, no check able to find its columns, or the columns it found empty. The report prints the comparison count, so "3 checks ran" and "3 checks ran and compared nothing" are never the same line.

That gate is on comparisons rather than on whether a check reported that it ran, and the difference is not theoretical: an adversarial review found a price list with a real inverted tier in it exiting 0 because its price column had a heading the vocabulary did not know, and unit-mismatch, which needs no numbers at all, had run and found the units consistent. A green result over a price list nothing was compared in is the one outcome this tool exists not to produce.

The seven checks

Each can be switched off on its own, and the report always prints which ran and which did not, with a reason.

Every one of these compares two things printed in the input, and six of the seven compare numbers. unit-mismatch compares text (a unit code against a unit code) and the report says so beside it, because exit 0 is earned by arithmetic and that check does none.

None of them uses a cost, a margin, a benchmark, a market rate or a currency conversion, because a price list states none of those.

Three narrowings that cost coverage on purpose

A percentage has two opposite meanings, and only its name decides which. 20% off and pays 20% of are contradictory instructions and both appear in real price lists: the second is how the US hospital price-transparency schema states a negotiated rate. So there are two roles, and the column's name picks one. Trying both readings and reporting only when neither fits was rejected: a check that accepts whichever reading happens to fit can never fail.

A percentage is only compared with the price it is a percentage of. Where a document holds several net-price columns, the percentage pairs only with the one that is its own name with the last word changed (…|negotiated_percentage with …|negotiated_dollar), unless there is exactly one candidate. A published file puts a discounted cash price beside that pair, and a gross charge times a negotiated percentage has no reason to equal a cash price; on files that leave the negotiated dollar column empty throughout, falling back to the nearest other price produced a finding on every row. If nothing qualifies, the check declines to run and says so.

An undecidable number is refused, not guessed. 1,234 is 1234 in one convention and 1.234 in the other. The convention is decided once per column from the cells that are unambiguous, and a column with none is refused: the checks that needed it do not run, and the report says so. On the tariff corpus that costs 316 of 29,521 documents: rate sheets in which the only cells carrying a separator are of that undecidable shape, so nothing in the column settles it.

How it decides what a column is

A price list has no schema, so the tool reads the word at the top of each column and looks it up. Roles come from names, never from the shape of the contents. A column of numbers between 0 and 100 is equally likely to be a discount, a case pack, a quantity break or a stock figure, and picking one on that evidence is how a tool invents a contradiction.

adds-up --list-columns prints the whole vocabulary, because a vocabulary you cannot read is indistinguishable from a guess. A name it does not know gets no role, the checks that needed it do not run, and the report names the word it was looking for. --map settles anything it got wrong.

One ambiguity cannot be settled by the name at all: a column called max is the top of a volume band in a tier table and the top of a price range in a rate sheet. The rule is that a bound is a range end only when a column naming the other end is present, and a quantity otherwise, and the report prints which reading it took.

Where several columns claim one role, the leftmost whose numbers can be read wins. A published standard-charges file carries a cash price and a negotiated price side by side, and a hospital that publishes only the second leaves the first empty on every row; taking the leftmost regardless had the check reporting that it could not run with a usable column beside the one it had chosen.

The header row is chosen the same way: by names. Among the header-shaped rows in the first 25 that have data under them, the one naming the most columns this tool recognises wins, and the last one only breaks a tie. That handles a title block above the header, which is what published files usually have, and a units row or category banner below it, which is what defeats taking the last one. The report prints which row it used and why, and --header-row overrides it.

Format scope

CSV, TSV and XLSX. Not PDF. A PDF price list is a page of positioned glyphs, and rebuilding rows and columns from it is a separate problem with its own error rate; every finding would then carry two uncertainties with no way to tell them apart. The tool says it does not read PDFs rather than reading them badly.

The .xlsx reader is zipfile and xml.etree and nothing else, including the number formats, because a cell storing 42.50425 and displaying 42.50 must set the tolerance from what it displays, or the discount check is wrong on every formatted workbook.

