Metamath in Sophia: the first real corpus (B-16 rung 1), and the first language whose analysis Sophia has checked against an independent reference implementation. 2026-10-20.
Sources: Megill and Wheeler, Metamath: A Computer Language for Mathematical Proofs (2019); set.mm and iset.mm (github.com/metamath/set.mm); metamath-exe (github.com/metamath/metamath-exe, commit
9898f5d, built from source);crates/sophia-ingest/src/metamath.rs(metamath); Real Code Corpora §3–§4;sophia-mm-check(sophia-mm-check).
Status by level
| Level | Status | Evidence |
|---|---|---|
| L1 Parsed | yes: the whole format (scoping blocks, $c $v $f $e $d $a $p, normal and compressed proofs), except file inclusion $[ … $] (refused) | set.mm (51 159 assertions) and iset.mm (17 398) parse |
| L2 Normalised | yes (sophia-metamath@2): a statement is its mandatory hypotheses in order, typecode, math string and disjoint-variable pairs, variables numbered by first occurrence; a proof is a separate record | renaming invariance tests; 12 927 of the 13 138 labels set.mm and iset.mm share are one statement |
| L3 Resolved | yes: labels are global and declared before use, so resolution is exact | — |
| L4 Lowered | stored as statements, proofs (evidence) and library collections (axioms, definitions, proofs); not translated into sophialang | — |
| L5 Executable | not applicable in the usual sense; proof checking is not done by Sophia | — |
| L6 Verified | dependency analysis checked: “which axioms and definitions does theorem T rest on” agrees with metamath.exe on every theorem checked | below |
The cross-check (SR-128, 2026-10-20)
Sophia’s answer is a query: the closure of a proof’s references through proofs, filtered to the library’s axioms (or definitions) collection. metamath.exe’s answer is show trace_back T /axioms /essential. Compared by statement id.
| Library | Theorems checked | Agree | Dependencies compared |
|---|---|---|---|
| set.mm | 302 (seed 1, incl. 2p2e4, fta) | 302 | 8 503 axiom, 24 471 definition |
| set.mm | 3 000 (seed 7) | 3 000 | 83 671 axiom, 242 975 definition |
| iset.mm | 1 001 (seed 3, incl. 2p2e4) | 1 001 | 35 676 axiom, 48 995 definition |
| set.mm | all 48 083 | 48 083 | 1 346 571 axiom, 3 934 015 definition |
| iset.mm | all 16 483 | 16 483 | 587 185 axiom, 802 318 definition |
Six theorems prove syntax ($p wff …: weq, wel, bj-0, sn-wcdeq); for them metamath.exe also lists the syntax constructor they use. Not a disagreement: the question is about axioms and definitions.
So the earlier claim “2 + 2 = 4 rests on 15 axioms in set.mm” is now checked: the same 15 (ax-mp, ax-1 to ax-9, ax-gen, ax-ext, ax-1cn, ax-addcl, ax-addass) and 29 definitions.
Storing is not checking
Sophia stores proofs and walks their references. It does not verify that each proof step is a valid substitution. A store that held an invalid proof would answer dependency queries about it just as confidently. Proof checking would be:
- cheap (metamath.exe verifies all 48 083 proofs of set.mm in 2.85 s, measured 2026-10-20), so it is a natural first checker for SPEC-006 §5’s trust records:
checked by metamath.exe <version>, or by a small Rust verifier in Sophia; - the right place to test “evidence with trust levels” end to end.
Not done yet; listed in Semantic Boundary Experiments E-0.
Limitations
- File inclusion (
$[ … $]) is refused; set.mm and iset.mm do not use it. - Being an axiom is a library’s role, decided by the
$akeyword and the label prefix (df-for definitions), as metamath.exe’s own listings do. - The store is about 11× the source (Current Inventory).