Guidelines
What this edition contains, how it was produced, how accurately it was measured — and where the measurement itself reaches its limits. All figures on this page are computed from the data, not maintained by hand.
What this edition is and what it is not
This digital edition reproduces Hegel’s text as the Gesammelte Werke (GW) present it, and makes it free to read, searchable, and citable by volume, page, and line of the GW. The text-critical work – comparing manuscripts and prints, establishing the text and giving reasons for it – was done by the editors of the GW. This edition adopts their result and checks that it is reproduced exactly.
It is therefore not a critical edition in its own right, and it does not replace the GW. It does not contain the editors’ apparatus, variant readings, editorial reports, and notes; the apparatus remains protected by copyright and is neither transcribed nor stored. Anyone working with the text in a scholarly way cites the GW and consults their apparatus; the references given here lead exactly there (How to cite).
Only GW volumes whose protection as a scholarly edition under § 70 UrhG has expired are included: 25 years after publication, counted to the end of the year (Copyright).
This edition is prepared by Giuliano Infantino, independently and without affiliation with the publisher, an academy, or the Hegel-Archiv. The text is transcribed by machine and measured on samples; a double-blind check by human readers is still outstanding (section 7).
1. Source
The basis is the text of the Gesammelte Werke (GW), the authoritative historical-critical edition. It has been published since 1968 by Felix Meiner in Hamburg, edited by the North Rhine-Westphalian Academy of Sciences, Humanities and the Arts in association with the German Research Foundation (DFG), and prepared at the Hegel-Archiv of the Ruhr-Universität Bochum. Only Hegel’s own text in its original spelling is transcribed — not the editors’ apparatus.
2. The holdings in figures
Sixteen volumes, 5,976 pages, 1,940,118 words. In addition, 1,614 footnotes by Hegel and 3,252 headings.
| Volume | Pages | Words | per page | letter-spaced | Density | Fn. | Hdg. |
|---|---|---|---|---|---|---|---|
| GW 1 | 397 | 133,514 | 336 | 746 | 0.56% | 242 | 59 |
| GW 3 | 232 | 38,704 | 167 | 365 | 0.94% | 42 | 38 |
| GW 4 | 499 | 209,320 | 419 | 5,275 | 2.52% | 26 | 78 |
| GW 5 | 475 | 134,253 | 283 | 1,568 | 1.17% | 40 | 91 |
| GW 6 | 329 | 67,307 | 205 | 4,278 | 6.36% | 95 | 68 |
| GW 7 | 347 | 128,495 | 370 | 1,200 | 0.93% | 63 | 58 |
| GW 8 | 305 | 89,558 | 294 | 6,283 | 7.02% | 686 | 72 |
| GW 9 | 434 | 180,162 | 415 | 11,284 | 6.26% | 1 | 100 |
| GW 11 | 391 | 155,701 | 398 | 9,166 | 5.89% | 5 | 210 |
| GW 12 | 301 | 114,306 | 380 | 8,148 | 7.13% | 32 | 180 |
| GW 15 | 266 | 85,823 | 323 | 2,363 | 2.75% | 7 | 142 |
| GW 17 | 328 | 89,200 | 272 | 6,826 | 7.65% | 209 | 98 |
| GW 18 | 325 | 98,783 | 304 | 3,323 | 3.36% | 89 | 63 |
| GW 19 | 419 | 135,435 | 323 | 9,105 | 6.72% | 29 | 887 |
| GW 20 | 556 | 149,882 | 270 | 9,708 | 6.48% | 31 | 889 |
| GW 21 | 372 | 129,675 | 349 | 6,671 | 5.14% | 17 | 219 |
| Total | 5,976 | 1,940,118 | 325 | 86,309 | 4.45% | 1,614 | 3,252 |
On 4,773 of the 5,976 pages (80%), the print contains an editors’ apparatus. It is not transcribed; the fact that it is there is recorded in the data.
3. Letter-spaced type — and what its distribution shows
Hegel emphasizes by letter-spacing, not by italics. Across the whole holdings there are 86,309 letter-spaced passages, each recorded individually. On average this affects 4.45% of all words.
