One paragraph scored 6.4 and 30.6. The gap is not noise

Ninety-seven words about compiler optimisations. Feed them to the seven readability indices a checking tool shows side by side, and the grades come back running from 6.4 to 30.6. A span of 24.2 grade levels, from the same 97 words.

The reflex is to shrug, call it noise, and pick one. That is wrong in a specific way. Six of the seven disagree far less than that: four land inside a 5.8-point band. One index sits 24.2 levels below the pack, and it is the only one built on a word list rather than syllable or character counts.

Every number below was recomputed against the shipped engine behind the readability checker, then checked back against what the sources say.

Key Takeaways

• 97 words, seven indices, 6.4 to 30.6. Four cluster inside 24.8–30.6; Dale–Chall sits alone at 6.4.
• On easy prose Dale–Chall inverts — hardest of seven, against ARI's 0.6.
• Reconciling them needs 139.3% and 164.7% of words flagged difficult. Neither exists.
• SMOG alone has a published minimum, 30 sentences, and alone declines below it.

What these formulas actually count

Each index takes two measurements: sentence length, and something standing in for word hardness. Flesch, Flesch–Kincaid and Gunning Fog count syllables. ARI and Coleman–Liau count characters instead, because that is what a machine can scan. Dale–Chall asks a third question: is this word on a list fourth-grade American students reliably knew?

The seven, and who each was aimed at

Index	Who built it	What it counts	Published minimum sample
Flesch Reading Ease	Rudolf Flesch, 1948	ASL + syllables per word	none found
Flesch–Kincaid Grade	Kincaid's team, 1975, U.S. Navy contract	ASL + syllables per word	none found
Gunning Fog	Robert Gunning, 1952, newspaper and textbook	ASL + % of words with 3+ syllables	none found
SMOG	G. Harry McLaughlin	count of 3+ syllable words	30 sentences
Coleman–Liau	Meri Coleman & T. L. Liau	letters/100 words + sentences/100 words	none found
Automated Readability Index	not stated in the source I read	letters + digits per word, ASL	none found
Dale–Chall	Edgar Dale & Jeanne Chall, 1948; rev. 1995	% of words off a familiar-word list, + ASL	none found

"None found" is a negative result from the sources I read, not proof that no minimum exists. Two cells are incomplete on purpose: the ARI article names no author or year, and the Coleman–Liau article names authors but no year. I would not print what I could not verify.

The dates came from bibliographic records rather than memory. ERIC's reprint collection The Classic Readability Studies lists "A New Readability Yardstick" by Rudolf Flesch, June 1948, in the Journal of Applied Psychology vol. 32 no. 3. That is the anchor for Reading Ease's 1948 date. It also dates Dale and Chall's two papers to February 1948. The 1944 date that circulates for Dale–Chall belongs to Irving Lorge's paper in the same collection. The 1995 revision is a book, Readability Revisited, built on "an updated familiar word list."

One paragraph, seven numbers

The passage under test is 97 words on compiler optimisation, five sentences, free of abbreviations, initials and decimals so that nothing about segmentation is arguable. A second passage of 28 plain-English words is scored alongside it for contrast.

Index	Hard passage (97 w / 5 s)	Easy passage (28 w / 3 s)
Flesch Reading Ease	−48.3 (higher is easier)	100.7
Flesch–Kincaid Grade	24.8	1.5
Gunning Fog	30.4	3.7
SMOG	declined — 5 sentences, needs 30	declined
Coleman–Liau	30.6	2.7
Automated Readability Index	26.6	0.6
Dale–Chall (approximate)	6.4	5.2

Sentence length roughly doubles across the two passages, 9.3 words to 19.4, and syllables per word more than doubles, 1.1 to 2.8. The syllable family punishes the second jump, the character family the first. That produces most of the spread before any formula gets a say.

The outlier inverts. On the hard passage Dale–Chall returns 6.4, the lowest grade of the six reporting, which reads as this is the easiest text in the set. On the easy passage it returns 5.2, the highest grade of the seven, which reads as this is the hardest, while ARI puts that same passage at 0.6, the easiest reading anyone returned. An index that merely disagreed would stay on one side of the pack. This one changes sides.

SMOG is the only one that refuses to answer

Wikipedia's SMOG article gives the procedure as: "Take three ten-sentence-long samples from the text in question. In those sentences, count the polysyllables (words of 3 or more syllables)." Three times ten is 30, and the same article is explicit about why the floor is not negotiable: "tables for texts of fewer than 30 sentences are statistically invalid, because the formula was normed on 30-sentence samples."

No other index here publishes a minimum. I looked and found none, which is a negative result rather than proof.

On the passage above, the checker renders that refusal as a sentence inside the tile rather than greying it out:

SMOG is calibrated on 30 sentences — this has 5

The other six return confident numbers from the same five sentences. A 6.4 and a 30.6 arrive with identical authority, and exactly one of the seven has published grounds for saying I need more text first. That is the whole warning.

