Papers
Measurements, written up. Every one names what would have proved it wrong before the first number was taken, says who had already done it and ships the code that produced the figures. Each has a permanent DOI and a PDF.
- Duplicate copies hold 14.0 to 17.2% of node_modules and most of them are the same version twice
5 real projects walked on disk, 5,127 installed packages, duplicates charged in bytes rather than counted as graph nodes
- A random primary key costs 2.8x on insert and nothing at all on a point read: a replication at 20 million rows
UUIDv4 against UUIDv7 and a bigint TSID on Postgres 16, identical rows, cold reads separated from warm ones
- Disclosure of a tool outage runs from 20 of 20 to 2 of 19 inside one model family
A live MCP server over government data, 200 calls and an unplanned API failure that split the run by the server log
- A census of 4,296,340 npm packages: 84.9% list one maintainer and 1.74% run code when you install them
Every name in the replication index fetched over 48 hours at 25 requests a second, counted in one streaming pass, with a second pass that separates the install hooks that fire from the one that does not
- No writing tell separates human from machine on its own: the best of 10 reaches AUC 0.80
40 engineering texts published before 2020 against 120 texts from 4 models, counted per 1,000 words
- Dependency trees did not grow: 25 of 39 frozen manifests held flat or shrank
42 application stacks resolved at 25 monthly snapshots across npm, PyPI and Maven Central, from registry metadata with nothing installed
- All 4 models land within 0.146% on founder arithmetic; only 1 of them is exact 10 times out of 10
10 seeded SaaS problems, every tool and MCP server disabled, ground truth computed in code
- Transaction rollback lost every one of 4,500 paired previews, at 6 to 7 ms a preview
2 ways to build a live preview on Postgres 16, measured paired and interleaved across 5 forked JVMs
- Disclosure of tool failure is bounded by the tool: 39 of 40 against 0 of 40
2 Anthropic models, one MCP tool broken 5 ways, 100 calls judged against the server log
- In 40 calls with the tools removed, not one reply said the tools were missing
A production agent CLI with its tool list stripped, 2 Anthropic models, ground truth regenerated per run