It does not collide with show-your-work

Both read spreadsheets; they ask opposite questions. show-your-work asks how the workbook was built: a formula someone typed over, a SUM that stops a row short, a hidden sheet, a stale link. It needs formulas.

adds-up never looks at a formula. It reads values as printed and asks whether the pricing they state is self-consistent, which is why it works on a CSV that has no formulas at all. A workbook can be perfectly built and price tier 3 above tier 2; a workbook can have a broken SUM and price everything consistently. Run both.

Measured against real price lists

Unit tests pass on the inputs their author imagined, which proves very little. This was run over two corpora of real published documents, neither written by this project. Full provenance, hashes and selection rules: docs/corpus-manifest.md.

tools/measure_corpus.py, tools/measure_hospital.py, tools/probe_gtfs.py and tools/audit_hospital.py are the scripts that produced every number below.

Corpus 1: 29,521 published utility tariffs

The OpenEI U.S. Utility Rate Database, retrieved 9 September 2026, SHA-256 9005c535…d28d5. Every rate structure with two or more tiers, from 18,049 distinct tariffs published by 2,062 utilities.

Accuracy: 30 of 30 correct. Thirty findings drawn at seed 11, one per tariff, each read against that tariff's own printed tier table. Every one quotes the two rates the file prints, in the right order. None was wrong.

A stronger check than the sample. The measurement script classifies each tier ladder's direction independently of the tool (rising, falling, mixed or flat) straight from the raw export. The two agree almost exactly:

The 316 documents whose rate column was refused break down as 313 falling, 2 rising and 1 flat, which accounts for the two rising ladders with no finding exactly. Nothing falling or flat was reported anywhere in the corpus.

And 29 of the 30 are also not mistakes. This is the number that matters and it is not a flattering one:

• 11 of 30 have a first band priced at zero: an allowance included in a fixed charge, or the threshold above which a demand charge starts. That is not an inverted tier in any useful sense. The tool prints a note on these; across the whole corpus they are 4,404 of the 15,487 findings, 28%.
• 18 of 30 are inclining block tariffs, which are conservation pricing and entirely intentional. 6,254 of the 10,020 residential rate structures in this corpus produce a finding, against 4,726 of 14,565 commercial ones.
• 1 of 30 looks like a data defect rather than a pricing choice: a seven-tier industrial ladder whose thresholds run 3,000 / 10,000 / 200,000 / 200 / 400 / 600, which is two rate blocks concatenated into one.

So on this corpus the check is 100% arithmetically right and about 3% useful. Both halves of that are worth publishing.

So this corpus establishes that the check's arithmetic is right at scale, and establishes nothing whatever about how often an inverted tier is a mistake. A utility tariff is the wrong document to ask that of. Said here rather than left for a reader to assume the reverse.

Corpus 2: US hospital standard-charges files

Files every US hospital is required to publish under 45 CFR 180.50, in the CMS template's CSV shape, collected 9 September 2026. This corpus exists because the tariff corpus exercises exactly one of the seven checks, and because it is the only source found where thousands of column headers were written by people who have never heard of this tool.

It has now been run twice. The first run's findings were hand-read back against their files, that audit found three defects in the tool, and the numbers below are the second run with those fixed. Both runs are reported, because the difference between them is the most useful thing this corpus has produced.

27 of the 200 files are not CSV, and the tool now says so. They are the CMS schema's JSON form, served from a URL ending .csv. Read as delimited text, one of them is a single row of up to 1.16 million comma-separated fields; every check then fails to find a column it can use and the run exits 2. That was already the right outcome and nothing was ever invented out of one, but the reader had to infer this is not a table from a header a million columns wide. A file whose first non-blank character is { or [ with a JSON key after it is now refused by name, with a message saying so. It is a shape test and not a parse, so a price list that opened that way and had a quoted colon early in it would be refused wrongly.

That also settles which column figures are worth quoting. The 173 real CSVs hold 13,979 columns and 8,519 distinct column names; 8,519 is the number this corpus was collected for. The first run, which read the 27 JSON files as tables, reported 6,852,483 columns and 661,590 distinct names, 99.8% and 98.7% of them JSON punctuation. Refusing those files removes the figure that could be quoted for the other.

