Trabeculae are the muscular ridges that line the heart’s ventricles, and the lattice inside bone that remodels where load falls on it. TRABECULAE groups the studies in which an autonomous campaign remodels its own conclusions as evidence arrives: a borrowed value proves wrong, a value turns out to describe a different processing state, an upstream model is re-released, runs are lost, or a reader needs to check that a revision did what it claims. All five work on one question, heat transport in SiGe, SiGeC and (Si)GeSn semiconductors, so each study removes one excuse for rebuilding everything.
NablaSDL: graded revision¶
∇ nabla. Bone remodels along the gradient of the load it carries; the campaign revises along the gradient of the evidence, and only where it points.
Nine graded actions on source-tagged evidence, from labelling or reweighting rows to forgetting them, confine each correction to conclusions named in advance. On the lattice thermal conductivity of SiGe and SiGeC, four of twelve flagged values were errors; the other eight were kept as special conditions or as leads to missing physics. The campaign showed that 2.5 nm SiC particles replace 14–16% of the germanium at equal thermal conductivity.
Use for a campaign: issue a targeted erratum instead of refitting from scratch when one source proves wrong.
CircSDL: processing history¶
∘ circ. Compact bone grows in concentric rings, each one a layer of its history; every value here carries the ring of processing it passed through.
Every value is stored as an event in a time-resolved evidence hypergraph that records the film’s state before and after each anneal. Literature activation energies for Sn diffusion spanned 1.13 to 3.10 eV, but fell into groups no wider than 0.7 eV once anneal regime and starting state were recorded.
Use for a campaign: treat two disagreeing values as possibly two different samples before calling either an error.
LeadstoSDL: reach of a model release¶
⇝ leads to. Load travels strut to strut along a path; a model release travels result to result, and the record says where it can lead.
When MACE-MP-0 was replaced by MACE-MPA-0, the campaign’s provenance record predicted which conclusions could change before any rerun. Recomputing only the four named units settled 20 conclusions, and none changed outside the predicted set.
Use for a campaign: recompute what a new release can actually change, not everything downstream of it.
ForallSDL: coverage under lost runs¶
∀ for all. A trabecular lattice braces every direction a load might come from; a covering batch samples every pair, even when a strut is lost.
Covering designs from software testing sample every pair of variable levels at least once, or twice so that any single lost run leaves the guarantee intact. For doped nanostructured SiGeC, 33 runs covered every two-factor combination of six coupled variables.
Use for a campaign: state before the first run which combinations a batch speaks to, even if runs are lost.
TurnstileSDL: checkable revisions¶
⊢ turnstile. Each strut carries what it can prove it carries; each revision passes the gate only after anyone can re-derive it.
Designs are hashed onto an append-only record before runs start, raw outputs are committed before analysis, and every revision is verified by deterministic re-execution. An independent verifier re-checks a campaign in about three seconds.
Use for a campaign: let anyone check every revision without trusting the campaign.
Code and data¶
Each article has a public repository under Sparks