szl-typesafe-triage

A deterministic, receipt-bearing triage engine, and an honest account of what it cannot do.

Read this first. The engine does not yet perform triage by meaning; it matches vocabulary. The useful content of this repository is the measurement discipline: 16 retractions recorded in an append-only ledger, a phrasing guard that mirrors the estate's CI rule, and a leakage gate that refuses rather than advises. Release is BLOCKED at 11/12.

Measured, not asserted

What is not claimed

• No winner, no baseline beaten, no comparison against any other system.
• No kernel verification is performed here; Lean symbols are bound by name only (out/lean_binding.json).
• Lambda is cited as Conjecture 1 and disproved as stated; conditional Theorem U is the proved result.
• No locked-formula count is asserted; the estate's own sources disagree (out/phrasing_guard.json).
• SLSA L1 honest, L2 roadmap. Not L2-verified, not L3, and no federal or hardened-image accreditation.

Audits in this repository

• docs/ESTATE_AUDIT.md - ten passes over 88 public estate repositories
• docs/PRIVATE_AUDIT.md - private-tier leads via authenticated code search
• docs/DEEP_FINDINGS.md - why the predecessor adapter is NOT PROMOTABLE
• docs/STATE_OF_THE_REPO.md - eight bandaids and six gaps, named
• docs/DATASET_CARD.md - contamination field filled, per the org template

Apache-2.0. Copyright 2026 SZL Holdings.

Use the local runtime

szl-triage serve --port 8765 starts a loopback console and JSON API around the deterministic pipeline. The installed package carries its default policy, so the CLI works outside this checkout. Decisions include verifiable payload, policy, and implementation hashes. See local runtime for installation, API requests, and the separate fresh GPU challenge command. The model-release block above remains in force.

Deploy the public deterministic software lab

The Docker Space entry point is python -m szl_triage.public_server, explicitly bound to 0.0.0.0:7860. It uses only the standard-library deterministic pipeline. /readyz and /v1/decide identify the runtime as DETERMINISTIC_SOFTWARE_LAB, report model_loaded: false, and retain model promotion as HOLD. Publishing this software lab does not qualify or publish a language model.

Before startup, the publisher must create SOURCE_BINDING.json from the immutable canonical GitHub commit. Its exact fields are schema (szl.triage-source-binding/v1), github_repository (szl-holdings/szl-typesafe-triage), github_commit (the full 40-character commit SHA), and source_files (relative paths mapped to SHA-256 digests). The file set must contain every deployed src/szl_triage/**/*.py file and src/szl_triage/data/triage_policy.v3.json, with no extra declarations. Startup rejects absent, changed, omitted, extra, or symlinked source files before binding. Readiness reports the declared GitHub commit and verified manifest digest; these unsigned digests bind contents and do not authenticate the publisher.

TRIAGE_TRUSTED_AUTHORITIES is a required comma-separated list of exact lowercase DNS authorities. The Docker image explicitly configures szlholdings-szl-typesafe-triage.hf.space. A domain proxy must preserve one configured authority or add its exact authority to this setting. Host wildcards and forwarded host trust are unavailable. An Origin, when supplied, must be HTTPS with the same authority as the request Host. Duplicate headers and cross-origin requests are rejected. TLS terminates at the managed hosting proxy.

TRIAGE_ALLOW_LOOPBACK_PROBES=1 explicitly permits only local-peer probes with localhost:7860 or 127.0.0.1:7860. Its default is 0; the Docker image enables it for managed container health checks. Framing is disabled by default. The image explicitly sets TRIAGE_FRAME_ANCESTOR=https://huggingface.co to permit the managed Hugging Face page to embed the lab while retaining nonce-based scripts and styles. No other frame ancestor or wildcard is accepted. The image copies only Python package source, the default policy, and the source binding; weights, training outputs, and credentials are outside its copy set.

Replay the five-seed study

python scripts/codex_finish.py audit checks the saved study offline, including raw predictions, recomputed metrics, split hashes, and the exact five seeds. It reports integrity separately from promotion. A bounded GPU smoke test can write a new receipt without replacing the historical results. See study replay and inference for commands and limitations.