Gnosis Morph Bench

Live window into the Zer0pa lab. None of the work in this repo is a final release. Zer0pa GitHub repos are open as a matter of course for visibility and community participation. Always-in-beta is the posture: useful now, improving continuously. Open gates and honest blockers are surfaced in plain text below — see §5 "What We Don't Claim" and §11 "Upcoming Workstreams".

Licensing

This repository is part of the Zer0pa Gnosis Portfolio.

Code in this repository is licensed under the Apache License 2.0. See LICENSE for the full text. SPDX identifier: Apache-2.0.

Documentation, reports, and written materials are licensed under Creative Commons Attribution 4.0 International. SPDX identifier: CC-BY-4.0. Canonical terms: https://creativecommons.org/licenses/by/4.0/.

Data and fixtures are handled per dataset and artifact family. See DATA_POLICY.md for this repository's data boundary. The code license does not license raw corpora, image-bearing cultural-heritage assets, private HF artifacts, model weights, endpoint logs, or operational transcripts.

Trademarks - "Gnosis", "Zer0pa Gnosis", and distinctive sub-marks are trademarks of Zer0pa. Apache-2.0 and CC-BY-4.0 do not grant trademark rights. See TRADEMARKS.md.

Public visibility is a separate repository-setting action. The license files in this repo define the intended open-source/open-documentation terms for released Gnosis code and written materials; they do not publish rights-gated data.

What This Is

gnosis-morph-bench is the benchmark-first methods lane of the Gnosis extraction program. It owns neutral route scoring, permutation-null metrics, a five-mode stability battery, deterministic replay, and SHA-256 reference-freeze helpers — without owning Indus or cuneiform scientific verdicts. Domain corpora connect through admitted adapters; the Indus Phase 4 v1 adapter is the first in-scope family.

Headline metric (repo-custody, clone-reproducible): 37 passed pytest in a clean Python 3.11 venv + byte-equality cross-environment smoke (macOS 3.11.15 ↔ Linux/RunPod 3.11.13) + 0/6 forbidden-monorepo-pattern lint hits + 9/9 ADAPTER_CONTRACT_v1 MUST-clause coverage. Evidence path: artifacts/blind_clone/ and tests/. Live Indus Phase 4 measured values (NMI 0.5793, Sigma 5.65, Jaccard 0.4351, Replay 3/3) are source-authority citations only — see docs/PROMOTION_READINESS.md trust boundary.

Honest blocker: Live Indus Phase 4 reproduction is blocked on admitted Phase 3c manifest access (Blocked-1); heavy-data release policy for image-bearing assets is open (Blocked-3). The synthetic smoke is what runs end-to-end today. See ## Commercial Readiness below and docs/PROMOTION_READINESS.md.

Key Metrics

Metric	Value
pytest -q (clean Python 3.11 venv)	37 passed (local closeout verification, 2026-04-25)
GitHub Actions CI on main	pytest workflow configured at .github/workflows/ci.yml; latest status checked during closeout
Forbidden-monorepo-pattern lint hits	0 / 6 patterns across src/gnosis_morph_bench/
ADAPTER_CONTRACT_v1.md MUST-clause coverage	9 / 9
Committed smoke report SHA-256	020f97b83b2948c2cd529b975010e6e5132799d89e395539d6f6f928c97c184e
Cross-environment smoke byte-equality	PASS (macOS 3.11.15 ↔ Linux/RunPod 3.11.13)
Deterministic replay all_identical (synthetic fixture)	True (3/3)
Tracked-file leak scan (concrete RunPod IPs / SSH commands / local home paths)	0 hits

These are repo-truth metrics, not product or scientific claims. Live Indus Phase 4 measured values (NMI 0.5793, Sigma 5.65, Jaccard 0.4351, Replay 3/3) remain source-authority citations; see docs/PROMOTION_READINESS.md trust boundary.