The mean hides what is interesting. The density varies between 0.56% (GW 1) and 7.65% (GW 17) — by a factor of 13.7.
This is not an artifact of transcription but a finding. The lower group consists of drafts and posthumous material, the upper of the fully worked-out prints. Whoever revises a text several times decides more often which word carries the weight — and sets it letter-spaced. The density of letter-spaced type is thus a rough measure of the degree of elaboration.
4. Transcription procedure
The text is transcribed page by page from the scan of the printed page, guided by fixed rules for spelling, emphasis, and structure. There is no separate intermediate OCR step. Misprints in the source are retained and not corrected.
5. How it is measured — the calculation in detail
5.1 The model
From each volume, pages are drawn at random, with a fixed seed, and read line by line against the print. Each word is thus a trial with two outcomes — correct or wrong —, and the number of errors k among n words follows a binomial distribution with unknown error rate p:
P(k | n, p) = C(n, k) · pk · (1−p)n−k
What we want is not k but p — and not as a number but as a range. The assumption behind this is the independence of the trials. It holds here only approximately; section 5.6 works out what that costs.
5.2 Why the point estimate alone says nothing
From 2 discrepancies in 2,990 words (GW 5) follows the point estimate
p̂ = k / n = 2 / 2,990 = 0.000669 → 99.933% correct
But with the same text quality there could just as well have been one or five. The point estimate conceals how far off it may be. It is an observation, not a statement about the volume.
5.3 Why the normal approximation fails here
The usual Wald interval is
p̂ ± z · √( p̂ (1−p̂) / n )
and it presupposes that n·p and n·(1−p) are both sufficiently large — as a rule of thumb at least five. With error rates around 10−4 this is never satisfied. Two consequences show it immediately:
- At k = 0 the square root becomes zero and the interval shrinks to the point [0, 0] — the calculation would claim 100% without any uncertainty, from zero observations.
- For small k the upper bound extends beyond 100%, that is, beyond the range in which a probability can lie at all.
5.4 The exact interval (Clopper-Pearson)
Clopper-Pearson turns the question around: we look for the two values of p at which what was observed is just barely not surprising. The lower bound is the smallest p under which k or more errors still occur with probability α/2; the upper bound is the largest p under which k or fewer still occur with probability α/2. Through the relation between the binomial and the beta distribution this can be written in closed form:
plower = B(α/2; k, n−k+1) pupper = B(1−α/2; k+1, n−k)
where B is the quantile of the beta distribution. Worked through for k = 2, n = 2,990, α = 0.05:
p ∈ [0.000081; 0.002414] → 99.759% to 99.992%
The interval is asymmetric, and that is as it should be: there is more room below than above. It is also deliberately conservative — because k is an integer, the actual coverage lies somewhat above the promised 95%. For a statement of quality this is the right direction: better too cautious than too favorable.
5.5 The case without errors — the rule of three
Several volumes show zero errors in the sample. That does not mean “error-free,” only that the error rate is small enough for this sample to have missed it. How small follows from a single line. The probability of seeing no error in n trials at rate p is (1−p)n. It should be at most α:
(1−p)n ≤ α ⟺ p ≥ 1 − α1/n ≈ 3/n (for α = 0.05)
The approximation comes from ln(0.05) ≈ −3. For GW 19 with n = 12,774 words this means: an error rate of up to 3/12,774 = 0.000235 — that is, up to three errors in ten thousand words — is still compatible with a zero finding. At 85,755 characters the same bound drops to 3.5 in a hundred thousand.
The difference from the table above is intentional: there the two-sided Clopper-Pearson interval is given (α/2 in the upper tail), here the one-sided bound. Both are correct; they answer different questions.
5.6 What the word count conceals: clustering
Here lies the greatest weakness of all the figures on this page, and it is rarely stated in reports of this kind. Pages are drawn, words are counted. The binomial model from 5.1 assumes that each word was drawn independently — in fact the words come in clusters of about three hundred each, and errors accumulate within a cluster: whoever misreads one line often misreads the next as well, and whoever misunderstands a typesetting convention applies it wrongly across the whole page.
The finding for GW 19 shows this immediately: of four errors in letter-spaced type, three are on the same page. An interval that treats them as four independent events comes out too narrow.
