CKC examples
Every example below is a valid Conventional Commit. Footers use the git-trailer form
(Token: value, repeatable). The proof examples use a formalization (Intrinsic Green's Learning);
the science examples use its experiments.
proof profile
Add a definition or statement, with no proof yet
state(green-kernel): define the elliptic Green's kernel operator
Lean: IGL.greensKernel
Status: math.conjectured
Conjecture a result
conjecture(master-formula): the Green's integral factorises as a finite rank-K×R sum
Claim-ID: conjecture:master-formula
Status: math.conjectured
Formalize and close it cleanly
formalize(fubini): close the d-fold separable factorization
Lean: IGL.fubini_dfold, IGL.fubini_factorization
Status: math.machine-checked
Axioms: propext, Classical.choice, Quot.sound
Verified-By: Lean 4.30.0; Mathlib v4.30.0
Closes: conjecture:master-formula
Cite an axiom (note the ~ and the AXIOM: footer)
axiomatize~(exp-sum): cite the logarithmic exponential-sum rank bound
The target library has no elliptic-Green's exponential-sum theory, so posit it as a citation.
Lean: IGL.expSumRank_logBound
Status: math.axiomatised
AXIOM: IGL.expSumRank_logBound (separable rank K = O(log 1/ε), dimension-independent)
Source-Ref: Braess & Hackbusch 2005, Numer. Math., Thm 3.1
Cites: Braess & Hackbusch 2005, Numer. Math.
Cite an axiom with its fidelity anchors
axiomatize~(cheb-interp): cite the Chebyshev interpolation error bound
The target library has no multivariate Chebyshev interpolation theory, so posit the bound as a
citation. The source requires f analytic in the Bernstein ellipse E_ρ, ρ > 1; both hypotheses kept.
Lean: IGL.chebInterpError
Status: math.axiomatised
AXIOM: IGL.chebInterpError (sup-norm error ≤ C·ρ^(−n), f analytic in E_ρ, ρ > 1)
Source-Ref: Trefethen 2013, Approximation Theory and Approximation Practice, Thm 8.2
Refutation-Attempt: none-found: ρ → 1 collapse, merely-continuous f, n = 0
Fidelity-Reviewed: aq, 2026-07-03
Cites: Trefethen 2013, Approximation Theory and Approximation Practice
The same axiom without its anchors (and what goes wrong)
axiomatize~(cheb-interp): cite the Chebyshev interpolation error bound
Lean: IGL.chebInterpError
Status: math.axiomatised
AXIOM: IGL.chebInterpError (sup-norm error ≤ C·ρ^(−n) for continuous f)
Cites: Trefethen 2013
This commit validates under v0.1.0: a known type, a status that matches the AXIOM: footer, a real
citation. But the transcription dropped the analyticity hypothesis, so the axiom claims a geometric
rate for every continuous function, which is false, and no kernel gate can see it, because the
kernel checks the statement as written, not the statement Trefethen proved. Rule 14 makes the
missing Source-Ref: a violation; see the case study for a real instance.
A clean proof that inherits a caveat through a dependency
formalize(approx-error): bound the approximation error via the exp-sum rank
Lean: IGL.approxError
Status: math.machine-checked
Axioms: propext, Classical.choice, Quot.sound
Depends-On: IGL.expSumRank_logBound
Its own #print axioms is clean, but its effective status is axiomatised, because it Depends-On
an axiomatised node. The ClaimGraph computes this, so the commit need not restate
it.
Leave an explicit gap
formalize~(rank-bound): scaffold the rank-bound proof; one case is still open
Lean: IGL.expSumRank_logBound
Status: math.open
OPEN: the r→0 boundary case is still a `sorry`
Refute a conjecture (a breaking change)
refute(separability)!: a counterexample blocks naive separable factorization
A radial counterexample shows the kernel does not factor under the affine-invariant metric.
Disproves: conjecture:naive-separable
Status: math.disproved
BREAKING CHANGE: conjecture:naive-separable is withdrawn; the compensation result depends on it.
Strengthen (additive; the old form still follows)
strengthen(gauge): prove gauge invariance for all C¹ diffeomorphisms, not only linear maps
Lean: IGL.gauge_invariance
Status: math.machine-checked
Supersedes: IGL.gauge_invariance_linear
science profile
Register a hypothesis (pre-registered)
conjecture(tensor-rank): tensor rank K>1 beats K=1 on non-additive targets
Claim-ID: conjecture:tensor-rank-helps
Status: sci.hypothesis
Pre-Registration: osf.io/abcd1
Record an experiment
experiment(non-additive): K=1 vs K>1 on XOR, radial, and multiplicative targets
Status: sci.measured
Metric: MSE
Sample-Size: n=1000
Seed: 0..4
Hardware: M4 (MPS)
Effect-Size: ΔMSE −0.41 (K=8 vs K=1)
Pre-Registration: osf.io/abcd1
Protocol-Deviations: none
UNREPLICATED: single machine, 5 seeds; needs an independent run
Closes: conjecture:tensor-rank-helps
A failed replication (a normal commit)
replicate(non-additive): an independent run fails to reproduce the K>1 advantage
Status: sci.not-replicated
Dataset: D-2026-014 sha256:aa80…
Sample-Size: n=1200
Effect-Size: ΔMSE −0.03
CI: 95% [−0.09, 0.04]
Refutes: conjecture:tensor-rank-helps
A code fix that changes the numbers but not the finding
repro-fix(scaling): correct an off-by-one in the epoch windowing
The direction is unchanged and the effect size is revised down. This is not a new finding.
Affects: result:igl-linear-scaling
Effect-Size: +6.8pp → +5.1pp
Seed: 0
Impact: patch
A data correction that flips a conclusion (a breaking change)
data(bnci)!: fix swapped treatment and control labels in the BCI dataset
Dataset: D-2025-019
Invalidates: result:cross-subject-transfer
Status: sci.falsified
BREAKING CHANGE: result:cross-subject-transfer no longer holds under corrected labels.
tooling and prose (plain Conventional Commits still apply)
docs(spec): clarify the trusted-base marker rule
chore: tag v0.1.0