ledger-grok

SAPIENS LEDGER demo — Kepler-10 b re-derivation

What this is: a public validation demo of the SAPIENS hash-chained evidence ledger. It re-derives the already known exoplanet Kepler-10 b from a small public Kepler light curve and records every step — data hash, hypothesis, analysis, adversarial challenge, verdict — in a tamper-evident JSONL chain.

This is a re-derivation of a published result. It is not a discovery, and nothing here claims one. Kepler-10 b was discovered by Batalha et al. 2011; we use the published ephemeris purely as the answer key for validation.

Naming and endorsement

"LEDGER" is a SAPIENS demo. The hypothesis / adversarial-challenge steps sit behind a small adapter interface whose optional live backend is Grok (xAI). Grok and xAI are named only to identify that API; no affiliation, sponsorship, or endorsement by xAI is implied. The committed demo run used the deterministic offline mock adapter — every model-text block you see is a recorded fixture, not live model output, and no API key or network was used.

Three distinct lanes (don't mix them up)

1. Reference data — the real Kepler Q1 light curve (NASA/MAST) and the published ephemeris (NASA Exoplanet Archive). See data/README.md.
2. Model text (recorded fixture) — the hypothesis and adversarial_challenge ledger entries. In this run they come from grok-mock, a deterministic offline stand-in seeded by the dataset hash. They are labelled "adapter": "mock" in the ledger and visibly labelled as fixtures on the demo page. A live RealGrokAdapter exists (GROK_API_KEY from the environment, never stored), but was not used for the committed artifacts.
3. Deterministic verification — the pipeline (stdlib-only BLS transit search), the ledger hash chain, and verify_ledger.py. Everything here is reproducible byte-for-byte except timestamps.

Quickstart (zero install, Python ≥ 3.10, stdlib only)

python demos/ledger-grok/run_demo.py        # ~25 s: pipeline + ledger + results
python demos/ledger-grok/verify_ledger.py   # verify chain + data hash
pytest demos/ledger-grok/tests              # 19 tests (also wired into repo CI)

run_demo.py regenerates out/ledger.jsonl + out/results.json (committed artifacts are the reference run).

Result (committed reference run)

Quantity	This demo (Q1 only)	Published	Check
Period	0.8373346 d	0.8374912 d	Δ = 1.6×10⁻⁴ d (≤ 5×10⁻⁴) ✓
Mid-transit	BKJD 131.579626	BKJD 131.574858 + k·P	Δ = 7.0 min (≤ 30) ✓
Depth	158.3 ppm	191 ppm	ratio 0.83 (0.4–2.5) ✓
Transits / SNR	40 / 34.0	—	—
Odd/even depth	156.2 / 160.2 ppm (0.42σ)	expect ≈ equal	✓
Secondary dip	2.3 ppm (0.49σ)	expect ≈ 0	✓

Verdict: MATCH — validation (re-derivation), not a discovery. Honest caveat: with a single 33.5-day quarter the periodogram has several near-equal peaks; the measured period is the global BLS maximum and sits 0.019 % off the 17-quarter published value, inside the demo's stated tolerance. The demo proves the pipeline + ledger end to end, not new astronomy. One Kepler-10 c transit in the segment is masked and disclosed (see data/README.md).

Ledger format

One JSON object per line, canonical JSON (sorted keys, no NaN):

{"seq", "ts", "actor", "kind", "payload", "previous_hash", "entry_hash"}
entry_hash = sha256(canonical({seq, ts, actor, kind, payload, previous_hash}))

previous_hash of entry 1 is 64 zeros (genesis). verify_ledger.py recomputes every link, catches any modified/reordered/removed entry, and can also check that the CSV still matches the sha256 in data_ingested (--data). Kinds: data_ingested, hypothesis, analysis, adversarial_challenge, challenge_response, verdict.

Grok adapter

from ledger_grok.grok_adapter import get_adapter
adapter = get_adapter("mock", data_sha256)   # default: offline, deterministic
adapter = get_adapter("real", data_sha256)   # needs GROK_API_KEY in env

• MockGrokAdapter — fixed templates seeded by the dataset hash; zero credentials, zero network; used for the committed run and in tests.
• RealGrokAdapter — calls the xAI chat-completions API (GROK_API_KEY or XAI_API_KEY from the environment, never stored; GROK_MODEL override, default grok-4.20-0309-non-reasoning, temperature 0).

Bounded live proof (2026-07-19): exactly one minimal live call was made with tools/grok_live_probe.py — HTTP 200, 274 tokens total, key read from the environment only. Model, status, usage, and the sanitized one-sentence reply are recorded in out/grok-live-probe.json. The committed ledger and all tests remain on the deterministic offline mock; the probe is a separate one-call proof, not part of the reproducible run.

Layout

data/     committed CSV + provenance
src/ledger_grok/  fits.py pipeline.py ledger.py verify.py grok_adapter.py run.py
tools/    fetch_data.py (one-off networked fetch; everything else is offline)
out/      committed reference run: ledger.jsonl + results.json
site/     static demo page (GitHub Pages)
tests/    19 tests incl. artifact-anchoring (committed ledger must verify)
SHA256SUMS  integrity manifest for data + artifacts

Credit

Developed by The Beast AGI engineering fleet in collaboration with the ASTRA research program (Prof. Glenn White) — attribution only; no ASTRA code is used in this demo. Kepler data: NASA/MAST. Ephemeris: NASA Exoplanet Archive. License: Apache-2.0 (see repo root).