What can be estimated cleanly from a sample of pages is therefore, first of all, a quantity at the page level. In GW 19, 3 of the 40 sampled pages have at least one finding:
3 / 40 → share of error-free pages 92.5%, 95% interval 79.6% to 98.4%
Across all volumes together, 12 of 316 sampled pages have at least one finding. Calculated at the page level, that is 96.2% error-free pages with a 95% interval of 93.5 to 98.0% — an entirely different order of magnitude of uncertainty than the four decimal places of the character table.
This is considerably wider than the word-level interval — and it is the quantity that the sample really measures. The word-level rates of sections 6 should therefore be read as a lower bound on the uncertainty: the true confidence interval is at least this wide, more likely wider. To make it narrow, one does not need to read more words per page, but more pages.
5.7 A flag is not an error
The machine measurement of letter-spaced type (section 6) reports flags. How many of them are errors is a conditional probability:
P(error | flag) = P(flag | error) · P(error) / [ P(flag | error) · P(error) + P(flag | correct) · P(correct) ]
The denominator is dominated by the second term: P(error) is small, P(correct) close to one. Even a false-alarm rate of a few percent then produces more false than genuine flags. That is exactly what is observed — in GW 16, on recounting, about nine out of ten flags were false alarms, and in GW 19 the machine stands at 94.5%, the judgment decided against the page image at 99.5%.
What follows from this is not a correction of the figures but a rule for reading them: the machine rate is a lower bound on quality, never an estimate of it.
5.8 What does not count as an error — and why
Three phenomena come up regularly in checking and are not counted in the error rate. They are listed here because a number without its boundaries is worth nothing.
The page-break bar. The character |
marks the page break of the first printing. The more recent volumes record it; the older ones — GW 4 to 9 — do not record it in a single place. That is a volume-wide decision about the scope of the transcription, not a misreading. If it were counted, the older volumes would systematically look worse without a single word being wrong.
Marks on small capitals and full capitals. By the convention of this edition, small capitals carry no letter-spacing mark. The print, however, sometimes really does letter-space such lines further — the mark then renders it correctly and at the same time breaks the house rule. Found in GW 1, 3, 6, 7, and 19. As long as the convention is undecided, a number that counts such cases as errors measures not the care of the transcription but an open rule.
Misprints in the source. If the print has erfoderte or tranquilllité with three l’s, then exactly that is correct in the edition’s text. This edition does not emend.
5.9 What became of the findings
On 27 August 2026 the confirmed individual findings were corrected — not as a batch correction, but each passage separately, with proof that the replacement applies exactly once:
- GW 1, p. 159: Gebäude → Gebaüde.
GW 1, p. 305: en dash → em dash in the Nathan quotation. - GW 4, p. 225: mark removed from unter — the word is set tight, its neighbors letter-spaced.
- GW 8, p. 275: missing period in Hegel’s marginal note.
- GW 9, p. 284: mark added for in.
- GW 18, p. 42: wrong hyphenation point — as a result the running text read einfachchen instead of einfachen.
- GW 19, p. 258: mark added for Fits.
GW 19, p. 423: z.B. → z. B. - GW 20, p. 288: mark removed from zu-.
- GW 12: 220 em dashes on 27 pages (see above).
On 6 September 2026 the 235 letter-spacing findings of the reassessment followed (section 5.10): 130 marks added, 74 removed, plus 31 from the recall sample — each passage addressed by its printed line, with proof that the replacement applies exactly once, and checked against the mirror check of pruefe_gemeinsam.py afterwards. Most concern short words in letter-spaced word groups (zur, des, und), most often in GW 20, GW 17, GW 9, and GW 8.
One misstep belongs here too: on the first attempt, the correction in GW 4 hit all thirteen occurrences of unter in the volume instead of the one on page 225. It was reverted and repeated with the exact page; twelve marks stand unchanged. A correction without a location is no correction.
5.10 Why two volumes cannot be compared
Two overlapping intervals do not mean “no difference,” and two separate ones do not automatically mean “difference established.” Anyone who wants to compare two volumes must test the difference itself — with sample sizes like these, using Fisher’s exact test, not by eyeballing two bars.
