Repository navigation
Description
Problem
The policy stage of lexlean verify stops at the first axiom-policy violation. A project with several wrong policies learns about them one verification at a time. Since a policy is exact and transitive (§17.12, §22.6), changing one helper's body commonly changes the policy of every caller, so several violations together is the usual case, and each extra round costs another verification.
Exact, transitive policies themselves are as specified and are not part of this report.
Reproduction
Run at pinned 3877351 (lexlean 0.4.0, language 1.2, Lean 4.32.1). origin/main and origin/release/v0.5 were read, not run: the same six LLV7005 sites in crates/lexlean/src/verify/mod.rs, each return Err(fail(...)).
\begin{lexlean}{M}
\useglossary{lexlean.std.nat@1.2.0}
\title{Natural number addition}
\begin{semanticmodule}
definition{
name: first,
parameters: [xs : (List String)],
result: String,
body: (prim join xs "," : String),
executable: true
}
definition{
name: second,
parameters: [xs : (List String)],
result: String,
body: (prim join xs ";" : String),
executable: true
}
definition{
name: third,
parameters: [b : Bytes],
result: (Option String),
body: (prim utf8_decode b : (Option String)),
executable: true,
axioms: [propext]
}
\end{semanticmodule}
\end{lexlean}
Observed, with all three wrong:
error[LLV7005]: `K1.M.first` violates its none axiom policy: observed [propext]
--> lexlean.toml:1:1
--diagnostic-format json gives one entry in diagnostics.
Expected: one LLV7005 each for first, second and third, in canonical order. The audit has already produced every observed set by this stage (§22.1 stages 9-11), and limits.max_diagnostics (default 256) already bounds a failure's output. LLV7002 is reported through LexLeanError::from_diagnostics, which can carry several and sorts them canonically. §22.6 does not ask for a stop at the first violation.
Second observation: the location
The diagnostic above is placed at lexlean.toml:1:1, although the declaration has a source position, which LLV7002 for the same module does report (src/M.lex.tex). The language-1.0 fixture tests/negative/axiom-policy-excess has its primary in src/Main.lex.tex.
From reading the code on release/v0.5: only the first of the six sites (line 1984) calls with_span. The semantic site that fires here (line 2163) builds its Diagnostic with no span, and so do the other four. Separately, declaration_spans is frozen from document.declarations(), which walks the document's blocks, while the semantic site iterates semantic.declarations, so the map may also need entries for semantic declarations. That last point has not been confirmed.
Proposed fix
- In stage 11 of
verify, collectLLV7005diagnostics across all six sites, and return them withLexLeanError::from_diagnosticsonce the stage has finished. Attestation is still written only if there are none. The internal "absent from the axiom audit" failures can stay immediate. - Freeze a source span for each semantic declaration alongside
declaration_spans, and attach it withwith_spanat the semantic sites.
Accepted when
- The module above reports all three violations in one run, in a stable order, in both human and JSON output.
- Each language-1.2 violation has its primary in the declaration's source file.
- A project with one violation produces the same result as today, other than the corrected location.
- A register row and fixture cover this (R3).
Activity
afflom commentedon Nov 3, 2026
I am taking this bounded diagnostics fix on release/v0.4. I will preserve immediate internal/process failures, collect policy violations through the existing canonical diagnostic path and limits, and attach declaration source spans. This work will avoid the kernel/replay optimization in #391; the issue-closing PR will include focused regression evidence.
afflom commentedon Nov 3, 2026
Completed through merged PR #403 on
release/v0.4, merge5a44fe49e5d73569c3bb377bb28b56e463ee0547. The final reviewed head215a9d591passed all 43 jobs in CI 41658939956, plus honesty 41658939965 and reproducibility 41658939998. Actual merge tree88d8b657f1ab621bf736b9d7e5d97fcb9c834e5eexactly equals the reviewed/tested head and synthetic merge tree. Explicitly closing this issue because the landing is on a non-default release branch.