Evidence

The compiler doesn't take your word for it.

Five claims with receipts. Every section links to a file in the repository, a command you can run, and — where possible — a direct quote from the compiler or test runner.

The ladder stops where the evidence stops.

E0 Syntax → E1 HIR Type → E2 Effect Lattice → E3 GUM Variance Bounds → E4 Lean 4 Formal Proof → E5 Closed-form Theorem. TRUSTWORTHY is a cyan diamond at the last gate that holds. UNBOUNDED is amber: a proved model cited past its scope. REFUSED is carmine: the band is not inside the window. No card on this page reaches E5.

gum_k95 finite-dof · Student-t, not silent 1.96

TRUSTWORTHY

  1. E0Syntax
  2. E1HIR Type
  3. E2Effect Lattice
  4. E3GUM Variance Bounds
  5. E4Lean 4 Formal Proof
  6. E5Closed-form Theorem

Type-A-dominant budget (n=5) gives ν_eff≈4 and k95≈2.776. The witness integer is 2776. Wave10 closed the bitcast that printed 1.960 on this path. E4 is unreached: this is an executable gate, not a Lean theorem of the emitted factor. E5 is unreached.

Verified snapshot from repository artifacts

Vancomycin ε gate Compile-time

Compile-fail fixtures reject dosing below ASHP threshold (ε ≥ 0.82)

Stdlib reliability —

0 active entrypoints

Release provenance vunpublished

Unsigned or unverifiable in current workspace

GPU public contract —

GPU public contract · not in CI · artifact 2026-03-09 · engine unnamed

Boundaries

What is still intentionally bounded.

Source artifacts

Files this page reads.

  • Release provenance artifacts/omega/souc_release_provenance.v1.json unknown
  • GPU public contract artifacts/omega/gpu_public_contract.v1.json unknown
  • Stdlib reliability status artifacts/stdlib/stdlib_reliability_status.v1.json unknown
Release attestation

Published provenance reports unknown for vunpublished.

Operational note

The trust story is strongest where release provenance, stdlib reliability, and the public GPU contract overlap. This hub does not claim equal maturity across every research lane.

§0

Live artifact status

These tables are generated directly from committed artifact JSONs in the repository. They are not hand-curated marketing copy. If a feature is marked Verified, the artifact gate passed. If it is Blocked, the artifact gate failed and the reason is quoted from the test runner.

Compiler & Toolchain

Synced from artifacts/
Lexer / Parser / Type Checker ✓ Verified

Verified (gate pass); pre-1.0

Native Backend (x86-64 ELF) ✓ Verified

selftest_passed=true, scalar_smoke_present=true, fail_closed=true

Self-Hosted Compiler (default path) ✓ Verified

172 files, 105058 lines, cycle parity=true

Cranelift JIT (retired) ⊘ Stub

No longer shipped since 2.1.0 (CHANGELOG.md). Madaros native x86-64 is the only shipped engine.

LSP Server ✓ Verified

smoke status=pass, strict_no_rust=true

GPU Codegen (PTX) ✕ Blocked

status=fail, blockers=[remote_env_missing]

Bootstrap Chain ? Unknown

full_concat=pass, knowledge_bootstrap=pass

Generated 2026-10-27T10:02:46.672Z

Standard Library & Science Lanes

Synced from artifacts/
Core Standard Library ? Unknown

Stdlib end-to-end: 196 of 533 test programs pass (165 fail, 172 skipped), measured 2026-10-06 with `bash scripts/stdlib/run_stdlib_e2e.sh`; result in `artifacts/stdlib/stdlib_e2e_result.v1.json`

Scientific Pipelines ✓ Verified

lanes=2, status_summary=pass

fmri darwin_pbpk
Hyper-Execution Neural Lanes ✓ Verified

lanes=7, status_summary=pass

nn onn qnn snn spnn quantnn math

Generated 2026-10-27T10:02:46.672Z

§1

The 168 Theorem

While developing Sounio's octonion backend, we found a combinatorial fact that does not appear to have been stated explicitly in the literature: the number of ordered triples (i, j, k) in {1,…,7}³ for which the octonion associator [ei, ej, ek] is nonzero is exactly 168 = |PSL(2,7)|. The decomposition reads 343 = 133 + 42 + 168.

The computation was performed in Sounio and independently reproduced in Python/NumPy. Verification repository: agourakis82/168-sedenion-bijection (14/14 pytest passing).

Paper: "The 168 Theorem: PSL(2,7) Governs Non-Associativity and Zero-Divisor Structure in the Cayley-Dickson Tower" — Agourakis & Gerenutti (2026). Submitted to Advances in Applied Clifford Algebras.

