The archive test · method

How the archive figure was counted — and what you can and cannot check for yourself.

You can inspect this count. You cannot re-run it — and this page will not pretend otherwise.

01The figure

145 945of 146 132

values sit in a row whose hand-written total the parser reproduces exactly. The number is the tool's own output — nothing about it is hand-derived, including the breakdown below.

All 146 132 values, accounted for

in a row whose written total matches145 945

in the two rows whose total disagrees64

in three rows the director never totalled123

values parsed, from 34 monthly workbooks146 132

146 132 values parsed from 146 136 non-empty cells. The four cells that could not be read are not in this count — an unreadable cell never parsed, so it was never a member to subtract. They are named below.

02The method

What the parser had to learn.

“We wrote a script” is not a method. Each rule below exists because a number refused to reconcile until the rule was found — the work was driven by its failures, which is why the failures get their own band.

Three layouts, sixteen years

Dates as text, dates as Excel serials, one column per day — then a collector header, with several stacked down one sheet. The layout is detected per file, never configured per era.

The shift split is per file — and once, per mid-month

Not even monthly: one month runs one column per day to the 24th and two from the 25th. That is the evening shift starting mid-month, visible only in the columns themselves.

Four notations for two shifts

Including one pair reversed, and one pair nobody can now explain. All four resolve to the same two shifts.

The totals triple moves

Read positionally, a farmer's 237.5 litres becomes a price of 237.5. So the triple is never read positionally — it is located by arithmetic: the rightmost triple where litres × price = sum.

The date convention slips

Two cells written back-to-front land two years away from the records around them.

03The exceptions

Everything that did not reconcile, named.

Named, not counted. The figure above is credible only if this list is complete — so this is all of it.

July 2010 · row short by 8.0 L8.0 Lrecovered, unread

This row holds both unreadable cells. The director's own hand-written total therefore tells us what those two cells said between them — 8.0 litres — without anyone being able to read either one. The method recovers a value it could not see.

July 2010 · row short by 33.5 L33.5 Lthe one slip

A different farmer, and no unreadable cell to explain it. The single arithmetic slip in sixteen years — named here rather than hidden.

The four cells that could not be read

2010-0713 · morning_

2010-0726 · morning_

2020-103 · morningო

2026-0313 · day?

the two July 2010 underscores are the pair recovered above · glyphs quoted as written · “day” means a month written with one column per day, not a third shift

absence — not a disagreement

Three rows carry no written total at all: 123 values with nothing to reconcile against. Those values parsed and are counted above — nothing about them failed. Counting an absence as a failure would be as dishonest as hiding it.

04Check this

What you can check, today.

The source is one company's sixteen years of finances, with about five hundred real farmers' names in it. It is private and it stays private. So the property on offer here is inspectability, not reproducibility: the working in full — every rule the parser was forced to learn, and every value that refused to reconcile, named.

The tool's own repository is private and stays private. So the method travels here instead: the working in full above, and below, the SHA-256 digest of all 34 source files. If you hold this archive, hash your copy and compare — that check needs nothing from us, and no repository at all.

The partition adds up.

145 945 + 64 + 123 = 146 132, and 146 132 + 4 unreadable = 146 136 non-empty cells. Redo it now.

The exceptions are enumerable.

Two disagreeing rows, four unreadable cells, three untotalled rows — each one named on this page. Nothing is summarised into a count you cannot unpack.

The figures are unrounded, here and on the homepage.

The pair of counts is the claim. No rounded or percentage form of it appears anywhere we publish.

If you hold this archive, confirm we read the same thing.

34 source files, one SHA-256 digest each, listed in full below — unshortened, because a shortened digest cannot be compared with anything. Hash your own copy and compare.

