Lupine Science accelerating materials discovery

Research note

The Savings Stack: A Booklet on Doing Less, and the Commons That Makes Everyone Faster

Seven savings layers from the literature, one measured sharing economy, and the theorem commons — as a free, designed booklet with a hashable data pack

Many pale compute lanes converging into a compact shared evaluation backbone that feeds several materials programs: reuse instead of brute force.
Brand film: The Savings Stack: A Booklet on Doing Less, and the Commons That Makes Everyone Faster

Video page · Download MP4 · Download captions

Today we are releasing The Savings Stack — a sixteen-page booklet on how atomistic simulation learned to do less, what that savings stack is still missing, and the measured economics of sharing the expensive part. It is a free download, built to be read: the literature digests, the union-anchor measurement, and the citation audit are all in the pack behind it, hashed and reproducible.

Download: The Savings Stack — 16 pages, US Letter, free.

Get the booklet (PDF, 748 KB) →

Inside the booklet

The booklet is the designed edition of the Savings Stack v1 release: typography-led, one idea per page, every number anchored to the committed record.

The seven-layer stack, drawn to scale — each layer with its strongest verified literature number, the shared crack “no certificate” through all seven, and the missing capstone layer

Seven chapters of the underlying book survey the savings layers the field invented independently — surrogate transition-state search, active learning, Δ-learning, abstention economics, systems-level DFT acceleration, electronic-structure surrogates, and path and sampling algorithms — each with its strongest verified number and its named hole. Two backdrop chapters cover four decades of supercomputing economics and today’s compute-resourcing landscape. The hole is the same shape in every layer: plenty of signals for where the model is probably wrong, no certificate for where physics is definitely violated.

The big-number page: 72.4% fewer DFT evaluations when four models share anchors instead of each paying for its own

The chart: naive per-model anchor cost explodes 139.5 → 558 as models go from one to four, while the shared union stays nearly flat at 139.5 → 154

The measured core, in one paragraph

On the locked 30-path Z1 panel, path 14 failed CI-NEB convergence in all four model artifacts — the honest denominator is 29. Across those 29 analyzable diffusion paths, the frozen sparse-anchor protocol would consume 558 naive per-model DFT anchors; shared across the four models, the union is 154 — 72.4% fewer evaluations, a 3.62× reduction. The scaling law is the headline: as guidance models go from one to four, the naive bill multiplies ~4× (139.5, 279, 418.5, 558) while the union stays nearly flat (139.5, 147.8, 152, 154) — four models of guidance cost ~10% more DFT than one. The numbers were recomputed from recorded campaign artifacts under the frozen anchor rule; an earlier informal figure was retracted as arithmetic drift, and the analysis note in the pack reconciles it line by line. No new DFT was run to produce any of it.

A sequence of three physical screening fixtures on one bench: a wide tray of candidate coupons, a smaller thermal-test rack, and one final reference-calculation instrument — Successive screening fixtures narrow the candidate set before the expensive reference instrument

Branching simulation paths reusing the same sparse set of shared anchor evaluations — the blank space is calculations deliberately not run

What it is — and what it is not

The booklet makes the case for a theorem commons: teams contributing formally verified physical-law theorems back, so everyone’s gates get sharper. Theorems are non-rival and leak nothing about the contributor’s chemistry; the measured union economics are the evaluation-side instance of the same sharing. The case for restraint is designed with equal care: the measured basis is one 30-path panel, four universal MLIPs, one DFT engine, one chemistry family; the multiplicative stacking figure is a derived estimate and is labeled as such wherever it appears; the digests keep their [UNVERIFIED] flags; and none of this is peer-reviewed — it is a research release with committed, hashable records.

The citation audit behind the book verified every identifier against its registry: 115/115 arXiv, 48/48 DOI, zero unresolved — with three corrections applied on the record, including the VASP MLFF venue reconciliation. Identifier verification does not upgrade abstract-only evidence, and the booklet says so.

Data and reproduction

The booklet is the cover story; the pack is the evidence. The v1 pack ships all ten chapters, the machine-readable union-anchor record (JSON + sha256 sidecar), the analysis note and script, the citation audit, and a MANIFEST.sha256 over every file. Structured data is released under the Open Data Commons Open Database License 1.0 (ODbL-1.0), mirroring the source repository’s data license; prose and code keep their repository licenses.

An open rack of identical unmarked sample cartridges feeding two separated battery test instruments through standardized mechanical sockets — The standardized cartridge socket lets the same physical reference sample serve independent instruments

sha256sum -c MANIFEST.sha256 verifies the pack; python3 union_anchor_economics.py --local <artifacts> --out z1-union-anchor-economics.json recomputes the record.

Receipts

Next: read the booklet, hash the pack, recompute the record.

Download The Savings Stack →