Pharos v0.1.0 — IR / Astrometric Confounder Benchmark

Pharos v0.1.0 — IR / Astrometric Confounder Benchmark

Released: 2026-05-11
Pre-registration: pre_registration/v0.1_ir_benchmark.md — FROZEN 2026-05-11, SHA-256 0784859d8cad230d98a43c14e84862396855f590586dd72b29c7e2676963d38c

What v0.1 establishes

Pharos v0.1 is the wedge release of the Technosignature Lead Graph: an end-to-end pipeline that takes Gaia DR3 + AllWISE + 2MASS archival data and produces a confounder-penalised re-observation ranking within the WISE-only coverage class. The release demonstrates that the system can demote seductive false positives before promoting anything.

This is not a technosignature detection. It is a calibrated archival ranking instrument with its first scientific gate passed.

Pass-criterion results

The pre-registered v0.1 benchmark required correct classification of the seven Suazo et al. (2024) "Project Hephaistos II" candidates. All seven cases pass against live Gaia archive data:

Candidate	Required class (pre-reg §7)	Pharos v0.1 output
G (confirmed background AGN)	discard_confounded	discard_confounded (β·P = 4.69) ✓
B (suspected, W1↔W3 offset 3.21″)	discard_confounded or needs_human_review	discard_confounded (β·P = 5.18) ✓
A (radio-archival contamination)	Not ir_followup	no_action ✓
C, D, E, F (unresolved)	Not ir_followup	All no_action or discard_confounded ✓

Candidate G is demoted purely on the W1↔W3 photocentre offset (5.59″ RA per Suazo et al. Table 7) — the same signal that high-resolution e-MERLIN/EVN follow-up later confirmed as background AGN contamination. The confounder model mechanises the original-paper internal flag.

Synthetic-injection recovery (§7 secondary)

W3-excess injections at 5σ / 10σ / 20σ into a held-out stratified pool drawn from the quiet-negative population:

σ	n_injected	n_recovered	rate	median IR evidence
5	61	1	1.6%	1.12
10	61	32	52.5%	2.14
20	61	61	100.0%	2.75

Recovery is defined as ir_evidence ≥ 2.0 (the needs_human_review evidence floor, pre-reg §8). Monotonic with σ, 100% at 20σ — the calibration is correct. The lower 5σ recovery rate is set by the empirical-p-value tail depth of the 1,225-source locus population and is the expected behaviour.

FDR calibration on null sample (§7 secondary)

BH q-value rejection counts on a 612-source held-out null sample drawn from the quiet-negative population:

q threshold	n_total	n_rejections	rate
0.01	612	0	0.0%
0.05	612	0	0.0%
0.10	612	0	0.0%

Zero rejections at all q thresholds is the correct conservative behaviour for a pure-null sample. BH does not falsely promote control sources.

Data products

All artifacts in this release are CC BY 4.0 (see LICENSE-DATA). Code is Apache 2.0 (see LICENSE).

Artifact	Path	Description
Quiet-negative controls	controls/cache/quiet_negative.parquet	2,451 Gaia + AllWISE + 2MASS sources
Hephaistos join	controls/cache/hephaistos_join.parquet	7 candidates with Suazo Table 5 / 7 fields
Injection recovery	controls/synthetic_injection/	Per-injection rows + per-σ summary
FDR report	controls/fdr_report/	Per-target FDR rows + per-q summary
Hephaistos registry	controls/hephaistos.yaml	The seven candidates and their labels

Code modules

Module	Purpose
pharos.sources	Gaia archive TAP client; split-query path; OAuth-style auth
pharos.ir_sed	Data-driven stellar-locus residual scoring
pharos.confounders	P_IR confounder vector (incl. P_wise_offset)
pharos.calibrate	Empirical p-values + Benjamini-Hochberg FDR
pharos.injection	Synthetic-injection recovery harness
pharos.fdr_report	BH calibration on held-out null samples
pharos.cli	CLI: fetch_quiet_negative, fetch_hephaistos, run_injection_recovery, run_fdr_report

What v0.1 does NOT do

• Does not score time-domain photometry, spectral residuals, optical pulse, or radio archives. These are the v0.2 / v0.3 / v0.4 layers (see manuscript §9).
• Does not run an unWISE re-extraction pipeline; W1↔W3 photocentre offsets for the seven Hephaistos candidates are loaded from Suazo et al. (2024) Table 7. General-population offset measurement is deferred to v0.2.
• Does not calibrate β weights — the v0.1 weights are pre-registered initialisation values. β calibration against the contaminant-positive set is scheduled for v0.1.1.

Reproduce v0.1

# Authenticate to the Gaia archive (one-time setup):
mkdir -p ~/.config/pharos && chmod 700 ~/.config/pharos
# Place Cosmos username on line 1 and password on line 2 of:
#   ~/.config/pharos/gaia_credentials.txt
chmod 600 ~/.config/pharos/gaia_credentials.txt

# Install in a venv:
python -m venv .venv && source .venv/bin/activate
pip install -e .[dev]

# Fetch data and run benchmarks:
python -m pharos.cli fetch_hephaistos --output controls/cache/hephaistos_join.parquet
python -m pharos.cli fetch_quiet_negative --output controls/cache/quiet_negative.parquet --limit 5000
python -m pharos.cli run_injection_recovery --max-pool-size 150
python -m pharos.cli run_fdr_report

# Verify all benchmarks pass:
pytest

Citation

A Zenodo DOI for this release will be minted via the GitHub-Zenodo integration on tag publication. Cite this software release alongside the methodology manuscript.

Next

• v0.1.1: β-weight calibration against the contaminant-positive control set; HOT DOG X-ray-null cross-match.
• v0.2: Time-domain integration (TESS, Kepler, ZTF). Per-transit residual scoring.
• v0.3: Radio coverage (Breakthrough Listen, COSMIC). Enables full demotion of Hephaistos candidate A.
• v0.4: Spectral residual integration.
• v1.0: Public cross-modal leaderboard.