For the machine rates, comparison is ruled out anyway, and for a substantive reason: how sharply the measurement discriminates depends on the typesetting of the volume — on typeface, normal spacing, and line density. It then measures the volume, not the transcription. How costly this fallacy is, the section “A withdrawn number” shows.
6. Measured accuracy
Word accuracy
| Volume | Pages | Words | Discrepancies | Accuracy | 95% interval |
|---|---|---|---|---|---|
| GW 1 | 20 | 7,194 | 2 | 99.972% | 99.900 – 99.997 |
| GW 3 | 20 | 2,699 | 0 | 100.000% | 99.863 – 100.000 |
| GW 4 | 20 | 8,725 | 0 | 100.000% | 99.958 – 100.000 |
| GW 5 | 20 | 5,161 | 0 | 100.000% | 99.929 – 100.000 |
| GW 6 | 20 | 3,620 | 0 | 100.000% | 99.898 – 100.000 |
| GW 7 | 20 | 7,022 | 0 | 100.000% | 99.947 – 100.000 |
| GW 8 | 20 | 5,665 | 1 | 99.982% | 99.902 – 100.000 |
| GW 9 | 20 | 8,483 | 0 | 100.000% | 99.957 – 100.000 |
| GW 11 | 20 | 7,922 | 0 | 100.000% | 99.953 – 100.000 |
| GW 12 | 20 | 7,210 | 15 | 99.792% | 99.657 – 99.884 |
| GW 15 | 20 | 7,166 | 0 | 100.000% | 99.949 – 100.000 |
| GW 17 | 25 | 6,585 | 0 | 100.000% | 99.944 – 100.000 |
| GW 18 | 20 | 5,967 | 1 | 99.983% | 99.907 – 100.000 |
| GW 19 | 40 | 12,774 | 2 | 99.984% | 99.943 – 99.998 |
| GW 20 | 20 | 5,288 | 0 | 100.000% | 99.930 – 100.000 |
| GW 21 | 16 | 5,646 | 0 | 100.000% | 99.935 – 100.000 |
Since 27 August 2026, every one of the 16 volumes has been measured. 386 randomly drawn pages were read, each one in full, line by line, against the print.
Character accuracy
For “practically error-free” text, the DFG Practical Guidelines set a threshold of 99.97% — measured in characters, not words. A word in the source has a little over four characters on average; a word error therefore does not count as a single character error.
Since 27 August 2026, the character rate has been counted for all eighteen volumes, no longer converted from the word rate. Anyone comparing line by line against the print anyway can count characters at no extra cost. The denominator is the character count of the drawn pages, recalculated from the transcription files. 815,203 characters were read on 386 randomly drawn pages.
| Volume | Characters | Discrepancies | Character accuracy | 95% interval | Source |
|---|---|---|---|---|---|
| GW 1 | 46,253 | 3 | 99.9935% | 99.9810 – 99.9987 | counted |
| GW 3 | 18,081 | 0 | 100.0000% | 99.9796 – 100.0000 | counted |
| GW 4 | 58,229 | 0 | 100.0000% | 99.9937 – 100.0000 | counted |
| GW 5 | 34,789 | 0 | 100.0000% | 99.9894 – 100.0000 | counted |
| GW 6 | 24,279 | 0 | 100.0000% | 99.9848 – 100.0000 | counted |
| GW 7 | 45,994 | 0 | 100.0000% | 99.9920 – 100.0000 | counted |
| GW 8 | 37,069 | 1 | 99.9973% | 99.9850 – 99.9999 | counted |
| GW 9 | 54,819 | 0 | 100.0000% | 99.9933 – 100.0000 | counted |
| GW 11 | 52,408 | 0 | 100.0000% | 99.9930 – 100.0000 | counted |
| GW 12 | 48,853 | 21 | 99.9570% | 99.9343 – 99.9734 | counted |
| GW 15 | 48,967 | 0 | 100.0000% | 99.9925 – 100.0000 | counted |
| GW 17 | 41,567 | 0 | 100.0000% | 99.9911 – 100.0000 | counted |
| GW 18 | 39,638 | 2 | 99.9950% | 99.9818 – 99.9994 | counted |
| GW 19 | 85,755 | 1 | 99.9988% | 99.9935 – 100.0000 | counted |
| GW 20 | 35,696 | 0 | 100.0000% | 99.9897 – 100.0000 | counted |
| GW 21 | 38,161 | 0 | 100.0000% | 99.9903 – 100.0000 | counted |
Sixteen of the eighteen volumes lie above the DFG threshold of 99.97% with their lower interval bound. Two do not:
- GW 12 stands at 99.957%, and its interval does not reach the threshold. The cause is a single one: the print sets Hegel’s syllogistic formulas (E—A—B) with em dashes, while the edition’s text had en dashes. Re-measured on the character boxes of pages 98, 103, 108, 309, and 312: the formula dash measures 35 to 40 px, the dash in running text on the same pages 18 to 22 px — twice as much.