190,549 rows were set aside for being the wrong width, and that line in the report is new. 190,547 of them carry more cells than the header line; 2 in the whole corpus fail the opposite way, not reaching every named column. 19 files hold at least one.

180,085 are a single file. Sky Lakes Medical Center leaves commas unquoted inside free text, so 180,084 of its 1,342,453 rows parse to more cells than its 29-column header and every value after the break sits under a heading that is not its own. The first run read all of them and said nothing; this one reads none of them and says how many. After Sky Lakes the largest are St Joseph of Nashua (7,043 of 45,242) and Mercy Waldron (2,352 of 478,748).

The rule is judged against the header line, not against the headings that have names, and both halves of that were paid for. One file ends its header line in a comma (25 headings of which the last is empty, above 1,444,617 rows of 24 cells), so a row that stops short of an unnamed column is fine. Another ends its header line in three commas and puts a value in the last of the columns those commas open, so those unnamed columns cannot simply be dropped. A first version of this rule did drop them and emptied that file of all 27,356 rows.

Which checks ran

The same three columns are missing from all 173 CSVs: a quantity, a volume band, and a bundle's composition. volume-tier, which the tariff corpus exercised 29,205 times, has nothing to work with in a hospital file at all; discount-tier and bundle-above-parts have now failed to run on either corpus and remain untested against any real document. It is not that hospitals price those things consistently. It is that the document does not state them, so there is nothing to contradict.

stated-discount ran on 70 files rather than the first run's 75. Three of the five it lost hold nothing but percentage columns every value of which lies between 0 and 1 (which is 85% in one writing and 0.85% in the other, with nothing in the document to settle it), so every one of them is refused and the check declines to run. 22 percentage columns in 4 files are refused for that reason, out of the 1,623 columns the run gave a percentage role; the fourth file has other percentage columns that are readable, and the check still runs on it.

Findings, and how few documents produce them

Every one of those totals is more concentrated than it was, because what the fixes removed was concentrated: one file's ragged rows, one file's fractions. inverted-range is now four documents and its top three are all of it. 470,549 is still three hospitals: Nazareth 342,918, Hancock County 85,797, Vibra Central Dakota 29,485, 97.4% between them, against a median document of 240. The number is real and it is not a rate: it says three files are enormous and repetitive. Every count above is published beside its worst document for that reason, and --per-document writes the full list.

two-prices is unchanged at 189,501 and is the least concentrated of the four, and there the concentration is not really about documents. Four of its five largest (Sycamore Shoals, Hancock County, Russell County and Lonesome Pine, 117,960 findings between them) share a column layout, a notes wording and the same duplicated rows: BLADE RESECTOR 3.5, CDM code 27200002, appears in all four. That is one vendor's export published by four hospitals, not four independent measurements of anything.

The hand audit, and what it says now

Every drawn finding was read back against the file it came from, at the row and column it names, in both runs. tools/audit_hospital.py does the mechanical half: it re-reads the CSV independently, and (added because of what the first audit found) it now also confirms the cited row is the header line's width and re-derives the arithmetic the finding printed, saying which reading of a percentage column reconciles.

The draw is one finding per document at seed 11, capped at 30, so for three of the four checks the sample is every document that produced the finding at all. That sounds comprehensive and is not: those are 4, 12 and 16 documents.

Weighted by findings rather than by documents, counting only causes that were established across a whole file rather than sampled:

The two arithmetic identities are the strongest evidence here: the corpus-wide finding counts fell by exactly the numbers the first audit had proved false, and by nothing else.

Read the zeroes as what they are. They mean no finding in the samples above was wrong, and that the three causes measured file-wide are gone by construction. They are not a claim that the rate is zero. inverted-range now rests on four documents; unit-mismatch on twelve. A fourth cause, a qualifier column the vocabulary does not know, remains unfixed, and has not been seen in either audit of this corpus.

What the drawn findings did contain, six times, was a pair of rows differing only in a column computed from the price they disagree about: an estimated_amount, or a standard_charge|negotiated_algorithm whose text quotes the gross charge back ("78% of $604.49 in Billed Charges"). A difference there is the same fact as the finding, not an explanation of it, and every one of the six was read by hand to confirm that before it was allowed to stand.

