Skip to content

Latest commit

 

History

History
103 lines (87 loc) · 6.3 KB

VERSIONING.md

File metadata and controls

103 lines (87 loc) · 6.3 KB

Versions, review snapshots and scientific releases

Identity Current value Scope
Python package 0.4.0a1 Supported alpha compiler API
Controlling technical release abi-final-validation-v2-repaired-r7-2026-08-30 Bounded R7 runtime/conformance proof
Current research review snapshot abi-research-r108-resource-stop-2026-11-14 Incomplete resource-budget stop; no English qualification or recipient result
Previous completed diagnostic snapshot abi-research-r107-2026-11-13 Natural-text residual intervention; no trained-route fluency or product promotion
Earlier research review snapshot abi-research-r106-2026-11-11 Failed narrow grammar control; no recipient comparison or English promotion
Previous research review snapshot abi-research-r105-2026-11-11 External-label evaluator rejection; no acquisition or host promotion
Previous public research review snapshot abi-research-r104-2026-11-10 Finite evaluator tests; failed control gate; recipient comparison not run
Previous research review snapshot abi-research-r103-2026-11-10 Raw control qualification failed; recipient comparison not run
Previous research snapshot abi-research-r102-2026-11-10 Negative coverage/exposure factorial; no English-core promotion
Earlier diagnostic snapshot abi-research-r101-2026-11-05 Execution/continuation diagnostic; no English-core promotion
Earlier capacity snapshot abi-research-r100-2026-11-05 Preserved R99-D1/R100 negative capacity control and public reconstruction
Earlier research snapshot abi-research-2026-11-05 Preserved R98/R99 research and public reconstruction
Default branch master Moving development and research history

The package API has not changed in this research update, so its package version stays unchanged. Experiment identifiers, canonical artifact/interface versions, package versions and Git tags are different things. Never use the highest experiment number as a measure of scientific progress.

Review tags are annotated and immutable. Do not replace a tag or silently overwrite evidence under its old path. Verify remote availability before citing a local tag as published:

git fetch origin --tags
git show --no-patch abi-research-r100-2026-11-05
git show --no-patch abi-research-r101-2026-11-05
git rev-parse 'abi-research-r101-2026-11-05^{commit}'
git lfs pull
git lfs fsck

Review checks

The supported API checks in Getting started do not require a sibling research host. Full research replay additionally requires published Git LFS payloads and the LayerCake Git history named by each protocol.

python scripts/run_review_tests.py
python scripts/check_docs.py
python -m build

The review runner executes the default suite except the exact historical Phase 7 module, then executes that unchanged module against 46 isolated files. It reads historical LayerCake bytes from commit 662c5a9b7264a1a5478c9dfb656f35c450e2504f, validates every registered binding, and never checks out or modifies the live sibling. Missing Git objects, payloads or hashes fail closed. Direct pytest against a newer sibling still rejects stale Phase 7 bindings; this is intentional, not a gate to relax.

R99 snapshot local review: 203 tests pass with four declared skips, plus 16/16 historical Phase 7 tests in their exact-source fixture. See the default receipt and Phase 7 receipt. These checks are not live inference, human ratings or independent-hardware reproduction.

R100 local review: 220 default tests pass with four declared skips, plus 16/16 historical Phase 7 tests. See the default receipt, Phase 7 receipt and 57-test supported CI surface. The package version is unchanged because no supported API changed.

R102 local review: 271 default tests pass with four declared skips, plus 16/16 exact historical Phase 7 tests. The supported CI list passes 108 locally. The annotated local tag points to d5a8957b152ce883779436103bd61a7ba4cd2f35; its tag-object identity is 1ef35b43aa14c23e1f3836e07009dcf0bd5ee1b5. See the R102 packet for complete dependency custody and the negative scientific result. The user subsequently authorized publication; the same tag and complete payload set are now public. Original inventory fields retain their historical local-only state. No package API or LayerCake source changed.

Promotion policy

R103 is sealed at source commit cc6b59cd8b0b72730a0ae847015c42a9c7a22ab5, annotated tag object db1d4b999e6351b28e552115603516a3d3f1c1fc. Its local supported review passes 313 tests with four declared skips; 12 additive completed-evidence checks pass separately. The R103 packet records the 41/48 prerequisite failure and complete public manifest. No recipient ran, no new model checkpoint was made, and the package API remains unchanged.

R101 local review: 246 default tests pass with four declared skips, plus 16/16 exact historical Phase 7 tests. The supported CI list passes 83. See the R101 packet, default report, Phase 7 report and CI-surface report. These checks and the same-machine diagnostic do not fill independent external or product-quality gates. No package version change is warranted.

  • Change the package version for supported API changes and describe migration.
  • Assign a new artifact/interface identity for incompatible execution contracts.
  • Keep every research failure and preregister the next bounded intervention.
  • Publish exact raw evidence and required artifacts before claiming public replay.
  • Require scientific gates separately from tests, tags, builds and CI status.

R7, V1089, R97, R98/C1, R99, R100, R101, R102 and R103 retain separate claim scopes. The whole-moonshot verdict is FULL_MOONSHOT_NOT_PROVEN. LayerCake owns host certification; ABI does not inherit its runtime certificates.