§2

Self-hosted bit-identical bootstrap

Sounio bootstraps from a 2,792-line C compiler (bootstrap/stage0.c) through a multi-stage chain into a self-hosted compiler. Stage 2 and Stage 3 produce byte-identical binaries — the canonical definition of a self-hosted fixed point. This was achieved on 2026-03-23.

Reproduce:

# Stage 0 → gen1 (from C compiler)
./artifacts/self-hosted/souc-self-hosted-x86_64 \
    self-hosted/compiler/lean_single.sio gen1.elf

# gen1 → gen2 (first self-hosted compile)
./gen1.elf self-hosted/compiler/lean_single.sio gen2.elf

# gen2 → gen3 (fixed-point check)
./gen2.elf self-hosted/compiler/lean_single.sio gen3.elf

# Verify
sha256sum gen2.elf gen3.elf

Expected: identical hashes. Recorded fixed-point SHA-256 (CHANGELOG): 7b588877f870e0011a9ace62eb12c7cd (741 KB). Cycle parity artifact: true.

§3

Compile-time rejection of under-confident dosing

ASHP 2020 §8.3 requires ε ≥ 0.82 for AUC-guided vancomycin dosing. The Sounio test tests/compile-fail/vancomycin_low_conf.sio attempts to pass a value with ε = 0.40 to a function with the signature fn prescribe_vancomycin(dose: Knowledge[f64, ε >= 0.82]). The compiler rejects the program. Not a lint, not a test suite — the type checker refuses to emit the program.

tests/compile-fail/vancomycin_low_conf.sio
//@ compile-fail
//@ error-pattern: ε

// Vancomycin prescription safety: compile-time ε-bound enforcement.
//
// ASHP 2020 §8.3 mandates ε >= 0.82 before AUC-guided dosing is permitted.
// A dose derived solely from Cockcroft-Gault (ε=0.40, CV ~60%) is too uncertain.
// The compiler must reject assignment to the prescription slot at compile time.

// Prescribing function: only accepts dose estimates with ε >= 0.82.
fn prescribe_vancomycin(dose: Knowledge[f64, ε >= 0.82]) with IO {
    println("Vancomycin prescribed")
}

fn main() with IO {
    // CrCl-derived dose: 40% confidence only (severe renal impairment + age 84,
    // CG formula CV >> 28%, multiple co-morbidities).
    let risky_dose: Knowledge[f64, ε=0.40] = Knowledge { value: 500.0, epsilon: 0.40 }

    // ERROR: ε=0.40 violates the ε >= 0.82 lower bound required by
    // prescribe_vancomycin.  The compiler must reject this call.
    prescribe_vancomycin(risky_dose)
}

Compiler output (captured from ./bin/souc check tests/compile-fail/vancomycin_low_conf.sio):

error: Type mismatch — expected confidence level not satisfied: Knowledge ε boundary violation at line 27

typecheck: failed

Exit code: 1. Output captured 2026-04-20. The gate criterion //@ error-pattern: ε is satisfied.

§4

Standard library test gate

The committed stdlib reliability gate reports stdlib reliability · not in CI · artifact 2026-05-12 · engine unnamed (source: artifacts/stdlib/stdlib_reliability_status.v1.json). If this page disagrees with README.md or the artifacts, those sources win.

$ bash scripts/stdlib_reliability_gate.sh

stdlib reliability · not in CI · artifact 2026-05-12 · engine unnamed
# numeral withheld — engine unnamed, not in CI
§5

Lean 4 verification of epistemic invariants

The formal/ directory contains Lean 4 proofs of the core invariants of the epistemic type system — in particular, that uncertainty propagation preserves monotonicity and that confidence gates are sound under composition. No sorry. No Mathlib.

Selected theorems

  • knowledgeLeq_trans — confidence ordering is transitive SounioEpistemic.lean
  • derive_decreases_confidence — derived knowledge is never more confident than its source SounioEpistemic.lean
  • combine_monotone — knowledge combination is monotone in confidence SounioEpistemic.lean
  • interval_monotone — uncertainty intervals grow monotonically under GUM propagation Epistemic.lean
  • composition_eps_bound — confidence gates are sound under sequential composition Epistemic.lean
  • confMeet_assoc, confJoin_comm — confidence lattice laws (meet/join associativity, commutativity, idempotence) SounioEpistemic.lean

If you can falsify any claim on this page, open an issue. The repository is public, the tests are runnable, the theorem is in submission. This page is under the same discipline as the language.