Find evidence first. Spend computation second. Let physics decide.
VESPER (named for the evening star) is a research program that tested — in a pre-registered, hash-sealed, single-evaluation experiment — whether evidence-first routing can reduce the computational cost of exoplanet transit detection without sacrificing recall. Instead of folding every star's light curve at thousands of trial periods (BLS/TLS), it detects individual transit-like events directly, infers the orbital period from their spacing, gates candidates with a bootstrap-calibrated period false-alarm probability, confirms with a transit-model likelihood ratio, and reserves the full periodogram search for stars showing no local evidence.
Status (2026-07-20): Phase I (TESS, Sectors 1–3) is complete and sealed; an
independent audit (2026-07-19) found the sealed compute verdict was not produced by the
frozen measurement rule — "H1 FALSIFIED (compute branch)" is withdrawn (decision
record DR-003) and the frozen-rule
re-measurement is in progress on branch phase1/audit-remediation. The recall result
(E1) survived the audit and is robust. Phase II has been re-scoped away from Kepler
routing-scaling — see Where this is going below.
- Recall non-inferiority (E1): PASS, robust. On 15,000 sealed-test injections the occurrence-weighted recall difference vs full TLS is −0.48 pp; the one-sided 95% lower bound clears the pre-registered −2 pp margin under all three interval constructions, including a host-cluster bootstrap (−0.82 pp).
- Scoped compute (E2): the originally recorded "FAIL (24.4% < 30%)" rested on a
12-star timing subset in deviation from the frozen measurement rule and was
statistically undecided (ratio CI [0.42, 1.14]). It was re-measured under the frozen
rule in the audit remediation — see
research/m4_evaluation/M4_ERRATUM_2026-07-19.mdfor the corrected verdict and the full deviations register. - Integrity: thresholds and protocol were sealed (git tags
phase1-prereg-v2/v3) before the single test read; the test split was read exactly once; sealed documents are byte-identical to their tags modulo the TRINETRA-X→VESPER rebrand strings.
A full technical audit (docs/audits/PROJECT_AUDIT_2026-07-19.md)
and an idea-level scientific review (docs/reviews/DEEP_SCIENTIFIC_REVIEW_2026-07-19.md)
concluded that per-star routing cannot materially cut survey-scale transit-search
compute for any router of this class — occurrence is dominated by planets below
single-event visibility (measured break-even prevalence π* ≈ 0.68 vs TESS-realistic
≈ 0.03). Phase II is therefore re-scoped (docs/VESPER_PHASE2_PROGRAM.md,
draft pending owner adoption as DR-004):
- Track A — Theory: formalize the triage impossibility bound, with the sealed Phase-I run as its empirical witness.
- Track B — Infrastructure: release the sealed injection-recovery machinery as VESPER-Bench — a pre-registered, leakage-safe benchmark for transit-search pipelines (Kepler DR25 extension).
- Track C — Science: an event-wise monotransit detection pipeline — the K=1 regime where fold-based search provably cannot operate and photometric physics is structurally forced to be the arbiter.
- G0 (gating experiment): a fast-folding / single-event-statistic search is priced against TLS first, with pre-registered decision rules.
Near-term execution (verdict completion, robustness sensitivities, engineering
hardening, publication) is governed by docs/ROADMAP_TO_10.md.
The original Kepler routing-scaling sketch is superseded and archived on its branch.
docs/VESPER.md— master charter.docs/SCIENTIFIC_HYPOTHESIS.md— the falsifiable claims (sealed).docs/VESPER_PHASE1_VALIDATION.md— the pre-registered protocol (sealed).research/m4_evaluation/M4_TEST_RESULT.md— the sealed result + addendum.research/m4_evaluation/M4_ERRATUM_2026-07-19.md— corrections, deviations register, re-measured E2.docs/audits/PROJECT_AUDIT_2026-07-19.md— full second-pass audit.docs/VESPER_PHASE2_PROGRAM.md— the re-scoped future.papers/phase1_evidence_first_triage.md— manuscript draft.
| Path | Contents |
|---|---|
docs/ |
Canonical specs + theory; decisions/ DR-001…DR-003; audits/ + reviews/ (2026-07-19 audit & review); Phase-II program + roadmap |
research/ |
Milestone tooling M0–M6 (m0_manifest … m6_reality_check), Phase-I plans (phase1/) |
data/manifests/ |
Sealed manifests, thresholds, and the single-test artifacts (tracked); light-curve caches are gitignored |
papers/ |
Manuscript draft |
tests/ |
Fast unit tests (run in CI) |
hackathon/ |
BAH 2026 PS7 track (separate, submitted 2026-07-01) |
archive/ |
Prior-project audit (historical; do not modify) |
vault/ |
Obsidian research memory (mirrors the repo; repo is authoritative) |
python -m venv .venv && .venv/bin/pip install -r research/m4_evaluation/requirements.txt
.venv/bin/python -m pytest tests/ -q # fast unit tests
.venv/bin/python research/m4_evaluation/e1_corrected_inference.py # E1 re-analysis (needs sealed artifacts)Light curves are fetched from MAST (TESS SPOC 2-min); the sealed target manifest and
all thresholds are content-hashed in data/manifests/ (verify with shasum -a 256,
noting the rebrand caveat in docs/decisions/F1_DECISION_RECORD.md §5a).
MIT — see LICENSE. Charter author: Ansul Suryawanshi.
Math in these documents uses LaTeX ($…$); view in a math-aware renderer (Obsidian, VS Code, GitHub).