diff --git a/VERSIONING.md b/VERSIONING.md index 4d5b479..21b89fe 100644 --- a/VERSIONING.md +++ b/VERSIONING.md @@ -257,8 +257,19 @@ heading — so a reader arriving the day after the flip would reasonably conclud ### 6.1 Repository visibility — what the flip itself does -The seven repositories become public. That is the whole of it, and it is enough to unlock three -things that are unavailable while private: +**Six repositories become public** — `astro-mine-platform`, `astro-mine-cli`, `astro-mine-api`, +`astro-mine-ui`, `docs` and `.github`. Two qualifications, because "make the org public" is not the +same act: + +- **`astro-mine/embargo` stays private, permanently.** It holds the sealed held-out benchmark seed + sets that Bench discloses only at evaluation time (§`bench.md`). Publishing it would destroy the + anti-gaming property the leaderboard depends on, and no amount of later rotation undoes a + disclosure. It is not part of the flip and never becomes part of it. +- **`.repo-template` was deleted rather than flipped.** It existed to scaffold a per-component + repository, which the four-distribution model retired; publishing an empty template for a layout + the project no longer uses would be publishing a wrong instruction. + +Visibility on those six is enough to unlock three things that are unavailable while private: - **Secret scanning, push protection and branch rulesets** (`conventions.md` §9). These are not deferred by choice: the API refuses them outright for a private repository on the current plan. diff --git a/architecture/bench.md b/architecture/bench.md index a651a13..d330080 100644 --- a/architecture/bench.md +++ b/architecture/bench.md @@ -316,7 +316,11 @@ REST/OpenAPI (conventions.md §3, §4). rather than trusting reported numbers, eliminating the largest class of fabrication. - **Held-out seeds** and **hidden test scenarios**: each spec carries a public seed set for development and an embargoed held-out set disclosed only at evaluation time, so methods are - scored on instances they could not have overfit. + scored on instances they could not have overfit. The sealed sets live in the private + **`astro-mine/embargo`** repository, read at evaluation time via + `$ASTRO_MINE_BENCH_EMBARGO_ROOT`; it is the one repository that **stays private after the + public flip** (`VERSIONING.md` §6.1), because disclosure is irreversible in a way rotation + cannot repair. - **Determinism enforcement**: a sampled fraction of submissions is **re-executed** from the provenance bundle; mismatching the recorded result flags the entry (non-determinism or tampering).