hyperresearch Blocks a Hallucinated Quote From Shipping. Its Own Headline Number Ships Unchecked.
A deep-research harness for Claude Code with genuinely strict verification gates, and what the caption under its benchmark chart teaches you about reading any repo README.
There is a line buried in the hyperresearch README that most research tooling would never write down. Describing how the project's tone controls work, it says the cite-checker and the ship gate "receive no shim at all. Verification never softens by mode." You can tell the harness to write a gentle explainer or an aggressive advocacy piece, and the register changes what the critics look for, but the part that checks whether the quotes are real does not move.
That is a real engineering commitment, and it is rarer than it should be. It also sits about 400 words below a sentence claiming the project "currently leads the DeepResearch-Bench RACE leaderboard," whose supporting caption calls the result a "forward-looking projection from a stratified pilot" and adds that "third party validation is pending."
Both of those things are in the same file. Holding them next to each other is the most useful thing you can do with this repo, and not because the project is dishonest. It isn't. It is because hyperresearch happens to be the clearest available illustration of a gap that exists in nearly every agent tool you will install this year: the standard a project enforces on its output and the standard it applies to its own marketing are set by two completely different mechanisms, and only one of them is code.
What the thing actually is
hyperresearch, by Jordan Gibbs, is a 16-step research pipeline that runs inside Claude Code. pip install hyperresearch && hyperresearch install, then /hyperresearch <anything> in a Claude Code session. MIT licensed, copyright 2026 Jordan Gibbs, v0.11.1 tagged on September 11, Python 3.11 to 3.13 (not 3.14 yet). It was sitting at Trendshift's #18 spot this morning. Its star count is one of the few figures I will not quote, because cache-busted shields.io and ungh.cc disagreed by roughly 10% across the morning's reads.
The pipeline itself is the interesting part. A query gets decomposed into atomic items and classified into a tier. light runs five steps in about half an hour. full, the default, runs all sixteen plus a citation check in roughly 90 minutes to two and a half hours. dissertation is opt-in and writes 25,000 to 80,000 words across chapters from 300 to 450 sources over four to eight hours. Orthogonal to tiers are "gears," which set source targets and depth budgets, and "levers," which set the register the report is written in: teach, survey, analyze, or advocate.
The subagent roster is explicit and published: fetchers on Sonnet running 8 to 12 in parallel per wave, draft orchestrators and four adversarial critics on Opus, a cite-checker on Sonnet. Every model assignment is profile config rather than hardcoded, so you can drop every fetcher to Haiku with one line in .hyperresearch/config.toml.
The gates are real, and they are unusually specific
Here is what makes this project worth your attention rather than a shrug. Most research harnesses promise accuracy. This one names the mechanisms that enforce it, and several of them are structural rather than instructional.
Patch, never regenerate. After the synthesis step produces the report, the only permitted modification is a surgical Edit hunk. The patcher, the polish auditor and the readability step are tool-locked to [Read, Edit] at the Claude Code allowlist level, so they physically cannot Write a new draft. Per-hunk caps make "just rewrite the whole section" mechanically impossible, and a critic finding that will not fit in a small hunk escalates as a structural issue instead of triggering a rewrite behind your back. This is the difference between telling a model not to do something and removing the tool.
Quoted text must exist. A quote-integrity lint blocks any quoted span that does not appear verbatim in a vault note. A hallucinated quote cannot ship, not because a reviewer might catch it but because the gate fails.
Retractions are a hard error. Citing a retracted source without acknowledging the retraction blocks the ship, and a retraction sweep re-checks every cited DOI fresh at ship time, including on vault sources reused from old runs. Retracted sources get their quality score floored to near zero.
Syndication does not count as consensus. An independence audit clusters derivative copies, so five reprints of one press release argue with the weight of one source. If you have ever watched a research agent "confirm" something across five outlets that all rewrote the same wire story, you know exactly why this exists.
Web content is data, never instructions. Every body fetched from the web is served inside an <untrusted-source url="..."> fence with a treat-as-data preamble, on both the note-reading and search paths. Forged fence tags inside a fetched body are neutralized and left visible for forensics. The url attribute is HTML-escaped with control characters stripped. In search, the wrapping happens after token-budget truncation so the closing fence cannot be severed by a long page. That last detail is the kind of thing you only write after somebody's page nearly got away with it.
Paywalled papers are handled honestly. A closed paper would normally enter a vault as a 1,500-character abstract that the report then cites as though it had been read. hyperresearch asks Unpaywall, Europe PMC and CORE for a legal open-access copy and stores the full text instead, and it discloses the substitution in four separate places: a banner in the note body, four frontmatter fields, an oa block in the JSON with body_is_not_from_source: true, and a line in the CLI output. When the source could not be read at all, the note is marked rescued with nothing_from_source: true, and the README tells you to take that literally: the title, the authors and every word of the body came from the substitute copy.
That last section carries a warning most tools would leave out. Unpaywall will hand back an accepted manuscript or a submitted preprint when no published copy is open, so if oa_version says acceptedVersion, check any direct quotation against the published paper before it reaches a report. This is a project that has thought carefully about the ways a research agent lies to you.
And then there is the chart
The README's first substantive sentence reads: "Hyperresearch turns Claude Code into a deep research agent: one that currently leads the DeepResearch-Bench RACE leaderboard (benchmarked internally)." Below it sits a chart showing hyperresearch at the top, ahead of several named commercial deep-research products. Below the chart, in small type, is the caption: "Forward-looking projection from a stratified pilot against the DeepResearch-Bench leaderboard snapshot. Third party validation is pending."