Fixed on 27 August 2026, volume-wide: 220 dashes on 27 pages, after which no en dash between capitals remains in the volume. The number in the table nonetheless stays. It describes the state at which the measurement was taken, and a guideline that retroactively improves its own measurements is no guideline. The volume today is better than the number — after the correction, no finding remains on the drawn pages. - GW 1 and a few others just fail to clear the threshold with their lower bound — there it is not a matter of text quality but of sample size. Anyone who wants to establish the threshold as a lower bound as well has to read more pages; see section 5.6.
Letter-spacing
The DFG sets no threshold for emphasis. Emphasis can, however, be checked geometrically: a script measures the letter spacing of every marked word in the scan and compares it with its neighbors in the same line.
This measurement runs over the whole volume, but it reports flags, not errors. A large share of them are false alarms: tight letter pairs such as ei or kt push the average of long words below the page threshold, and the report “too much” then hits a word that is indeed letter-spaced in the print. In GW 16, on recount, about nine in ten reports were such false alarms. The “correct” column is therefore a lower bound of quality, not the quality itself — the true value lies above it, not below.
| Volume | checked | in the volume | flagged | at least correct | 95% interval |
|---|---|---|---|---|---|
| GW 1 | 478 | 746 | 17 | 96.444% | 94.367 – 97.915 |
| GW 3 | 318 | 365 | 21 | 93.396% | 90.082 – 95.866 |
| GW 4 | 4,503 | 5,275 | 445 | 90.118% | 89.209 – 90.974 |
| GW 5 | 1,240 | 1,568 | 69 | 94.435% | 93.010 – 95.645 |
| GW 6 | 3,964 | 4,278 | 111 | 97.200% | 96.638 – 97.691 |
| GW 7 | 957 | 1,200 | 64 | 93.312% | 91.540 – 94.812 |
| GW 8 | 4,996 | 6,283 | 410 | 91.793% | 90.998 – 92.540 |
| GW 9 | 10,752 | 11,284 | 332 | 96.912% | 96.568 – 97.231 |
| GW 11 | 8,533 | 9,166 | 715 | 91.621% | 91.013 – 92.200 |
| GW 12 | 7,949 | 8,148 | 253 | 96.817% | 96.407 – 97.192 |
| GW 15 | 3,930 | 2,363 | 203 | 94.835% | 94.096 – 95.506 |
| GW 17 | 9,481 | 6,826 | 201 | 97.880% | 97.570 – 98.160 |
| GW 18 | 4,514 | 3,323 | 112 | 97.519% | 97.022 – 97.953 |
| GW 19 | 8,917 | 9,105 | 490 | 94.505% | 94.012 – 94.969 |
| GW 20 | 9,253 | 9,708 | 536 | 94.207% | 93.712 – 94.675 |
| GW 21 | 2,302 | 6,671 | 143 | 93.788% | 92.723 – 94.740 |
Decided against the page image
What the lower bound is worth is shown by its counterpart: a recall sample of 6 September 2026. For each volume, ten pages were drawn at random (seed 20260906) and read in full for letter-spacing by a checker who did not see the edition’s text; every discrepancy between his list and the marks in the edition’s text was then decided individually against a zoomed crop. Unlike the tool’s measurement, this sample also finds what the tool does not report at all. The reference quantity is the number of marked words on the drawn pages.