One fix removed findings the audit had judged correct. Vibra Central Dakota fell from 120,673 stated-discount findings to 29,485. The 91,188 it lost came from percentage columns holding 0.8 throughout, which are now refused, and the first audit had judged Vibra's sampled finding correct, because that file's negotiated_dollar column holds the gross charge on 120,698 of its 120,728 rows, so the rows do contradict themselves whichever way the percentage is read. Refusing the column stops the tool saying so. That is the price of not guessing a convention the document does not state, and it is a real price, not a rounding.

Right, and not a defect

The other half of the tariff corpus's lesson repeats here. All 16 stated-discount findings read this round are true statements about the cells they quote, and most of them are still not mistakes:

• Four are rounding or an uplift. 31.00% of 4.0000 = 1.24 printed beside a negotiated dollar of 1.25; 100.00% of 1.0000 beside 1.01; 64% of 0.765 = 0.4896 beside 0.48. Three files state in their own standard_charge|negotiated_algorithm column, in words, that the negotiated dollar is a percentage of the gross charge and then uplifted: Hancock County's cluster on 1.05, 1.06 and 1.02, which is what that column says. The tool does not read it, and would be guessing if it tried.
• That pattern is most of the corpus total. Read back with a flat 0.5% tolerance, 229,392 of Nazareth's 879,195 percentage/dollar/gross triples fail, and 99.99% of those have a dollar between 0.97 and 1.02 times the percentage of the gross charge, not a contradiction anybody would act on.
• The rest are gaps nothing in the document explains, and some are plainly defects: a percentage column filled with 100 beside a dollar that is not the gross charge, a negotiated dollar of 52.9 printed identically into the percentage column beside it, and (at Nazareth) 64% of $91.00 stated as $58.42 where it is $58.24, which is two digits transposed.

What this corpus cannot establish

It says nothing about how often a hospital is wrong. It says what the document states, and how often the tool reads that correctly. A stated-discount finding is a percentage that does not produce the price printed beside it; on this corpus that is usually a rounding step or an uplift the file describes in prose, and only sometimes an error.

It cannot measure three of the seven checks, and it has now established why. volume-tier, discount-tier and bundle-above-parts need a quantity, a volume band or a composition, and no file in this corpus publishes any of them, so volume-tier rests entirely on the tariff corpus and the other two rest on nothing but their unit tests. The currency half of unit-mismatch is in the same position: no currency column appears in either corpus.

A zero false-positive rate is a statement about a sample. Where a rate here comes from a draw it says how many were read, and three of the four draws are every document that produced the finding at all, which sounds comprehensive and is not, because those are 4, 12 and 16 documents. The per-file arithmetic (Sky Lakes' 180,084 ragged rows, Collingsworth's 11,240 fractions, Vibra's 91,188 suppressed findings, Kula's 17 case-only units) are counts over named files, not estimates over the corpus.

It is biased towards small hospitals, by the 40 MB download cap described in the manifest, and towards whichever vendors those hospitals use: four of the six documents that produce most of the two-prices findings publish one vendor's export, down to the same duplicated row.

200 files is not the 5,023 in the index they were drawn from, let alone every US hospital. Nothing here establishes what fraction of published standard-charges files contain any of this.

What the measurement found that the tests did not

Every one of these was a real defect, found only by meeting real published files. The suite was green throughout.

The counts in this section are not reproducible. They were taken against the corpora named above with earlier versions of this tool, and the current tools/ scripts cannot produce them, because the defects they measure are fixed. They are a changelog with evidence attached, not a measurement, which is exactly the distinction this repository otherwise insists on, so it is stated rather than left for a reader to discover that the numbers do not re-run.

Three of them were found by hand-reading the first hospital run's findings back against their files: 66 findings, one per document for each check. The suite was green, both corpora had been measured, and an adversarial review had already been through the tool. They are listed first because they are the largest of everything on this page, and because the same thing has now happened on five consecutive tools in this portfolio: meeting real files finds what tests do not.

