A short list of places where a project's documentation contradicts itself, with the evidence.
For maintainers. It reads a project's docs, pairs the passages that restate each other, and uses the Clocked API to pick the pairs worth a person's time. Every item shows both passages with file and line. It is a list to review, not a verdict: about half of it is real (11 of 21, 52%), and on four projects it had not seen before, none of its 8 items was (0 of 8).
15 real projects measured, 2 more scanned later. 17 defects backed by source code. 5 pull requests open upstream, 2 merged.
Try it live
Send one pair of passages, or a public GitHub repository, through the same pipeline. Requests go to a small service that holds the Clocked key; it has a fixed credit budget, a cap per request and a rate limit, and it says so when the budget is used up. It logs the address, sizes and credits, not the text.
What request.body is inside the onRequest hook in fastify/fastify
73> In the [onRequest](#onrequest) hook, `request.body` will always be74> `undefined`, because the body parsing happens before the75> [preValidation](#prevalidation) hook.
154You can pass an **array of condition objects** to match multiple distinct sets of rows in a single call (conditions are OR'd together):
packages/typeorm/src/entity-manager/EntityManager.ts:897-927 has only update(target, criteria, partialEntity, options); no {criteria, data} form exists in src or test. The example came in with docs PR #12025; the feature PR #10910 implemented OR-ed condition arrays. The same PR added 'three separate queries' comments to the delete and softDelete examples, which the code also runs as one query.
184A utility used to enable `refetchOnMount` and `refetchOnReconnect` behaviors.
packages/toolkit/src/query/core/setupListeners.ts dispatches onFocus, onFocusLost, onOnline and onOffline; nothing in it concerns mounting. The same wording is at docs/rtk-query/overview.md:85 and packages/toolkit/README.md:107.
Each contradiction was checked against both passages and the project's source code before a fix was written. The final word belongs to the upstream maintainers. Status as of 2026-11-09.
Which entities em.flush() checks for changes; How to load SQLite extensions; What Type.toJSON() returns by default; Whether a raw() fragment can be serialized
What request.body is inside the onRequest hook; What request.body is inside the preParsing hook; Which HTTP methods Fastify supports; What an onSend hook may change the payload to; Which types the final payload may have
notes for 1 issue ready: Which methods a custom logger must have
The numbers, including the bad ones
1,878pairs of passages read across 15 projects
52flagged by /compare
21kept for a person to read
11real contradictions
11 of 13 (85%) on the 11 projects the listing rule was tuned on.
0 of 8 (0%) on the 4 projects scanned after the rule was fixed.
Read those two together. The rule that decides what makes the list was chosen on findings from eleven projects and then run unchanged on four more. There it listed eight items and none was real. So this is a tool that saves a maintainer from reading 1,878 pairs, not one that can be trusted to be right: about half of what it lists is real, the rate varies a lot between projects, and 9 of 15 projects produced nothing real.
Of the 17 defects, 10 were on the standard run's list, 3 came from reading more pairs in three projects, and 4 surfaced in other runs without making the standard list. Each card says which.
9 of 15: vitejs/vite, jestjs/jest, Automattic/mongoose, eslint/eslint, vitest-dev/vitest, sindresorhus/got, gofiber/fiber, pydantic/pydantic, sveltejs/svelte. In most of them it listed nothing. In jest, fiber and pydantic it listed items that turned out not to be contradictions.
What it listed and got wrong
10 of the 21 listed items are not contradictions. Four examples, one for each kind of mistake. All 10 are in review/labels_v2.json with the reason.
redirect's From is a regular expression (redirect.go:202 MustCompileRegexp), rewrite's From is path text (rewrite.go:30 MustCompilePath). Two middlewares, each documented correctly.
proxy.md:92 and whats_new.md:2036 say the same thing in different words: the body is dropped for the redirected hop, the request the handler holds keeps it.
It also dropped two real contradictions, because /compare labelled them scope differences: the rollup one on this page, and a second sighting of the MikroORM flush one.
How it works, and what each Clocked endpoint changes
Find passages that restate each other
Local, free. The docs are split at headings. Sentences from different sections that are worded almost alike but not identically become candidate pairs, and so do passages that mention the same option name.
Compare decides which pairs conflict
One /compare request per pair. Its conflict spans become the two quoted passages and their line numbers. Its conflict kind decides what happens next: value, version and rule conflicts go on, scope and name differences are dropped. On its own, 13 of 52 (25%) of what it flags is real.
Verify and a swapped Compare decide what makes the list
/grounding checks each side's sentence against the other passage, and /compare is asked again with the passages swapped, because its answer depends on their order. An item is listed only when one of the two agrees. On the tuning findings this took precision from 55% to 85%. On the four later projects it did not help.
Rerank finds where else the fact is stated
/rerank picks other passages that state the disputed fact, and /grounding checks which side they support, which names the side that is probably stale. With /rerank, 14 of 31 picked passages stated the fact, against 11 of 48 for a plain word-overlap pick. The side it named came out the same either way: of 7 real findings where the code settles which side is wrong, 4 were named with /rerank and 4 without, all correctly. So this step helps the reader more than it changes the result.
The other 14 defects
Marked words differ between the two passages. Line numbers link to the pinned commit on GitHub.
