Behavioral SETI — analysis code

Code and intermediate results for A combination-space search for embedded technosignatures in solar and heliospheric archives (R. Griffin, Dxtra Inc.) and its companion, Set and forget: initial conditions as a technosignature.

Every number in the papers is produced by a script here. The table below is the audit trail: paper claim → script → result file. Nothing is quoted in the papers that is not regenerable.

Requirements

pip install -r requirements.txt        # numpy, scipy, pyhdf

pyhdf is needed only for ACE SWICS, whose level-2 products are HDF4 Vdata tables. Runtimes below are for 80 cores.

Data

No data is committed. Everything is public; acquire/ fetches it.

python acquire/pull_chan.py            # ACE SWICS, sunspot area, WSO, Ca II K, LASCO
python acquire/getpsr2.py              # ATNF psrcat -> psrcat.json

GOES-R EXIS, OMNI2, SILSO, F10.7, NMDB and LISIRD products are listed in the paper's data availability statement and were retrieved from the archives named there.

Where it is written

Downloads, caches and intermediate products go to $BEACON_DATA, which defaults to $HOME if unset:

export BEACON_DATA=/srv/beacon        # optional; defaults to $HOME
python acquire/wind_fetch.py

Python scripts read it as

DATA = os.environ.get("BEACON_DATA") or os.path.expanduser("~")

and the shell ones as ${BEACON_DATA:-$HOME}. Code in this repository imports from this repository: nothing resolves a module through an absolute path outside it.

Audit trail

Paper claim	Script	Result
§3.3, §4.8 — 13-channel sweep, 229 tests	search/sweep13.py	results/sweep13.json
§4.8 — 30-channel sweep, 1,074 tests, 22 survivors, 0 of 510 across the boundary	search/sweep30.py	results/sweep30.json
§5.2 — tail-fitted null, pair space	search/sweepT.py --mode pairs --shifts 10000	results/sweepT_pairs_zeus.json
§5.2 — tail fit validated against counting	validate/validate.py	printed
§4.9 — triple sweep, 10,996 tests, 198 screened	search/sweepT.py --mode triples	results/sweepT_triples.json
§4.9 — direct-count confirmation, 93 of 198	search/confirm.py --shifts 500000	results/confirm_triples.json
§4.9 — form-level control that voids the sweep	validate/fptest.py	printed
§4.2 — fast-band limits	search/fastsearch.py, search/euvsfast2.py, search/nmfast3.py	printed
§4.2, Fig 2 — injection recovery curves	validate/inject2.py	results/injection.json
§4.6, Fig 3 — p-mode comb, 135.1 / 135.0 µHz	search/pmode.py, search/pmode_cutoff.py	—
Fig 3 — comb curves for plotting	figures/figdata3.py	results/figdata_comb.json
§5.7 — 53 unreachable pairs, all permanently blocked	validate/span30.py	results/reach30.json
§4.10 — self-keyed excess, ACE ρ = −3.52, Wind ρ = −3.07, rotation-matched null	validate/selfkey_replicate.py	results/selfkey_replicate.json
§4.10 — power by half, 100% / 94% at 0.10σ	validate/selfkey_power.py	results/selfkey_power.json
§4.11 — annual term, decoy-offset null, 8 channels	search/geoterm.py	results/geoterm.json
§4.12 — viewpoint control, lag RMS 1.97 d vs null 7.65 d, p = 5×10⁻⁴	search/viewpoint_l2b.py	results/viewpoint_l2b.json
§4.12 — slow-band limits, 0.02σ at 3 d	validate/viewpoint_limit.py	results/viewpoint_limit.json
§4.12 — fast band, 0 Earth peaks, 561 Mars	search/viewpoint_fast.py	results/viewpoint_fast.json
§4.12 — fast-band limits, Earth 7.2×10⁻⁶ / Mars 4.2×10⁻³ at 307 s	validate/viewpoint_fast_limit.py	results/vpfl.log
§4.13 — LLR inter-reflector, 24 of 29 series, all near-survivors lunar harmonics	llr/llr_interref.py	llr/llr_interref.json
§4.13 — LLR within-session, p = 0.114, residual 244 ps	llr/llr_model_free.py	llr/llr_mf2.log
§4.13, §2.5 — LLR cannot reach the modulator: 2.2×10¹¹ short	llr/llr_reach.py	printed
§4.14 — RSTN, 44 / 64 / 43 peaks, twin gate	search/rstn_search.py	results/rstn_search.json
§4.14 — the 5.9 s power is per-channel, not in the ratio	validate/rstn_5s.py	results/rstn_5s.log
Companion §3.1 — population back-extrapolation	search/backex2.py	printed
Companion §3.2 — subset ladder search	search/subset.py	printed
Companion §3.3 — Shklovskii + Galactic correction	validate/shk2.py	results/psrcat_shk.json
Companion §3.3 — subset search on corrected Ṗ	search/subset_c.py	printed

Audit, run against this commit

validate/audit.py checks every number the manuscript quotes against the result file that produced it. A claim passes only if the value is found in the repository and matches the manuscript; "not regenerable" is a failure, not a note.

