Σ-GLYPH

A portable check engine: send a computation by hash, get the same answer on any machine, with work and memory bounded up front.

result_hash = eval(term_hash, atp)     // deterministic, integer-only, total

Two independent machines given the same term_hash and the same budget return the same result_hash — bit for bit, with no shared runtime, no float, no clock, no network. If the budget runs out, that outcome is deterministic too. No input hangs it, and no input makes one implementation disagree with another.

Why anyone needs that. It is what lets you re-run a stranger's reason. Warrant records why an AI agent was allowed to do something; a reason there can be a Σ-GLYPH term, so a reviewer re-executes it on their own laptop and gets the same verdict — instead of trusting the log of whoever wrote it. Bounding work and memory up front is what makes running a stranger's computation safe by construction.

Three independent implementations agree on all 49 conformance vectors — Python, Rust and warrant-go, across serialization, byte-rejection and evaluation alike — and the evaluator's determinism, totality and memory bound are machine-checked theorems in Lean 4 (proofs/), not prose. A randomized differential fuzzer runs all three against each other on every push, and tools/x1_cross_repo.sh runs this repo against warrant's HEAD rather than a pinned snapshot.

python3 impl/sigma_glyph.py                        # Book I -> ALL PASS
python3 tests/spec_conformance/run_reference.py    # 49/49 vectors

Current: v0.6.6

Three Books (DRAFT STANDARD) + a constitution now at STANDARD. Hash-thunk evaluation with size-priced ATP (Book I), field-level wave pins with absent-wave semantics (Book II), selection-only annotation federation as a Warrant v0.3 profile (Book III: jurisdictions, machine-readable selection policies, permanent divergence by design), and governed Specification Anchors (spec/GOV-anchors.md v1.0.2, STANDARD: releases adopted by 2-of-3 threshold warrants, ADR-007; promoted DRAFT→STANDARD through a second 3-family gate). Machine-readable conformance suites for all Books (tests/spec_conformance/).

Beyond the compute core there is a second, optional layer: a wave/coordinate navigation view for finding related terms. It is strictly a view over identity and never part of it — ignore it entirely and Book I still holds. The names, the cosmology, and why FALSE sits at 270° live in spec/LORE.md: non-normative, and deliberately unhurried.

The Three Books

Core invariants, in one breath: hash is identity; phase is a coordinate; wave is a view; aggregate is never a field; ATP prices work AND memory (size − 1 ≤ spent); dead branches are never fetched; canonical failures are deterministic, local faults are not canonical.

Reference implementation

impl/sigma_glyph.py — Book I: serialization, validation, CAS, the v0.5 hash-thunk evaluator (lazy left-spine, size-priced ATP, genesis intrinsic), C1 λ→SKI compiler. impl/sigma_wave.py — Book II: arbiter-checked LUT, interfere() with entropy–coherence coupling. impl/sigma_federation.py + impl-go/ — Book III (Python oracle + independent Go). impl-rs/ — a third, independent Rust implementation of Book I (from-scratch SHA-256 + evaluator, no external crates) that replays the same oracle-generated vectors byte-exact.

python3 impl/sigma_glyph.py         # expected: ALL PASS (Book I)
python3 impl/sigma_wave.py          # expected: WAVE: ALL PASS (Book II)
python3 impl/sigma_federation.py    # expected: FEDERATION: ALL PASS (Book III)
(cd impl-rs && cargo build --release) && \
  ./impl-rs/target/release/book1 conformance tests/spec_conformance/vectors.json  # RUST-CONFORMANCE: ALL PASS (49/49)

Book I now has three independent implementations that agree on every vector — the Python oracle, warrant-go's native evaluator (via ski@v1), and Rust — plus a Lean 4 mechanization of the evaluator's determinism/totality and memory bound (proofs/EvalMachine.lean).

For AI reviewers

This repository is deliberately structured for multi-model review. If you are a model asked to critique this spec, read reviews/README.md first — it defines the protocol, and reviews/ contains prior reviews (Claude, Codex, Kimi) so you don't rediscover settled points.

Fastest way to be useful: run the reference implementation before critiquing the prose. Two of three prior reviewers filed "ambiguities" that were already resolved by executable test vectors.

Specification Anchors

Every published spec version is a citizen of its own system: SpecAnchor = NodeHash(LITERAL, atom = SHA-256(document_bytes)) — published detached in spec/ANCHORS.txt. A spec update is formally a fork with an explicit ancestor.

Provenance

archive/ preserves prior eras verbatim, including Era-1 (v0.2.12 "Titanium Monolith"), whose genesis forge method was reconstructed by brute force in 2026. Old hashes remain valid artifacts of their era. Dirty history is provenance, not shame.

License

MIT for the implementation; CC-BY-4.0 for the specification texts.

Part of the s0fractal mycelium federation (trinity / myc.md / OMEGA / liquid).