A bundle claims first_seen dates per detected skill. Libraries have real-world release dates. A commit that imports a library before that library existed is fabricated history. Period.
Example with our own taxonomy: auth/better-auth maps to the better-auth package, first published in 2023. A bundle claiming first_seen: 2022-06-x for that slug is self-refuting. No forensics needed, just a date comparison.
This is one of the strongest cheap filters available against fabricated history, because people who fake a repo almost never think about it. The data to build it is already public: signatures/package-map.json gives slug-to-package mappings, and registries (npm today, others as detection grows) expose first-publish dates.
Two places it can run:
- In the CLI, at
submit time (this repo). Submit already has network access, so it can fetch first-publish dates for the bundle's detected packages and warn before upload: "your first_seen for X predates its first release; this will not survive verification." A courtesy honesty check, never a block. scan stays zero-network, untouched.
- Server-side at verification time, independently. Out of scope for this repo; noted for completeness.
Scope for a first PR: submit-time warning, npm-only, using the package-map slugs present in the bundle. No schema change, no new data leaves the machine (the registry queries reveal only which public packages were detected, and they ride the same network consent submit already has; worth discussing whether even that should be listed in the pre-upload consent box).
Open questions:
- Should the registry lookup result be cached locally to keep submit fast?
- Monorepo edge: a vendored/forked copy of a lib predating its public release is legitimate; hence warn, never block.
Credit where due: this came out of a design conversation with an outside reviewer who proposed it as "cross the taxonomy against release dates and you have a brutal detector, because fabricators never think of this." They're right.
A bundle claims
first_seendates per detected skill. Libraries have real-world release dates. A commit that imports a library before that library existed is fabricated history. Period.Example with our own taxonomy:
auth/better-authmaps to thebetter-authpackage, first published in 2023. A bundle claimingfirst_seen: 2022-06-xfor that slug is self-refuting. No forensics needed, just a date comparison.This is one of the strongest cheap filters available against fabricated history, because people who fake a repo almost never think about it. The data to build it is already public:
signatures/package-map.jsongives slug-to-package mappings, and registries (npm today, others as detection grows) expose first-publish dates.Two places it can run:
submittime (this repo). Submit already has network access, so it can fetch first-publish dates for the bundle's detected packages and warn before upload: "yourfirst_seenfor X predates its first release; this will not survive verification." A courtesy honesty check, never a block.scanstays zero-network, untouched.Scope for a first PR: submit-time warning, npm-only, using the package-map slugs present in the bundle. No schema change, no new data leaves the machine (the registry queries reveal only which public packages were detected, and they ride the same network consent submit already has; worth discussing whether even that should be listed in the pre-upload consent box).
Open questions:
Credit where due: this came out of a design conversation with an outside reviewer who proposed it as "cross the taxonomy against release dates and you have a brutal detector, because fabricators never think of this." They're right.