A forward-looking projection from a stratified pilot is not a leaderboard result. It is an estimate of where the project expects to land, built from a subset of the benchmark, compared against a saved snapshot of other people's scores. That is a legitimate thing to compute and a legitimate thing to publish. It is not the same class of claim as "leads the leaderboard," and the two are stacked in a way where the strong version is in bold at the top and the weak version is in <sub> tags under an image.
I want to be fair about this, because the disclosure is there. "(benchmarked internally)" is in the headline sentence itself. The caption says validation is pending. Nobody is being deceived by a project that tells you twice. Compare that to the average trending repo, where the number appears with no qualifier at all.
But run the project's own machinery over the claim and watch what happens. The independence audit exists because one source repeated five times should count once; here one internal pilot is presented in a visual format that borrows the authority of a public leaderboard with multiple independent entrants. The cite-checker exists because a citation should actually support the sentence it is attached to; here the chart's underlying source is a snapshot of a leaderboard that hyperresearch does not appear on. The ship gate exists to stop a report making a claim the corpus cannot carry.
If hyperresearch had written that sentence in a report, the sentence would not have shipped.
Put this into practice
Two separate things to take away, and the second one matters more.
If you want to try the tool, the low-friction path is a per-project install: cd your-project, then pip install hyperresearch && hyperresearch install, then /hyperresearch <your question> in Claude Code. Skip the --global flag on the first run. The README is explicit that global costs roughly 15 lines in every session's system reminder, which is a permanent context tax on unrelated work. Start on the light tier by asking a bounded factual question, because a 30-to-40-minute run is a much cheaper way to find out whether the output is worth anything than a two-and-a-half-hour full run. Set HYPERRESEARCH_CONTACT_EMAIL before you do anything scholarly; it enables Unpaywall, which is otherwise skipped entirely, and it puts you in the OpenAlex and Crossref polite pools. And use hyperresearch run init --budget 50 on your first real run, because crossing the cap blocks the run rather than letting spend balloon unnoticed.
The more portable practice is the reading habit. When you evaluate any agent tool this week, separate the claims the project's own code enforces from the claims that live only in prose. In hyperresearch, "a hallucinated quote cannot ship" is enforced by a lint that fails the build. "Leads the leaderboard" is enforced by nothing. Both sentences are in the same README, set in the same font, and only one of them is a fact about the software.
Then ask the diagnostic question: does the project apply its own standard to itself? A project whose answer is "mostly, with a caption" is doing better than most. A project whose README makes a benchmark claim with no qualifier anywhere is telling you it has never had to think about the difference.
Honest limitations
I have not run hyperresearch. Everything above comes from reading the README at raw.githubusercontent.com with a cache-buster this morning, plus the repo's license file and release feed. I have not verified that the quote-integrity lint, the retraction sweep or the tool-locking actually behave as described, and a README describing a gate is exactly the kind of prose claim this article is telling you to be careful with. The irony is not lost on me and you should hold my account of the gates at the same distance I am asking you to hold the benchmark chart.
I also did not attempt to reproduce or evaluate the DeepResearch-Bench pilot. The README links a leaderboard snapshot on Hugging Face; I did not fetch it, and I make no claim about where hyperresearch would actually land.
Costs are real and unstated. The subagent roster puts Opus on the draft orchestrators, all four critics, the patcher, the polish auditor and the readability recommender. A full run on the default gear targets 55 to 80 sources; premier targets 100 to 130 with a doubled depth budget and the README estimates three to five hours. The project's own "what it doesn't do" section says usage scales with tier, gear and corpus size, and puts no number on it. The budget cap is the only guardrail, and it is an estimate.
The lint gate catches structural failures, not factual ones. The README says so directly: it cannot guarantee factual accuracy, that is still your call. Anthropic models only, at present. And the star count discrepancy I mentioned is a live reminder that even the simplest number attached to a repo is often two different numbers depending on who you ask.
What to do with this
The useful frame is not "hyperresearch oversells itself." It mostly does not. The useful frame is that a project can build genuinely serious verification machinery for the artifacts it produces and still have no machinery at all for the claims it makes about itself, because the first kind lives in code and the second kind lives in a file anyone can edit.
That gap is where almost all of your risk sits when you install an agent tool. You are not usually being lied to. You are reading two different categories of sentence printed in the same typeface, and nobody has labeled which is which.
Go look at the last three repos you installed. For each headline claim, find the mechanism. If there isn't one, you have learned something about the claim.
Sources: jordan-gibbs/hyperresearch README (read via raw.githubusercontent.com, 12 September 2026), repo LICENSE and release feed, Trendshift daily board read at 07:11 ET.
Medium metadata
- Title: hyperresearch Blocks a Hallucinated Quote From Shipping. Its Own Headline Number Ships Unchecked.
- Subtitle: A deep-research harness for Claude Code with genuinely strict verification gates, and what the caption under its benchmark chart teaches you about reading any repo README
- Tags: AI Agents, Open Source, Claude Code, Research, Software Development
- Canonical: import from the fervorai.dev URL
- Suggested kicker image: a factory quality-control line where every part passes under a scanner, and the sign above the line is hand-painted