| Volume | Pages | marked words | confirmed errors | correct | 95% interval |
|---|---|---|---|---|---|
| GW 1 | 10 | 11 | 0 | 100.000% | 71.509 – 100.000 |
| GW 3 | 10 | 20 | 1 | 95.000% | 75.127 – 99.873 |
| GW 4 | 10 | 188 | 1 | 99.468% | 97.072 – 99.987 |
| GW 5 | 10 | 79 | 0 | 100.000% | 95.438 – 100.000 |
| GW 6 | 10 | 118 | 3 | 97.458% | 92.750 – 99.473 |
| GW 7 | 10 | 72 | 1 | 98.611% | 92.503 – 99.965 |
| GW 8 | 10 | 183 | 3 | 98.361% | 95.284 – 99.661 |
| GW 9 | 10 | 255 | 0 | 100.000% | 98.564 – 100.000 |
| GW 11 | 10 | 323 | 1 | 99.690% | 98.287 – 99.992 |
| GW 12 | 10 | 286 | 3 | 98.951% | 96.965 – 99.783 |
| GW 15 | 10 | 140 | 1 | 99.286% | 96.085 – 99.982 |
| GW 17 | 328 | 10,022 | 66 | 99.341% | 99.163 – 99.490 |
| GW 18 | 10 | 154 | 0 | 100.000% | 97.633 – 100.000 |
| GW 19 | 10 | 222 | 0 | 100.000% | 98.352 – 100.000 |
| GW 20 | 10 | 167 | 1 | 99.401% | 96.709 – 99.985 |
| GW 21 | 10 | 214 | 2 | 99.065% | 96.665 – 99.887 |
The gap is considerable. GW 15 stands at 94.8% by machine and at 99.3% in the recall sample; GW 19 at 94.5% by machine and at 100.0% in the sample. Both numbers are correct — they measure different things.
The reassessment of 6 September 2026
Until 6 September 2026, the numbers in this table came from a smaller sample of 27 August 2026, which only decided the tool’s reports on the drawn pages against the page image. It could therefore say nothing about errors the tool overlooks. The reassessment did both: each of the tool’s 4,159 reports (tiers A and B across all eighteen volumes) was decided individually against a zoomed crop, and in addition came the recall sample described above, on 180 pages with 3,106 marked words.
The result of the full check: of 4,159 reports, 2,756 were false alarms, 187 concerned headings (which by convention carry no marks), 41 remained undecidable even at fivefold magnification, and 204 were errors in the edition’s text — 130 missing and 74 superfluous marks. On its 180 pages, the recall sample found 31 further errors that the tool had not reported. All 235 have been corrected to the exact page, in the reading text as well as in the line-by-line version. About 900 reports from GW 15 to 18 turned out to be reading errors of the tool itself: it had not recognized as marks those that extend over several words or a line break. This error was also present in this edition’s checking tool and has been fixed.
Not corrected were 24 letter-spaced words in foreign-language quotations that the GW set in semibold (GW 16 and 20): by convention the edition’s text leaves these passages entirely without marks, and marking a single word in them would be less consistent than marking none. Six further findings of the checker did not survive a second look and were left as they were.
GW 17 read in full
The reassessment showed where the remaining errors are: on the drawn pages, the machine measurement found only about one in ten of them. Further measurement runs therefore achieve little; only complete reading finds the rest. On 07/8 September 2026, GW 17 was therefore read in full — all 318 text pages, word for word, by checkers who did not know the edition’s text. The volume had the highest estimated residual error density of all eighteen.
9,912 named words against 9,688 marked ones yielded 947 discrepancies; each was decided individually against a zoomed crop. 810 were false alarms, 7 remained undecidable, 54 were errors in the edition’s text. Together with the findings of the targeted check, 96 marks have been corrected in GW 17. For this volume, the number in the table above is therefore not a sample but a full count: 66 errors in 10,022 marked words.
A by-product: the two rounds of judgment contradicted each other on eleven words. There the second applies, because its checkers measured the letter spacing instead of estimating it — with words like und, an, or ist, the eye errs even at fivefold magnification.