What We Prove

• The neutral benchmark contract (BenchmarkManifest) round-trips cleanly through load_manifest / evaluate_route / freeze_reference / deterministic_replay / leave_fraction_out / the four added stability modes.
• The Indus Phase 4 v1 adapter satisfies every MUST clause of docs/family/ADAPTER_CONTRACT_v1.md, enforced by tests/test_adapter_contract_coverage.py.
• The forbidden-pattern lint catches the hidden-import / path-coupling patterns named by the adapter contract, self-tested via a positive control in tests/test_forbidden_patterns.py.
• The smoke and replay outputs are byte-identical across macOS and Linux at the documented Python 3.11 floor — verified end-to-end through fresh-clone install, see artifacts/blind_clone/.

What We Don't Claim

• That the live Indus Phase 4 measured values have been reproduced from repo custody. They have not. Live rerun is blocked on admitted Phase 3c manifest access (Blocked-1). Those values are source-authority citations, not Morph-Bench-proven values.
• That the synthetic smoke path constitutes proof of any live finding.
• That the heavy-data image-bearing asset release is unblocked. It is not; Blocked-3 (heavy-data release policy) remains open. See ## Commercial Readiness and docs/PROMOTION_READINESS.md.
• That the cuneiform benchmark family is in scope for the v1 adapter contract. It is explicitly deferred to a separate future contract once the first real Indus replay lands.
• That this repo carries any descriptor, kernel, image-preprocessing, or domain-verdict ownership. Those belong to other Gnosis lanes.

Commercial Readiness

Field	Value
Verdict	STAGED
Posture	benchmark_methods_active_with_named_blockers

The repo-local benchmark framework is usable today on synthetic fixtures with a deterministic, cross-environment, lint-clean surface. Named blockers are surfaced below per the always-in-beta posture — open work is in-progress, not a reason to withhold visibility. Canonical blocker list: docs/PROMOTION_READINESS.md.

• Blocked-1: admitted Indus Phase 3c feature manifest access.
• Blocked-2: CLOSED — Apache 2.0 LICENSE landed 2026-04-28. (Previously: canonical license identity OWNER_DEFERRED.)
• Blocked-3: heavy-data release policy for image-bearing assets — open.

Blocked-4 (smoke byte-reference re-freeze) closed on 2026-04-24.

Tests and Verification

Clean-venv local verification (Python 3.11):

python3.11 -m venv /tmp/gnosis-morph-closeout
source /tmp/gnosis-morph-closeout/bin/activate
python -m pip install -e '.[dev]'
pytest -q
git status --short

Expected: 37 passed, working tree clean. Test suite covers: adapter CLI surface (11), adapter contract coverage (3), forbidden-pattern lint with positive control (2), CLI smoke + Phase-4-style end-to-end (4), stability modes battery (4), replay record shape (3), HF SHA-pinned cache loader (5), plus 5 supporting suites.

CI: minimal pytest workflow at .github/workflows/ci.yml. No HF fetches, no blind-clone, no public uploads.

Blind-clone verification (cross-environment): see artifacts/blind_clone/03-01_transcript.md and .gpd/phases/03-blind-clone-and-promotion-review/03-01-SUMMARY.md.

Proof Anchors

Anchor	Path
Sovereign PRD	PRD_GNOSIS_MORPH_BENCH_2026-04-23.md
Adapter contract v1 (Indus Phase 4)	docs/family/ADAPTER_CONTRACT_v1.md
Stability battery v1	docs/family/STABILITY_BATTERY_v1.md
First replay target decision	docs/family/FIRST_REPLAY_TARGET_DECISION.md
Promotion readiness (named blockers)	docs/PROMOTION_READINESS.md
HF custody register	docs/HF_CUSTODY_REGISTER.md
Legal review prep (questions for Zer0pa legal)	docs/LEGAL_REVIEW_PREP.md
HF storage strategy	docs/HF_STORAGE.md
Status report (2026-04-24)	docs/STATUS_REPORT_2026-04-24.md
Live Phase 4 source-authority boundary	docs/PROMOTION_READINESS.md, docs/STATUS_REPORT_2026-04-24.md
Forbidden-pattern lint evidence	tests/test_forbidden_patterns.py, artifacts/blind_clone/03-01_forbidden_pattern_scan.txt
Public-audit boundaries	PUBLIC_AUDIT_LIMITS.md
Blind-clone transcript (macOS + Linux)	artifacts/blind_clone/