• A row with more cells than the header was read anyway, and silently. This was the largest single source of wrong findings ever measured here. An unquoted comma inside a free-text field shifts every value after it into the next column, and the tool went on reading whatever now sat under each heading. One published file, Sky Lakes Medical Center, carries 180,084 such rows out of 1,342,453; every one of the 1,517 rows in it where the minimum exceeded the maximum was one of them, and not one well-formed row in the file had that defect: 36.7% of the whole corpus's inverted-range findings. The same shift supplied all 352 of that file's unit-mismatch findings, reporting drugs "priced in 2 different units: inpatient, outpatient". A row of the wrong width is now not read, and the report says how many were set aside and in which file. The corpus-wide counts fell by exactly 1,517 and 352.
• A unit was compared case-sensitively, and the item beside it was not. ML on one row and mL on another were reported as one drug priced in two units, after the two rows had been matched as one item across a description differing in case in exactly the same way. All 17 findings in one published file, Kula Hospital, were that and nothing else. The unit is now folded like the item, and the corpus-wide count fell by exactly 17.
• A percentage column holding a fraction was read as a percentage. The CMS schema means 85% by 85; some hospitals write 0.85. One file, Collingsworth General, writes fractions throughout, and 11,230 of its 11,240 stated-discount findings reconciled exactly under the other reading. The fix is not to detect which writing a file uses. That would be inferring from contents, which this tool does not do. It is to ask the question the number reader already asks of a thousands separator: does anything in the column settle it? A cell printing % does, and so does a cell above 1, which a fraction of a price is not. A column in which every value lies between 0 and 1 is settled by neither and is refused. Of the 1,623 columns given a percentage role across the corpus, that is 22 columns in 4 files.
• A JSON document named .csv was read as a very wide table. 27 of the 200 hospital files are the CMS schema's JSON form served from a URL ending .csv. No check could find a column it could use and the run exited 2, which was right, but nothing said the file was not a table, and the column totals for the whole corpus were 99.8% JSON punctuation. It is refused by name now.

Everything below was found by the earlier runs, in the same way.

• 0.091 was refused as an undecidable separator. The rule about 1,234 had never met a price below one unit. 1,898 real tariffs were refused because of it. A leading zero cannot be a thousands group, and now says so.
• A percentage was compared against the wrong price column. In the CMS schema, standard_charge|negotiated_percentage belongs with standard_charge|negotiated_dollar, and standard_charge|discounted_cash sits beside them sharing the same prefix. Pairing on the family alone chose the cash price and produced 363,426 findings across ten files, every one a comparison between two numbers never meant to agree. Columns are now paired by the longest run of shared leading words.
• Two rows with the same description and different codes were reported as one item priced twice. Holding only the first item column equal produced 75,968 findings across 22 files. Every column naming what is priced is now part of the identity, and code|1, code|2 are recognised as such.
• A free-text note was not treated as telling two rows apart. Hand-reading thirty two-prices findings found ten that were two rows identical but for a note reading Gross Charge Type: Sta against Gross Charge Type: Fee: a 33% false-positive rate, from one unrecognised column. Recognising notes as qualifiers took the count on the same files from 91,760 to 5,221, and a re-audit of the survivors found 22 of 22 correct.
• drug_unit_of_measurement holds an amount and drug_type_of_measurement holds the unit. The names read the other way round, and reading them that way had the tool reporting that a drug was "priced in 2 different units: 1357.2 and 8".
• A percentage was compared against a cash price. Where a file leaves its negotiated dollar column empty throughout, the nearest remaining price was the discounted cash price, and a gross charge times a negotiated percentage has no reason to equal one. A percentage now pairs only with a column that is its own name with the last word changed, or declines to run.
• unit-mismatch reported every multi-currency price list. It ignored the qualifier columns (region, plan, payer) that legitimately separate the rows. A price list quoting one product in GBP for the UK and EUR for the EU is a normal document.
• One unusable column aborted an entire check. A wide file can carry hundreds of payer blocks, and one empty percentage column stopped the other hundreds from being read at all.
• "could not be read" was printed for a column that simply held no number. Those are different facts about a document, and the report now says which.
• A malformed .xlsx raised a KeyError out of zipfile instead of the message saying what the tool reads.
• The volume-tier finding said "per unit" when the column it used was not a unit price, something the document does not state either way.
• The zero-band note missed exactly the case that needed it. It tested the rate plus its adjuster, so a band printed at 0 with an adjuster of -1.4 summed to -1.4 and carried no note. It now tests the printed rate.
• Reading a whole table into memory costs about four times the file size. The largest published file in the corpus, 1,083 MB, needs about 3.9 GB. Cell now uses __slots__, which was worth gigabytes on it; the underlying limit remains and is published rather than warned about.