What request.body is inside the preParsing hook in fastify/fastify
104> In the [preParsing](#preparsing) hook, `request.body` will always be105> `undefined`, because the body parsing happens before the106> [preValidation](#prevalidation) hook.
575The type of the sent payload (after serialization and going through any576[`onSend` hooks](./Hooks.md#onsend)) must be one of the following types,577otherwise, an error will be thrown:578579- `string`580- `Buffer`581- `stream`582- `undefined`583- `null`
lib/reply.js:183-188 and the final send path accept web streams and `Response` objects without throwing.
How it surfaced: not listed by the standard run; it surfaced in the both-orders run below the listing bar, and in the day-1 batch runUpstream: fastify/fastify #7193, open
Which methods a custom logger must have in fastify/fastify
366A custom logger instance. The logger must be a Pino instance or conform to the367Pino interface by having the following methods: `info`, `error`, `debug`,368`fatal`, `warn`, `trace`, `child`. For example:
The runtime check in lib/logger-factory.js:68 requires info, error, debug, fatal, warn, trace and child. The TypeScript type FastifyBaseLogger (types/logger.d.ts:27) also requires `level` and `silent`. Each page matches one of the two.
How it surfaced: not listed by the standard run; day-1 batch run only (with one pair per request /compare returns no passage for it)No patch: which list the docs should give is a maintainer decision. Notes for an issue are in upstream-patches/fastify/06-logger-methods-issue-notes.md.
Which JavaScript version TypeORM targets in typeorm/typeorm
5and can be used with TypeScript and JavaScript (ES2021).
PR #11382 (drop Node 16 and 18) changed the README from ES2021 to ES2023 and moved tsconfig.json to @tsconfig/node20; getting-started.md was not touched.
149 Due to the 7 bit limitation, once a chunk has at least 127 samples,150 `st_changed_on` is set to 127 (0xEF) and the 127th and further samples will151 have `st_i` present.
301Once a chunk has at least 127 samples,302`st_changed_on` is set to 127 (0x7F) and the 128th and further samples will303have `st_i` present.
127 is 0x7F; tsdb/chunkenc/st.go:19 has `maxFirstSTChangeOn = 0x7F`. The two passages also disagree on 127th versus 128th sample; the patch leaves that to the maintainers.
389Whenever you call [`em.flush()`](/api/core/class/EntityManager#flush), the ORM will iterate over the Identity Map, and for each entity it compares the original state with the values that are currently set on the entity.
85Now whenever you call `em.flush()` MikroORM will iterate over all entities you previously marked for persisting via `em.persist()`.
packages/core/src/unit-of-work/UnitOfWork.ts computeChangeSets walks the identity map (line 769) and then the persist stack (775). unit-of-work.md:34 itself says managed entities need no persist call.
7This fragment is represented by `RawQueryFragment` class instance that can be serialized to a string, so it can be used both as an object value and key.
packages/core/src/utils/RawQueryFragment.ts:100-102: `toJSON()` returns the string `raw('...')`; only conversion to a non-string primitive throws (lines 86-94).
How it surfaced: extended run (more pairs), one check agreedUpstream: mikro-orm/mikro-orm #8617, merged
The action types createAsyncThunk dispatches in reduxjs/redux-toolkit
52The specification is now fixed, but it is only implemented in modern browsers and not finalised in Node.js.
No source file settles this; the README paragraph is the newer wording of the same paragraph and matches Node.js since 12.17.
How it surfaced: not listed by the standard run: /compare flagged it but labelled it a scope difference, which the listing rule dropsUpstream: rollup/rollup #6643, merged
The default of keepAliveTimeoutThreshold in nodejs/undici
66 * `keepAliveTimeoutThreshold` {number|null} A number of milliseconds67 subtracted from server *keep-alive* hints when overriding `keepAliveTimeout`,68 to account for timing inaccuracies caused by e.g. transport latency.69 **Default:** `2e3`.
80 * `keepAliveTimeoutThreshold` {number} A number of milliseconds subtracted from81 server _keep-alive_ hints when overriding `keepAliveTimeout`, to account for82 timing inaccuracies such as transport latency. **Default:** `1e3`.
lib/dispatcher/client.js:311 uses 2e3 when the option is unset, and H2CClient passes it through unchanged; #3145 changed it from 1e3. types/client.d.ts:61 and types/h2c-client.d.ts:43 also still said 1e3. The same PR fixes the H2CClient pipelining default in types/h2c-client.d.ts:47, found while checking.
How it surfaced: later run on 2026-11-09, outside the measured 15 projects (47 pairs read): /compare flagged it; the checks did not run because the run's credit cap was reachedUpstream: nodejs/undici #6182, open
Limits
It only sees contradictions between passages that restate each other. A page that is simply wrong, with nothing else to compare it to, is invisible.
How much it misses is not measured. A spot check of about 50 unlisted pairs found no clear miss.
Sibling pages confuse it: two releases, two middlewares, the React and the Vue page of one API. Pairs of per-version guides are skipped by file name; the rest still gets through.
The listing rule was tuned on a small set. The 0 of 8 on later projects is the fairer guide to what to expect.
Run it yourself, without a key
Every API answer behind this page is recorded in the repository, so a scan of a project at its pinned commit replays offline and spends no credits.