Repo Shape

.
├── NOTICE                          # Apache-2.0 code and CC-BY-4.0 docs posture
├── PRD_GNOSIS_MORPH_BENCH_2026-04-23.md
├── README.md
├── pyproject.toml                     # Python ≥ 3.10, deps: numpy, scikit-learn
├── .github/workflows/ci.yml           # pytest on Ubuntu / Python 3.11
├── .gpd/                              # GPD orchestration state, plans, summaries
├── src/gnosis_morph_bench/
│   ├── adapters/indus_phase4.py       # v1 adapter (CLI: gnosis-morph-bench-adapter-indus-phase4)
│   ├── benchmark.py                   # route scoring, NMI + null + sigma + silhouette
│   ├── stability.py                   # 5-mode stability battery
│   ├── replay.py                      # ReplayRecord emission
│   ├── schema.py                      # BenchmarkManifest, freeze_reference
│   ├── cli.py / __main__.py           # smoke + replay subcommands
│   ├── hf_cache.py                    # SHA-pinned HF dataset fetch
│   └── _utils.py                      # repo-local JSON / hash helpers
├── tests/                             # 37 tests across 9 files
├── fixtures/tiny_benchmark_manifest.json
├── artifacts/
│   ├── smoke/smoke_report.json        # committed byte-reference (SHA-pinned)
│   ├── replay/                        # default replay output dir (gitignored content)
│   └── blind_clone/                   # cross-env blind-clone transcripts
├── docs/                              # promotion readiness, HF custody, legal prep, family contracts
└── code/                              # code-facing surface notes and current limits

Quick Start

git clone https://github.com/Zer0pa/Morph-Bench.git
cd Morph-Bench
python3.11 -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'

# repo-local pytest suite
pytest -q

# synthetic smoke (deterministic, byte-stable)
python -m gnosis_morph_bench smoke fixtures/tiny_benchmark_manifest.json \
    --output artifacts/smoke/smoke_report.json

# full 5-mode stability + replay
python -m gnosis_morph_bench replay fixtures/tiny_benchmark_manifest.json \
    --output artifacts/replay/replay_record.json

# Indus Phase 4 v1 adapter (against fixture inputs only until Blocked-1 clears)
gnosis-morph-bench-adapter-indus-phase4 --help

For the full audit walkthrough: AUDITOR_PLAYBOOK.md.

For what is and is not promotable today: docs/PROMOTION_READINESS.md.

For governance and status semantics: GOVERNANCE.md.

For release protocol and owner inputs: RELEASING.md.

Upcoming Workstreams

Open work across four categories. "Useful now, improving continuously" — every item here is surfaced honestly, not an apology.

Operations / External Dependency

• Blocked-1 — Phase 3c manifest admission: live Indus Phase 4 rerun from repo custody cannot proceed without an admitted Phase 3c feature manifest JSON with SHA-256. Owner-gated. When clear: produces artifacts/replay/indus_phase4_live_<date>.json.
• Blocked-3 — Heavy-data release policy: image-bearing benchmark payloads cannot enter the repo until the release boundary is written (which asset classes, under what terms, storage location and size bound). Owner-gated. When clear: produces DATA_POLICY.md §Image-Bearing Release appendix.

Active Engineering

• Cuneiform benchmark family adapter: deferred to a separate future contract per current scope decision; no active engineering until Blocked-1 live rerun lands and the v1 adapter is validated against a second source family.
• Smoke byte-reference re-freeze: local housekeeping only; re-run smoke and commit the new byte-reference when convenient. Not a promotion gate.

Research-Deferred — Investigation Underway

(none currently — all research surface is defined by the v1 adapter contract and the 5-mode stability battery)

Zero-Base Scientific Thinking — GPD Research and Planning Pending

(none currently — the Phase 04 cuneiform contract is the next planned GPD unit; not yet initiated)