The trimmed word list is not the explanation

The obvious excuse for Dale–Chall's 6.4 is that the shipped list is short. It is. Wikipedia's Dale–Chall article describes a 3,000-word list of what fourth-grade students could reliably understand, up from 769 in 1948; the shipped file holds 1,032 unique entries.

On the hard passage it flags 11 words of 97: attribute, benefit, budget, conservative, methodology, nevertheless, redundant, several, subsequent, therefore, underlying. Now notice what it misses: implementation, intermediate, representations, compilation, invalidation, amortised, optimisation, interprocedural. The passage is thick with specialist vocabulary and the list is silent on all eight.

Is a short list the whole story? No. Hold sentence length at 19.4 and walk the difficult-word percentage up:

Difficult-word %	Dale–Chall grade
0	1.0
5	5.4
11.3 — what the shipped list returns	6.4
20	7.8
30	9.3
40	10.9
50	12.5
70	15.7

Each point of difficult-word percentage buys 0.1579 grade levels. Flag every single word in the passage, which is arithmetically impossible, and Dale–Chall reaches about 20.4, still under Flesch–Kincaid's 24.8 and nowhere near Coleman–Liau's 30.6.

Run it backwards. What percentage would reconcile Dale–Chall with ARI's 26.6? 139.3%. With Coleman–Liau's 30.6? 164.7%. Both are percentages of words, and both exceed 100 by a wide margin.

So even a perfect, complete 3,000-word Dale–Chall list could not reconcile Dale–Chall with the character-based indices on technical prose. The divergence is structural, not a bundling bug. That is the finding.

Three places the docs and the source disagree

Checking the engine against its own documentation turned up three discrepancies. All are small and all checkable in the source, which is why this post runs the code rather than the manual.

1. ARI rounding. Wikipedia: "Non-integer scores are always rounded up to the nearest whole number, so a score of 10.1 or 10.6 would be converted to 11." The shipped code rounds to one decimal place, so 26.6 here would print as 27 under the published rule.
2. The Dale–Chall threshold. Wikipedia: "If the percentage of difficult words is above 5%, then add 3.6365." The code tests difficultPct >= 5. above and >= disagree at exactly 5.0%, and nowhere else.
3. The word-list count. The file header and the tool's documentation both say 1,071. The file holds 1,071 string literals and 1,032 unique words once new Set() deduplicates them.

None of the three changes the thesis. All three are the argument for reading a tool's source, since documentation is a claim about code.

What none of the seven measure

Wikipedia's Flesch–Kincaid article states the limitation in one sentence, and it is the one to keep:

"As readability formulas were developed for school books, they demonstrate weaknesses compared to directly testing usability with typical readers. They neglect between-reader differences and effects of content, layout and retrieval aids."

Content, layout, retrieval aids and between-reader differences all sit outside every formula here, which is why a grade level is not a person. "Grade 12" names a calibration point, not a reader.

The same article is candid about where the scale ends. "Due to the formula's construction, the score does not have an upper bound," and the lowest score in theory is −3.40, "belonging to the passage 'Go. See. Stop. Rest.' for example". A very low number is not a failure state and a very high one is not a wall. Wikipedia's article on readability generally counts typography as part of readability itself, listing "font size, line height, character spacing, and line length". Every one of the seven ignores all four.

A note on sourcing: that limitation rests on Wikipedia, which is tertiary, and no peer-reviewed critique was fetched here. Treat it as a responsible secondary summary, not settled scholarship.

The fitting epilogue is that the man who built the best-known formula was once asked whether it worked. IEEE Transactions on Professional Communication ran an interview with J. Peter Kincaid in 1987 titled "Readability formulas: Useful or useless?" Putting that question to the formula's own inventor is not a verdict. It is the right one to leave in the air.

What to do with seven disagreeing numbers

• Read the pattern, not the tile. Four indices inside a 5.8-point band is one finding. One index sitting 20 levels below it tells you which family it belongs to.
• Never rank your writing by one index against another tool's number. The scales are not comparable, and nothing on either screen says so.
• Go to the sentence. A 30.6 on a page with one 40-word sentence and twelve 12-word ones is a sentence problem, not a document problem. The checker flags that sentence: the output these formulas can support.
• Take the refusal seriously. When an index declines, that is the most trustworthy signal in the row.

One thing about why this page exists. /tools/readability-checker/ has zero rows in 90 days of Search Console data. Not a weak count, none at all. Its sibling the word counter earns 5,555 impressions over the same window doing something adjacent. People search for counting text; on this site's data they do not search for having text scored. This is editorial coverage, not a traffic asset.

Related Tools & Further Reading

• The readability checker computes all seven in the browser, flags the hard sentences, and shows SMOG's refusal rather than hiding it.
• The word counter renders Flesch, Flesch–Kincaid and Fog for a whole document.
• How many sentences in a paragraph covers the structural convention every formula here ignores. Sentence count is the one input all seven share.
• The online thesaurus is the next step: the checker names the difficult word, this replaces it. It does not rewrite your prose.

A readability score is a reason to read one sentence again. So is an index declining to answer.