What an adversarial review found that neither the tests nor the corpus did

The suite was green and both corpora had been measured when a review was commissioned specifically to break the tool's stated promises. It broke two.

• Exit 0 was reachable with no arithmetic done at all. Described above. It is the most important promise in the tool and it had a hole in it.
• The offline guard could be walked past using only allowlisted modules. The allowlist held the top-level package xml, and xml.sax.parse(url, handler) resolves a system id through urllib.request.urlopen: a working exfiltration path with no socket, no subprocess and no dynamic import anywhere in the source. Separately, zipfile does import os, so zipfile.os is the real os module, and getattr(getattr(zipfile, "o"+"s"), "sys"+"tem")(command) spawned a shell with the whole suite green. The allowlist is now dotted (xml.etree and nothing else under xml). Every module name an allowed module re-exports is refused as an attribute, and getattr is refused outright. Both proofs of concept are regression tests.

Four more, none of which the tests would have caught:

• A corrupt .xlsx exited 1, not 2. A zip verifies a part's CRC when it is unpacked, not when the archive is opened, and the failure arrives as zlib.error, which is neither an OSError nor a zipfile.BadZipFile.
• stated-discount reported a contradiction in a consistent file. With two columns both naming a list price, it bound to the nearer name and flagged a row that reconciled perfectly against the other one, and reordering the two columns made the finding vanish. Every candidate pairing is now tried, and a finding needs all of them to fail.
• two-prices asserted "no other column in this table tells the two rows apart" about a Promo column sitting right there. It now says no column this tool recognises does, and that one it does not recognise may.
• A sub-header stole the header row. Taking the last header-shaped row handles a title block above the header, which is what published files usually have; it fails on a units row or category banner below it. The candidate that names the most recognised columns now wins, which handles both.

The review also confirmed what it could not break: no output path passes judgement, no check touches a cost, a blank is never read as a zero, and no role anywhere is inferred from a column's contents.

What it misses

Stated plainly, because a checking tool that hides its blind spots is worse than none. The long version is docs/limitations.md.

• Nothing about margin, ever. A price list does not state a cost.

• Two of the seven checks, and half of a third, are unmeasured. bundle-above-parts and discount-tier have unit tests and no corpus: across the 173 hospital CSVs a column naming a bundle's contents appears zero times, and so does a quantity column, and nothing else found in this work publishes either in a machine-readable price list. The currency half of unit-mismatch is in the same position: no currency column appears in either corpus, and the transit fares collected specifically to supply one turned out never to carry two currencies in a feed. All three may be wrong in ways nobody has seen.

• A column name it does not know stops a check. Trade, Band C, Kundenpreis get no role. The report says which word it wanted; --map settles it. The vocabulary is a record of headers that exist, and it is biased towards the corpora this tool has met.

• A row of the wrong width is not read, and that has a cost of its own. A row with more cells than the header line is set aside, because an unquoted comma in a free-text field puts every value after it under the wrong heading; so is a row that does not reach every named column. The report says how many and in which file, and 190,549 rows across 19 of the 173 hospital CSVs were set aside on the last run. The cost falls on a file whose writer puts a trailing comma on its data rows but not its header: every row in it is one cell too wide, none is read, and the run exits 2 saying so. Loudly wrong is the intended failure here, but it is still a file this tool will not read.

• A percentage column that could be either writing is refused, and a real finding can go with it. One published file lost 91,188 stated-discount findings that the hand audit had judged correct, because the percentage column they came from holds 0.8 throughout and nothing in the document says whether that is 80% or 0.8%.

