"HyperResearch Claims vs. Reality: A Line-by-Line Verification"
It’s called HyperResearch.
Give it one prompt. It searches the web, checks sources, finds conflicts, and builds a detailed report.
→ Uses a 16-step research process
→ Checks whether citations support each claim
→ Runs four critics against every draft
→ Saves sources in a searchable research vault
→ Can resume a research run after a crash
→ Supports quick reports and long research projects
The developer says it leads an internal DeepResearch-Bench test. Third-party validation is still pending.
It requires Claude Code and Python 3.11 to 3.13.
The repo has over 1,000 GitHub stars and an MIT licence.
I found the actual repo (jordan-gibbs/hyperresearch) and cross-checked the claims against the README and a third-party listing. Here's the fact check:
✅ Confirmed as stated
| Claim | Verdict |
|---|---|
| 16-step research process | The README lists a 16-step pipeline (plus sub-steps 1.5 and 14.5), routed through a step-skill architecture. |
| Checks whether citations support each claim | A dedicated cite-check step verifies citation-sentence bindings via a skeptical LLM spot-check, backed by a "quote-integrity" rule that blocks any quoted span not found verbatim in the vault. |
| Four critics against every draft | Step 12 runs four adversarial critics in parallel — dialectic, depth, width, and instruction critics — producing findings that a tool-locked patcher applies as surgical edits. |
| Searchable research vault | Every fetched source lands in a SQLite-indexed vault, with markdown notes as the source of truth and SQLite as a rebuildable cache, plus full-text and semantic search commands. |
| Resume after a crash | Each run keeps a manifest, and a crashed run resumes exactly where it stopped via run resume. |
| Quick reports + long research projects | A light tier runs a ~30–40 minute fast path, full is the ~1.5–2.5 hour default, and an opt-in dissertation tier produces 25K–80K-word chaptered reports over 4–8 hours. |
| Requires Claude Code + Python 3.11–3.13 | Requirements are listed as Python 3.11+ and Claude Code, with the install instructions specifying Python 3.11–3.13 (3.14 not yet supported). |
| MIT license | The repo is MIT licensed. |
| 1,000+ GitHub stars | The repo currently shows 1.1k stars, 107 forks, and 9 watchers — matches "over 1,000." |
⚠️ Needs a caveat — the "internal benchmark" framing
The prompt's summary ("leads an internal DeepResearch-Bench test, third-party validation pending") is accurate but slightly undersells how loaded the claim is. The README itself flags it as a forward-looking projection from a stratified pilot against a DeepResearch-Bench leaderboard snapshot, with third-party validation still pending. In plain terms: this is the developer's own extrapolated estimate of where the tool would land on a leaderboard, not a completed, independently-run benchmark. That's a meaningfully weaker claim than "leads a benchmark" reads on first pass — worth flagging if you're citing it as evidence of quality.
One thing not in the original prompt but worth knowing: the tool runs entirely on Anthropic models via its subagent roster, and usage/cost scales with tier, gear, and corpus size — it doesn't replace your judgment on source selection, and its lint gate catches structural failures but can't guarantee factual accuracy.
Bottom line: every specific factual claim in the description checks out against the live repo. The one place to keep your skepticism up is the benchmark line — it's an honestly-labeled internal projection, not a verified third-party result, and the developer's own README says so.