Source files · SHA-256

2010-07: 2e253b21f5a97fb7daadcd3fd588c646ec6a2f0f637e377a58249fb3f392d0c0
2010-08: 0eb9e6a0305775cc657b058973b736e968c776cbcc0a8daf6b9940751b8cee81
2010-09: a2505902d91fd496c1d60396e052e018599e6eba25ebd8ecde958d34ee9837de
2010-10: bd3a2c840ac96a2a74bb39041c921753cab229f23bfad9f9ce0101f957f1bcf6
2016-01: ea800c19193fa83bd064af8c7e3d03636a8f48728308b1336df391a3d74fdb2e
2016-02: 2c8bda80f7efa571bb96f23a24806e94ec2c94c593576b129b01b59d0c69ca28
2016-03: 6ee71203e211a7a154980d702c33291acee182a53e3c322641e1d09deda28a2e
2016-04: e4f12ec50eb80eee9cd20075c33e604a38c2ca094c97f90dfcfb0f4bbdba638d
2016-05: 3e2251b0399a1ba94f84b40cc3e937ff621c6a31d3a0127dd6a2a922c79f473e
2016-06: 99eb60bf2b20cd6e26885a2c08786f56ae54b891abafbf6b1d7885cda41e9510
2016-07: 69a27df800d5f46ec9b8dc3430dde7ee8d1c130bafc1a252f103625b1a3ed926
2016-08: 1f8bc3d420ef62f8a83b14158166f7df5f13070cae2ff1af758c76f24e1345c6
2016-09: b3641e4a2f09be76f15e1ee641bc6011f6a9beab1fb16465cfcd7768914cab5d
2016-10: fb4c39d94f07b688185bceb77465a0afe95ac463e1bb3de35634bfc7ce5b681c
2016-11: c7cad9176550e3ac93379bc26bbbf8fbe3dd426dd9778b7e70f08ae880b88e0b
2016-12: 7abf58728a74fbf7cabdc9384848a31acca7e950372a3c438069f3442236d836
2020-01: 82d39ef13f873838de27ec1e5e608eabe611d801a0f14cb73ec4014225b28244
2020-02: b97f7b9aa321af870ddf66dc4211a282cb8846e3e1696a19427caac822919b94
2020-03: a2c0f9fed69e0f65bf2c2bc6c2b3d90fc6dd50a34c7fbe0fea86e80da8a2887b
2020-04: 2562ca2a4f5993eb5745a158baf1705fb40689aee91ed8fd48b419a36236d6f0
2020-05: 48839222fe084b5bddd13946be932ce8171b4c89f4d7c51608e16db281f49048
2020-06: 574a48e20d6398f0f782a71c36517d1cfccdcfece391c8cfacc082f7055ef43b
2020-07: fc6ccf51d50a17e2e3b61a312bb2a7ff05524169779e73f34e1d9c13d6b0fb1b
2020-08: 7c1c89e1cd5b5aa626f8a35508231ac642c4b07b1be78b10fe09519641e4175c
2020-09: 04ba25d673d10c109e70d93ab700b54533f300c1a7316e809cab4ecee54d5171
2020-10: abc67fea3b5f3b26dc8acc9562d27cd432e9700e3bd42ec3eab7fe828b95d570
2020-11: c1eb9d43312de9b5790972dbb675faa33bfa07c2269bb12b17bf8cb4af90d7af
2020-12: 8c038249eac75a4155aae3084eb233943cacb574023449db2d8e4815130ace3c
2026-01: e72298ca37ed67e7898d3f47ff7715c9759dba3293d2eabab1721628dcd97b46
2026-02: 57f7f074b167ccf3854e1fc2c6fb849f2c4ec8a948beb410e3341651b4278766
2026-03: 6217da346cbd4ea2c1154f55b1931a1527d4efdb39bcda0f0fb078bb33072d60
2026-04: ef9c44f354a4e46dfede6bdbacc306960a76ee9409f50516815cadde959303b4
2026-05: fe575feac7977ca186d8cd73bf2169270f7572b82917cf9048d9ac94bfa48d54
2026-06: fe8505d3b17b21ed700c5b6fabd1ed7119fc00e4b27ad7327a61cabdeb98b833

34 files, one line each, keyed by month. The whole set — nothing is elided.

The toolprivate repository — the method is published here instead

Back to the homepage