How the gap comes about was already shown in detail by the sample for GW 19 of 27 August 2026. On the drawn pages, the machine measurement reported marks as superfluous one after another that are clearly letter-spaced in the zoomed crop: words with 4 to 6 px internal spacing against a page threshold of 6 px, plus a shift in which Kants is letter-spaced and metaphysischen set tight. One reason could be named: the threshold is derived from the type size of the running text and does not fit the smaller type of the footnotes. Of the volume’s four confirmed findings, three are the same open question — marks on small capitals, which by convention carry none.
A limitation one needs to know: the confirmed cases were corrected afterwards. The drawn pages are therefore cleaner today than the sample shows them. The rate thus estimates not the present state of the checked pages but that of the unchecked remainder of the volume — and that is exactly what it is meant for.
A withdrawn number
Until 27 August 2026, this place held the proudest sentence on this page: GW 12 had been measured at 98.9%, GW 21 at 99.96%, the intervals did not overlap — and from this it followed that checking during transcription is more effective than checking afterwards.
The sentence is withdrawn. When the measurement first ran over all eighteen volumes on 27 August 2026, the same two volumes gave different values: GW 12 253 flags in 7,949 instead of 92 in 8,161, GW 21 143 in 2,302 instead of 2 in 5,400. The reason is not an error in the old numbers but their age: on 26.08., the measuring tool gained the test for partially letter-spaced words, the test for contiguous letter-spaced runs, the detection of words merged in the image, and a fix for a cropping error. Each of these finds additional cases. Comparing two numbers from different versions of the tool measures the tool, not the text.
Measured with one version, the order even reverses: GW 12 stands at 96.8%, GW 21 at 93.8%. That, too, proves nothing about the transcription, because how sharply the measurement discriminates depends on the typesetting of the volume — on typeface, normal spacing, and line density. The numbers in this table are not comparable between volumes. Only the samples decided against the page image are comparable; for those, fifteen of the eighteen volumes are still missing.
7. Known limitations
Everything measured so far is machine-based. Double transcription compares two machine readings with each other; the geometric letter-spacing check compares a script with a machine markup. A double-blind human comparison — two people who independently collate the same pages against the print without knowing each other’s reading — is still outstanding. Only that could confirm or refute the approximate rate; until then, all numbers on this page are machine self-reporting.
- Two readings share weaknesses. Double transcription finds what two independent passes do differently. What both get wrong in the same way, it does not find. Numbers from double transcription are therefore upper bounds of quality, not neutral estimators.
- The sample draws what is common. Randomly chosen pages are predominantly running prose; indexes, tables, and formula setting are underrepresented — and it is precisely there that the error rate is higher.
- Words are measured, not zones. Two readings of the same region never ask whether the region was chosen correctly. Whether a line belongs to the reading text or to the apparatus is not checked by this measurement.
- The geometric letter-spacing measurement has blind spots: words of fewer than four letters cannot be measured reliably, attached punctuation distorts the measure, and with several letter-spaced words in a row an internal gap is sometimes chosen instead of the word boundary.
8. How to cite
This edition is an open beta. Although the text is very accurate as measured on samples, we recommend citing the Gesammelte Werke in publications and consulting their apparatus.
The numbering of the Gesammelte Werke is authoritative: volume, page, line. The line is set in superscript — a period in between reads like a decimal point.
GW 21, 3514
For a passage spanning several lines or pages:
GW 21, 3514–19 · GW 12, 327–334
The reader adds this reference to the text itself when copying, if With source reference is turned on under the gear icon.
Bibliography entry
What belongs in the bibliography is the edition cited, and that is the Gesammelte Werke:
Georg Wilhelm Friedrich Hegel: Gesammelte Werke. Edited by the North Rhine-Westphalian Academy of Sciences, Humanities and the Arts in association with the German Research Foundation. Hamburg: Meiner 1968 ff.
This digital edition need not appear there. It provides no text and no numbering of its own: volume, page, and line are those of the Gesammelte Werke, and a reference such as GW 21, 3514 leads to the same place in the print. Anyone who nonetheless wants to document the access used can add it informally; it is not necessary.