22 claims checked, 22 reconcile. Three needed investigation and all three resolved:

• sweep30.json is a JSON list of 1,074 entries, not a dict — the count is right, the first audit script read it wrongly. The extractor is fixed, so the audit now reports 22 of 22 rather than three false failures; an audit that cries wolf is not an audit.
• reach30.json carries perm (53 permanently blocked) and ok (382 reachable), which is what §5.7 and §4.8 quote.
• The triple count was genuinely wrong in the manuscript. sweepT_triples.json holds 2,749 unordered triples at exactly 4 tests each = 10,996. The paper said 2,934 triples were attempted; 2,934 is the number reachable (2,934 × 4 = 11,736, also quoted). Only 2,749 completed. The manuscript now distinguishes the two.

The two checks that run from a clean clone

git clone <repo> && cd <clone>
python3 search/confirm.py --shifts 500
  -> REFUSING: floor is not below the threshold; raise --shifts     (expected)
python3 validate/validate.py results/sweep30.json results/sweepT_pairs_zeus.json
  -> === PASS - tail fit validated, proceed to triples ===          (exit 0)

Both pass. Note the second takes two arguments, counted then tail-fitted; an earlier version of this README implied it ran bare, and it does not.

Two things a reader should check first

The p-floor against the significance threshold. Counting exceedances among M circular shifts floors p at 1/(M+1). If that floor sits above α/N, no test can survive and a clean-looking null means nothing. This happened here: the 30-channel sweep was first run at 6,000 shifts against a threshold of 4.66×10⁻⁵ and returned "0 of 1,074 survive", which is reported in §5.2 as void. confirm.py refuses to run when the floor is inadequate.

Whether a detector can fail. Three statistics in this work could not, and were caught only by injection or by a form-level control: a log-ratio autocorrelation that sat at 0.954 for data and surrogates alike; a residual-product form sitting 4.5σ below its own null; and a phase-randomised surrogate used against a statistic computed from the power spectrum it preserves. validate/fptest.py is the template for testing a new form before trusting it.

Known corrections

Recorded because they are load-bearing, and because the papers report them rather than hiding them.

• The §4.2 limits were originally described as 99%-confidence figures established by injection. They are analytic threshold-crossing amplitudes, recovered about half the time, and no injection had been run. validate/inject2.py supplies the real curves; the cosmic-ray limit holds up at 95% recovery, Lyman-α and soft X-ray recover 1% and 16% at their quoted values.
• The continuum-normalised power is not exponentially distributed (mean 1.44, tail ~4×10³ heavier than Exp(1)), so the analytic threshold is itself optimistic.
• A re-implementation of the mode-comb detector returned 127 µHz where the original gives 135, from a 4001-minute median detrend and a linear continuum in place of a 60-minute mean and a log-space one. Use search/pmode_cutoff.py; figures/figdata3.py reproduces it with a strided median for speed.
• Mesh numerics in an early draft did not cohere; they are now derived from the link budget.
• A cost estimate was wrong by eight orders of magnitude by costing probe buses rather than modulators.

Licence

Code (acquire/, search/, validate/, figures/) — MIT, see LICENSE. Derived results (results/) — CC BY 4.0, see LICENSE-DATA.

The underlying observations are not relicensed: they belong to NOAA, NASA, LASP, SILSO, NMDB, ATNF and the other archives listed in LICENSE-DATA, and should be cited to them. No observational data is committed to this repository.

Software

Written and run with Claude Code (Anthropic). Scope and documented failures are set out in the paper's Software section. Responsibility for every number rests with the author.

Manuscript changelog

Moved here from the manuscript at the reviewer's request: a changelog belongs with the repository, not in a paper submitted to a journal.

Round 4

Changed since the last review. §3.6 is new and is the largest addition: a designer-priority table ranking fifteen channels by what a sender would choose, against what we can resolve and whether we have searched there. It sits before the Results deliberately, so its ordering is a prediction rather than a description of what was found, and §1.4 now carries its uncomfortable consequence — that rows 2–7 are capable and unsearched, so this paper has not yet tested the framework's own best guess. §4.12 gained the cross-viewpoint test as the 24th search and has since been extended to the fast band on 1,879 days of MAVEN L2: null, with the binding limit set by MAVEN rather than GOES and 580× weaker, and with the band capped at 6 h because Earth and Mars see solar rotation at different synodic periods. §4.13 reports lunar laser ranging outside the tally, with the reason — that no dynamical method can reach a 2.1×109 kg modulator — stated before the nulls. §2.4 gained the neutrino gate as content behind a later gate and §2.5 the coupling calculation behind §4.13. §5.5 proposes an untried neutrino-line search. Five references added, all verified against ADS and the publishers; the existing Learned et al. 2008 entry is a different paper from Learned, Pakvasa & Zee 2009. Earlier in this round: §4.10 replaced the note that dismissed the self-keyed excess on the sign of its correlation, §4.11 answered the annual-geometric-term objection, the stale 3.7% false alarm became the measured 6.3%, and the outcome table gained the false-alarm-and-power row it had been missing.