Hash-committed pre-registration and exact unlearning make belief revisions in autonomous SiGeC semiconductor campaigns checkable
Abstract
Autonomousmaterials laboratories revise conclusions fromevidence sources (calculations,machine-learned potentials, databases, furnace runs) whose errors are known in kind but not in size. These revisions are self-reported, so a reader cannot check that no batch was dropped or that a retraction removed what it claims; verifiable unlearning covers only erasure from trained machinelearning models. The aim is a record that lets anyone check every revision without trusting the campaign. Full checkability is hypothesised to be affordable when beliefs are compact symbolic models, whose clean retrain takes seconds. Designs are hashed onto an append-only campaign record before runs start, raw outputs are committed before analysis, and revisions are verified by deterministic re-execution. Four pre-registered campaigns on heat transport in SiGe and SiGeC semiconductors (molecular dynamics, machine-learned relaxations, phonon transport, an experimental database) test the scheme. The record keeps failed checks, refuted predictions and set-aside evidence in view, and an independent verifier re-checks it in about three seconds. It meets pharmaceutical data-integrity rules; separating a measurement’s influence on fitted laws from the measurement reconciles the European right to erasure with mandatory retention. Materials claims, such as howmuch germanium carbon can replace, can thus be adopted and extended on an inspectable record.
Related
- Charge density wave instability below the semimetal-to-semiconductor transition in Mo8O23
- Room-temperature multiferroic semiconductor with electrically switchable valley polarization in Zn-/Cd-intercalated CrSe2
- Stress induced semiconductor to semimetal transition and tunable optical anisotropy in puckered bilayer GeTe: a first-principles study
- OSC700: A GW+BSE Dataset of 700 Organic Semiconductor (OSC) Crystals with an LLM-Assisted Data Extraction Agent
- Colossal visible birefringence and ultraviolet hyperbolicity in wide bandgap semiconductor CuAlO2
BibTeX
@misc{ackolf2026hashcommitted,
title = {{Hash-committed pre-registration and exact unlearning make belief revisions in autonomous SiGeC semiconductor campaigns checkable}},
author = {Kinston Ackölf and Taylor D. Sparks},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010668/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010668/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Hash-committed pre-registration and exact unlearning make belief revisions in autonomous SiGeC semiconductor campaigns checkable
AU - Kinston Ackölf
AU - Taylor D. Sparks
PY - 2026
DA - 2026/10/19
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010668/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010668/v1
AB - Autonomousmaterials laboratories revise conclusions fromevidence sources (calculations,machine-learned potentials, databases, furnace runs) whose errors are known in kind but not in size. These revisions are self-reported, so a reader cannot check that no batch was dropped or that a retraction removed what it claims; verifiable unlearning covers only erasure from trained machinelearning models. The aim is a record that lets anyone check every revision without trusting the campaign. Full checkability is hypothesised to be affordable when beliefs are compact symbolic models, whose clean retrain takes seconds. Designs are hashed onto an append-only campaign record before runs start, raw outputs are committed before analysis, and revisions are verified by deterministic re-execution. Four pre-registered campaigns on heat transport in SiGe and SiGeC semiconductors (molecular dynamics, machine-learned relaxations, phonon transport, an experimental database) test the scheme. The record keeps failed checks, refuted predictions and set-aside evidence in view, and an independent verifier re-checks it in about three seconds. It meets pharmaceutical data-integrity rules; separating a measurement’s influence on fitted laws from the measurement reconciles the European right to erasure with mandatory retention. Materials claims, such as howmuch germanium carbon can replace, can thus be adopted and extended on an inspectable record.
ER -