Axiom-ZSpace

Axiom-ZSpace runs a light curve through a single deterministic pipeline:

ingestion → BLS detection (period-prior ladder, harmonic rejection, FAP
firewall) → ephemeris resolution → transit-physics audits → context/density
checks → false-positive gate engine (circuit breaker) → CVS classification →
discovery card with a full proof chain

Every tunable number lives in one place — config/production.yaml + zspace_engine/thresholds.py — and every number is measured: threshold changes must be re-benchmarked before shipping (the one rule, enforced in docs/CONTRIBUTING.md).

• Blind search with a prior, not a lock: a period hint only wins when its harmonic-ladder peak carries enough power; otherwise the global BLS peak wins.
• Physics audits: even/odd depth (Welch t-test), U/V shape ratio, per-transit depth consistency, ingress/egress, secondary eclipse, stellar-density mismatch, centroid (optional), MCMC posterior (optional).
• Proof chains: every card records why each gate passed or failed.
• Circuit breaker: a failed critical gate makes SOVEREIGN_PASS impossible — covered by tests.
• One catalog: three profiles (conservative / balanced / sensitive) with an auto-generated reference (THRESHOLDS_REPORT.md).
• Deterministic: same seed gives same results, checked by tests/test_reproducibility.py.
• C99 build included: the same logic ships as a dependency-free C99 binary (kernels generated from NumPy by Purce, 148/148 validator kernels, 90/90 cards). Use --engine {python,c99} — python by default, c99 for batch (42.8 ms/TIC light at 650×, 4.8 s/TIC heavy at 5.8×, 16 cores; see docs/BENCHMARKS.md:151).
• Offline-first: engine and test suite run without network; the real benchmark needs one MAST download, then works from pinned snapshots.

* Heavy Python ~27.8 s is a placeholder copied from light (not re-measured for 87k; expect higher). † Heavy 8/10 — two marginal FAP≈0.05 flips (fap 0.008→0.078, alias).

Recommendation: python for reference and single targets, c99 for batch (--engine c99) — docs/BENCHMARKS.md:151, docs/C99_ENGINE.md:18. Per-core: light 40.6×, heavy 0.36×.

FPR honesty: 0/80 is the fixed-seed sample; across 8 fresh samples (BIG400, 800 targets) the measured contamination FPR is 4.25%, dominated by high-SNR eclipsing binaries that pass every gate (pure noise: 0.6%). Both numbers are committed evidence — see docs/BENCHMARKS.md §1.6.

Evidence runs are versioned in the repo: benchmarks_controlled/evidence/{OVERHARM_FIX2,BIG400,PROBE_FPR68*} and benchmarks_real/evidence/REAL_FINAL — per-target JSONs, recomputable metrics, and the reports that derive from them. How to reproduce everything in two commands: docs/BENCHMARKS.md.

Re-run both benchmarks after any threshold edit — the numbers above only describe the shipped configuration.

pip install -r requirements.txt          # Python ≥ 3.10

python run_pipeline.py --synthetic       # offline self-test (no internet)
python run_pipeline.py --synthetic --engine c99   # same, C99 sovereign engine
python -m pytest tests/ -q               # 101 tests, ~5 min, offline-green

python run_pipeline.py --tic 260128333   # single TIC, end-to-end
python run_pipeline.py --sector 42       # whole sector (heavy; see PRODUCTION.md)

CLI: --tic ID · --synthetic · --engine {python,c99} · --sector N / --sectors A-B · --max-targets N · --output DIR · --mcmc · --tpf-centroids · --multi-sector (full reference: docs/QUICKSTART.md).

run_pipeline.py          CLI entry point (+ c99_bridge.py — Python↔C99 adapter)
zspace_engine/           the engine: ingestion, detectors, ephemeris,
                         auditors, context, validator, core, report,
                         thresholds, sector processor, output organizer
C99-Version/             the C99 sovereign engine (Purce-generated kernels,
                         zspace_core, zspace_card CLI, verification harnesses)
benchmarks_controlled/   synthetic benchmark (injector + runner) + evidence
benchmarks_real/         real Kepler benchmark (pinned NEA snapshots) + evidence
tests/                   the 101-test suite (+ validator demo helper)
scripts/                 snapshot fetcher, sector-aggregation utility
scripts/legacy/          pre-v1 tools (C engine, old harnesses) — NOT part of v1
config/                  canonical configuration (threshold catalog)
docs/                    this documentation set
archive/                 local-only pre-v1 history — git-ignored, never pushed

• Numbers trace to artifacts. Every figure in the docs comes from a committed file; CHANGELOG v2.x entries are marked legacy because their batch numbers were not preserved.
• Fixed samples. The synthetic benchmark is seed-fixed (--seed 20260814); the real sample is pinned by offline NEA snapshots (scripts/fetch_nea_snapshot.py). Same command, same result — checked by tests.
• Git-ignored by rule. archive/, run outputs, caches, discovery cards, and anything credential-bearing never get committed.

Read docs/CONTRIBUTING.md first — especially the "one rule": no threshold change ships without both benchmarks re-measured, and no number ships without an artifact.

_{Blind transit search with reproducible numbers. The evidence files say where each figure comes from.}