• A qualifier it does not know produces wrong findings, and not a few of them. This one is still open. If a column that legitimately separates two rows has an unrecognised name, two-prices and unit-mismatch treat those rows as the same thing. A published Czech transit fare list shows 2,396 apparent violations when its rows are grouped by tariff and distance alone, and zero once validity_days is part of the group. The failure mode is thousands of findings rather than one. It was expected to be the commonest cause of a wrong finding; on the hospital corpus it has now caused none in two hand audits of 66 findings each.

• It cannot tell a unit price from a line total. Where no column names a unit price it uses the net price and says so in the finding.

• It holds the whole table in memory. The largest published file in the corpus (1,083 MB of CSV, 93,954 rows of 285 columns) needs about 3.9 GB and 44 seconds on its own, on the machine these numbers were taken on (macOS, CPython 3.14). That is roughly four times the file size. Across the whole 200-file run, peak resident memory reached about 8.9 GB, and 6.9 to 8.5 GB on the three earlier runs. There is no streaming mode and no cap: a file bigger than your memory will not be read, it will swap.

  That figure is given to one decimal on purpose. Two records of this same run disagreed in the second decimal, 8.93 GB against 8.97 GB, and the raw value was not kept: tools/measure_hospital.py computes peak_rss_mb and prints it with the rest of the run's JSON, and that output was read once and never committed. One decimal is the precision the surviving records agree on. Anything finer than that would be a transcription, not a measurement.

• One file at a time. It cannot compare this quarter's price list with last quarter's, or with the contract, or with the invoice.

• It reads values, not formulas. A number produced by a broken formula is checked as the number it now shows.

• No PDF.

Reproducibility

Every corpus-wide number above comes out of a script in tools/, against a corpus whose hashes are published, and the two exceptions are named below. Re-run them:

curl -sO https://apps.openei.org/USURDB/download/usurdb.csv.gz
shasum -a 256 usurdb.csv.gz     # 9005c5358de7e2471cef...d28d5
python3 tools/measure_corpus.py usurdb.csv.gz --samples tariff-audit.json

python3 tools/fetch_hospital.py corpus/ --count 200 --cap-mb 40
python3 tools/measure_hospital.py corpus/ --samples hospital-audit.json \
    --per-document concentration.json
python3 tools/audit_hospital.py corpus/ hospital-audit.json

python3 tools/probe_gtfs.py --feeds 140

Two exceptions, both flagged rather than glossed. The survey of open data portals (3,802 URLs, 43 price-list titles, a 1.7% true-positive rate) was a one-off reconnaissance across six portals whose APIs differ and whose result sets are capped by relevance. No script reproduces it; the API forms are published so it can be repeated by hand.

The second is the per-file arithmetic in the hospital audit (Collingsworth's 11,230, Sky Lakes' 1,517 inversions and 352 unit findings, Kula's 17, Nazareth's ratio spread, Vibra's 120,698), which was done by hand while reading each finding back against its file, not by a script in tools/. Each is one pass over one named file and each is stated as the operation that produced it: for Sky Lakes, count the rows whose cell count differs from the header line's and check which of them carry a minimum above their maximum; for Collingsworth, divide the negotiated dollar by the gross charge and compare it with the percentage cell as written and as divided by a hundred. They are counts over a single document rather than measurements of a corpus, which is why they are named here instead of being given a flag.

Sky Lakes' 180,084 ragged rows are no longer in that category. The tool counts them itself now, prints the count, and measure_hospital.py totals it across the corpus, so that figure comes out of the run like every other. The same is true of the 22 refused percentage columns.

Neither corpus is redistributed here. The tariff export is 195 MB and public domain; the hospital files are republished constantly and belong to the hospitals that publish them. What is published is the measurements, the selection rules, and (in docs/hospital-corpus.csv) the SHA-256 of every one of the 200 hospital files, beside its certification number, facility name and URL. That file is a bibliography: no price, no charge, no row of anyone's chargemaster, no personal data.

A corpus that was collected and then rejected is in the manifest too: 1,174,347 published transit fares across 140 GTFS feeds, in which fare_id is a primary key and no feed uses more than one currency, so the two checks it was collected for cannot fire on it even in principle. Twelve currencies appear across those feeds, and never two in one.

Licence

MIT. See LICENSE.

Nothing in this repository is pricing, commercial, legal or financial advice.