{"id":"conjecture-ledger","title":"The Hypothesis Ledger","subtitle":"Every claim we have tested, its lifecycle status, and why it moved.","category":"conjectures","tags":["ledger","index","overview"],"source":"articles/docs/conjectures/ledger.md","lang":"en","words":532,"readMinutes":2,"toc":[{"depth":2,"text":"The ledger","id":"the-ledger"},{"depth":2,"text":"2026-06-07 Kimi MLIP Import","id":"2026-06-07-kimi-mlip-import"},{"depth":2,"text":"Why this shelf exists","id":"why-this-shelf-exists"}],"html":"

Conjectures & Proofs — The Hypothesis Ledger

Every claim Lupine has seriously tested, where it stands, and why it moved. This is\nthe structured counterpart to the narrative changelog: the changelog tells the story,\nthe ledger tells the state.

\n

Each hypothesis is its own entry with a lifecycle status. The status legend:

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
StatusMeaning
SupportedSurvives the strongest test we have applied so far.
OpenLive; evidence is partial or mixed.
Refuted by usWe tried to confirm it and the effect did not survive a fair test. The confounder is named.
Self-correctedWe announced something, then found our own error and retracted it.
Proven (Lean)Cross-checked by a machine-checked Lean 4 theorem.

The ledger

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
HypothesisStatusOne-line resolution
Hyper-ribbon universality (classical potentials)Supported · ProvenError vectors occupy a low-dimensional manifold; Lean-grounded.
Hyper-ribbon transfers classical → MLIPUnder re-audit (2026-06-11)Prior evidence: 14/15 on-ribbon under MACE/CHGNet/Orb-v3. Born screening (replication/error-geometry) excludes 7/45 foundation-model tensors (incl. CHGNet-Fe, MACE-V, Orb Al/Nb/Pb/Pt); per-element counts must be recomputed on screened inputs. Directional structure independently confirmed at n=8–11 models (rank-1 share 0.56–0.94).
Projected hyper-ribbon releaseOpenNew Lean-first release lane: prove projected-ribbon gates, then require replay plus cloud evidence before promotion.
Cross-MLIP orthogonal error modesSupportedMACE and CHGNet have orthogonal error directions on Ag/Nb/Pd.
Au escapes the ribbon under foundation MLIPsOpenConfirmed for MACE+CHGNet; Ag escape refuted.
Fe magnetic MLIP failure modeOpen · under re-auditThe old PR > 2 trio claim is frozen after Born screening excluded CHGNet-Fe; Fe remains a magnetic/mechanical-stability failure-mode target.
D-band controls error correlationRefuted by usSample-size confounder (full-sample ρ = −0.02).
MEAM is intrinsically 2-DRefuted by usMatched-n bootstrap: MEAM overlaps Tersoff.
BCC/FCC "causal shield"Self-correctedThe dramatic r 0.90 vs 0.04 was 1.5 % data contamination.
Simpson's paradox in BCC elastic constantsRefuted by us · LeannoSimpsonsInBccEam: the causal graph has no bypass.

2026-06-07 Kimi MLIP Import

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
HypothesisStatusOne-line resolution
Parameter-basis Vandermonde decay in foundation MLIPsRefuted by usThe 4-model Fisher sweep fails rho >= 1.5; MACE parameter-basis rho is flat while CHGNet/SchNet only reach about rho 0.4.
MACE irrep-basis Vandermonde thresholdRefuted by usIrrep coefficients show real geometric decay (rho 0.3865, R2 0.9807), but still fail the pre-registered rho >= 1.5 threshold.
Weak-form acceleration/refusal theoremOpenScalar weak-form gate now builds in Lean; full Lipschitz/reach formalization and deeper-model runtime evidence remain open.
Layerwise distance as MLIP/MD uncertainty signalOpenLayer-0 distance correlates with force error and helps mixed-reference refusal, but Cu-only reference tuning failed and force-calibrated follow-up is needed.
Kimi Cloud Run cross-MLIP v7Supported45 MACE/CHGNet/SevenNet elastic calculations are preserved; Fe is a MACE-disagreement sentinel, while Ta/V/Pt have the highest 3-MLIP PR values.

See also:\nKimi MLIP Universality Import ·\nCross-MLIP Cloud Experiment Runbook.

\n

Why this shelf exists

The most defensible thing Lupine produces is not a single result — it is a method that\ncatches its own mistakes. The d-band and MEAM refutations and the BCC/FCC\nself-correction all came from the same matched-n / contamination-gate discipline. Making\nthe refutations as visible as the confirmations is the point: a corpus you can trust is\none that publishes what it killed.

\n

See also: Formal Proof Ledger ·\nMethodology · Data & Provenance.

\n"}