Preserve native authoring evidence and compared semantic contracts

2026-09-11 02:04:38 -07:00
parent 80cf197be6
commit c6d296a7e2
59 changed files with 11952 additions and 76 deletions
@@ -0,0 +1,29 @@
+# Analysis context experiment
+
+This standalone Rust program compares explicit baseline/override studies (B) with complete independent case bindings (D). It is an isolated experiment, **not Studio syntax or a production model implementation**. Its expression language is deliberately limited to integer magnitudes in milliseconds, addition/sum, comparison, missing inputs and pure reusable functions.
+
+The source owns the fixture generator, parser, explicit context resolver, dependency-cycle check, evaluator, frozen-snapshot operation, provenance observations and line-diff counter. All comparisons use the same 20 fixed inputs, 2 varying inputs and 20 named cases. The formula/check variants are either inline or factored into reusable pure definitions.
+
+From the repository root, using the installed local toolchain:
+
+```powershell
+. .local/toolchain/env.ps1
+New-Item -ItemType Directory -Force .local/analysis-context-20260911 | Out-Null
+rustc --edition=2024 -D warnings -C linker=clang.exe harness/experiments/analysis-context-20260911/main.rs -o .local/analysis-context-20260911/analysis-context.exe
+if ($LASTEXITCODE -ne 0) { throw 'Experiment compilation failed' }
+& .local/analysis-context-20260911/analysis-context.exe harness/experiments/analysis-context-20260911/generated
+if ($LASTEXITCODE -ne 0) { throw 'Experiment assertions failed' }
+```
+
+The program overwrites its named generated fixtures/results in the output directory. It does not invoke Cargo, modify `target`, or import Studio modules. `EVIDENCE.json` was recorded separately after the verified run and contains exact compiler, program/executable/result hashes and every generated source fixture hash. Rerunning after editing the program requires refreshing that evidence rather than trusting the old manifest.
+
+- [Results](generated/RESULTS.md): executed counts and assertions.
+- [Evidence](generated/EVIDENCE.json): final run identity and 29 fixture hashes.
+- [B with factored calculations](generated/B-factored.study) and [D equivalent](generated/D-factored.study).
+- [B case provenance](generated/B-factored-provenance.md).
+- [Frozen B case](generated/B-factored-frozen-case20.study): complete independent snapshot, copied definitions and explicit captured row origins.
+- [Interpretation and limitations](../../release-analysis-context-experiment-20260911.md).
+
+The small dialect has explicit `study B/D revision`, a shared property catalog, optional pure `function` declarations, an explicit named `baseline` for B, and named cases declaring either `base engineering` or `independent`. `missing` is an entire binding hole. Removing a B override resolves the baseline; removing a required D row rejects. Function parameters cannot capture project properties. No case-to-case inheritance or implicit selected/default case exists.
+
+Generated studies do not have OA/SA subjects, UUID identity, quantity kinds, persistent source editing, JSON/REST, library import resolution or native UI integration. The experiment compares context semantics and the actual emitted text size/edit count, not human completion time or end-to-end product conformance.