[burtenshaw]

Summary

Separates @rycerzes's RFC 014 design from the experimental implementation in #1500, as requested in #1500 (review). Discussion and origin: #1382.

rfcs/
  014-environment-services.md  design, still In Review
  README.md                    index entry

The nine original RFC revisions are replayed with their author/date preserved and original commit SHAs recorded in each commit. The author credit and prototype results remain. The prototype link points to #1500.

Adds the proposed submit/inspect/cancel/logs/price/artifacts provider contract. A second backend and native no-network support are prerequisites before step 2; Docker and Modal remain proposals. Core manifest types and reserved jobs_ names are deferred. Unscored None rewards depend on RFC 013 acceptance and its final semantics. Examples use step(CallToolAction(...)); direct-MCP episode enforcement is tracked in #1507.

Evidence

• Before: RFC and implementation shared Add experimental posttrain_env prototype (RFC 014) #1500; examples used direct MCP calls.
  After: diff against main contains exactly the two RFC files; example executed with fake HF Jobs through server serialization/client parsing and returned final reward/spend. No paid jobs.
• Test plan: whitespace checks and RFC example execution pass. The env validation is reported on Add experimental posttrain_env prototype (RFC 014) #1500. Full local pytest selection: 3958 passed, 150 skipped, 79 deselected, 3 failed (259 s). Failures are in unchanged main files: two harness-process tests raised macOS os.killpg PermissionError during the full run, but passed in isolation on main; the Harbor occupied-port test failed identically on main (expected "already in use", observed startup timeout). These results do not establish a regression from this change. Full lint has 47 pre-existing main formatting findings plus the two known usort files; changed files pass. The test hook itself lacks timeout on this host, so the same pytest selection ran directly. CI was pending at the post-push check; no monitoring requested. Repository formatting has existing failures on main; no Python or runtime files change here.

Merge Danger

Door: two-way

Documentation only; this PR does not mark the RFC accepted or adopt core APIs.

Blast Radius: design

Provider contract, second-backend validation and RFC 013 semantics remain review decisions.

[rycerzes]

Thanks @burtenshaw! This covers everything @sergiopaniego asked for in #1500. Three small things:

1. Submit should require secrets for grader jobs only. The hidden split's seed depends on it.
2. The Evaluator section still reads the score from the logs, but Logs says the result comes separately. One line on the planned /data/result.json would fix it.
3. For public data, Sergio suggested "a contamination check or an allowlist". Only the check is in the RFC.

[burtenshaw]

@rycerzes Addressed the three wording fixes from #1508 (comment) in 2be7239:

1. Submit now requires support for secrets for grader jobs only, for the hidden split's seed; agent jobs receive no secrets.
2. Evaluator explicitly says the current prototype reads POSTTRAIN_SCORE from logs. Writing /data/result.json and reading it separately from diagnostic logs is a planned follow-up, not implemented.
3. Network now requires a contamination check after the run or an allowlist of permitted inputs for tasks with public test data.

Only rfcs/014-environment-services.md changed. RFC status remains In Review; attribution, history, prototype results, provider neutrality, second-backend/native no-network prerequisites and the RFC 013 caveat are preserved. No environment or core-type changes.

Validation: diff whitespace, Markdown structure/format, unchanged links/local anchors and single-file scope checks pass. No runtime tests run for these wording-only edits. At the post-push snapshot, no checks or reviews were reported; CI/review remain pending. No monitoring requested.