Retain the authenticated promotion and workspace adapter fixes. The prospective confirmation supports memory based algorithm selection on fresh input draws in twelve synthetic u16 sorting contexts. Broad autonomous compiler improvement, algorithm invention and transfer to independent kernel families are not proved.
Validation was completed locally on branch proof/authenticated-evaluation.
Publication was subsequently authorized by the user. The exact validated source
snapshot has tree 83c9f76c02a734fed40c05cf4a0d2c3cec72abef and local tip
9ccbbcd. The original local commit history, including preregistrations, is
preserved in history.bundle. No release, deployment or merge is included.
Training used 12 contexts and three universally applicable sorting algorithms: unstable sort, counting sort and two-pass radix sort. Every candidate was run through the independent evaluator. Seven successful training memory records covering three strategies were frozen. The unchanged HashedFeatureRanker chose one algorithm per test before any test outcomes were exposed. Feature metadata describes known size/range/order context; runtime feature extraction is not implemented or included in these measurements. This is a study harness using Rune's sealed recipe snapshot/ranking API, rather than a new autonomous proposer.
The original 34-kernel study's Insufficient and the new 36-case pilot's NoDifference remain in the record. The pilot measured 50.0% versus 36.1% random and 33.3% fixed, but its required Wilson intervals overlapped. Its acceptance threshold was not relaxed. The separately preregistered confirmation used the same frozen memory, twelve contexts and 24 fresh seeds, with no further training.
| Arm, budget one evaluation per case | Success | Wilson 95% |
|---|---|---|
| Frozen training memory | 154/288 = 53.47% | [47.70%, 59.15%] |
| Uniform random among three applicable algorithms | Expected 35.53% | [30.23%, 41.22%] |
| Counting, selected by training accept count | 106/288 = 36.81% | [31.44%, 42.52%] |
Memory clears random by 17.94 percentage points and the fixed control by 16.67. That is 50.49% and 45.28% relative success uplift. These are selection success rates, not an application runtime speedup. Random is the exact expected rate from all three measured outcomes, not thousands of independent Monte Carlo workloads. Other fixed controls were radix 33.68% and unstable 36.11%; counting was also the strongest fixed strategy observed in confirmation.
Both paired case and seed-block sign-flip tests returned Monte Carlo p about 0.00001 at 100000 draws (resolution floor), below the preregistered 0.0125 threshold for each control. Each seed block averages its twelve contexts. All 24 seed blocks favored memory over both controls. Required interval separation and >=10 percentage point margins also passed. The verdict is Better under the frozen rule; no optional sample extension or test driven tuning occurred.
There were 36 training, 108 pilot and 864 confirmation evaluator runs: 1008 fresh reports in total. Independent auditing matched baseline source, candidate source, contract structure, contract file, holdout and evaluator executable digests for every report. Every accepted report had 40 paired samples, a strict lower 95% speedup bound above 1.10 and a passing n+7 guard with at most 10% slowdown. Correctness is checked against Rust stable sort on empty, singleton, odd, random and extreme inputs. All strategies apply to every case, so simple applicability filtering cannot account for the result.
Protocols and raw evidence:
study/RESULT.md: original study, unchanged.study-transfer/PREREGISTRATION.md,result.json,audit.jsonandreports/.study-transfer-confirm/PREREGISTRATION.md,result.json,audit.json,selection.json,selection-freeze.json,freeze.jsonandreports/.
The pilot inputs were hash frozen before evaluation; its Git commit followed
the first training outputs. Confirmation inputs and choices were committed at
e7dfc44 before any confirmation evaluation. Generated confirmation source is
reconstructed by the frozen generator and verified against freeze.json.
Previously, an invented unsigned Accept promoted an unchanged hello fixture; the identical report could authorize x86_64 and aarch64. ADR 011 now requires a domain separated Ed25519 receipt and a separately supplied trusted public key outside the candidate workspace. Report bytes are read once and authenticated. The measured baseline, compiler, target, policy and evaluator must match.
Executable regressions now reject unsigned reports, changed reports, tampered receipt identity, wrong keys, candidate supplied trust anchors, ARM relabelling and trusted-signer context mismatches. A real sqdist evaluator run produced an Accept at 6.756x, 95% CI [6.455, 7.071], 40 paired samples; its receipt promoted the exact candidate and rollback restored the original plan. CLI controls against that real receipt rejected altered bytes, ARM relabelling, a wrong public key, and degraded evidence without explicit opt in. The ephemeral private key was removed; the retained public key and signature allow verification.
This proves authenticity checks assuming an uncompromised trusted signer. It does not protect a signing key from malicious processes sharing its account. Hostile build scripts require a separate signer service and stronger job isolation. Keys are never generated or trusted automatically by Rune.
Two failing-before/passing-after regressions fixed completion-order dependent artifact selection and duplicated relative -C target paths (ADR 012).
A snapshot of Rune's own workspace at e7dfc44 built offline with real locked
dependencies. Plain Cargo and Rune selected rune-study with SHA256
46274af1e47e4dbfaac3fd24d3c38815289784de4a02a589d06b9ce512c82162.
A fresh target-directory rebuild reproduced the same digest. The corrected
relative -C command used the intended target directory and matched that digest.
This is one explicitly selected workspace product, not the complete artifact set.
Targeted checks: 63 library unit tests (including four Cargo adapter tests), 23 CLI/promotion regression tests and 13 artifact envelope tests passed, 99 unique selected tests in total. Clippy passed for changed crates and all targets with warnings denied. This was not a full platform or adversarial sandbox audit.
Host: AMD EPYC 9V74, x86_64 Linux, rustc 1.99.1 / LLVM 23.1.1, shared environment. Network namespaces were unavailable. The strict evaluator correctly refused; known fixture/study runs explicitly used --allow-degraded, reporting environment scrubbing, plain copies, CPU/memory limits and timeouts. Production isolation is not established. Ordinary Cargo offline mode restricts dependency resolution, while existing build scripts remain trusted host code. No paid model call was used.
Training/pilot ran three core-pinned queues; confirmation used six. Timing calibration, algorithms, memory and gate thresholds were frozen. Confirmation invoked the prebuilt benchmark directly to avoid Cargo process startup per sample. The timed operation and five-batch median were unchanged. Shared-host load and within-family context reuse limit generalization.
cargo build -p rune-toolchain -p rune-toolchain-eval -p rune-toolchain-memory --locked
cargo test --workspace --lib --locked
cargo test -p rune-toolchain --test cli --test promote --test proof_boundary --locked
cargo test -p rune-toolchain-target --test envelope --locked
python3 study-transfer-confirm/run.py generate
python3 study-transfer-confirm/run.py analyze
python3 study-transfer-confirm/audit.pyThe last three commands reconstruct frozen source, analyze retained reports and
audit their bindings. A fresh performance rerun requires preserving the original
reports first, then running run.py run; new machine outcomes may differ. Do not
replace the committed confirmation result with a new run or reinterpret this as
a general self improvement benchmark.
Next acceptance gate: freeze an independent real-project workload family and demonstrate the same improvement over both random and training selected fixed controls, with trusted job signing isolated from candidate execution.