Kill Report2 min read

When All 13 Compute Results Were Error Payloads

Every single Modal QE DFPT and HSE06 result from our April 2026 GPU compute campaign was an error payload, not usable data. 13 files landed on the volume with status ERROR, PSEUDO_ERROR, SCF_FAILED, DFPT_FAILED, PARSE_ERROR, or NOT_CONVERGED. The honest-null pattern in our ledger is what saved rankings from corruption.

LG
Lattice Graph Research
|BiIrSeBaHfN2Sm3InS6Tb3InS6

When All 13 Compute Results Were Error Payloads

The Discovery

During a third-pass deep audit of our compute results, we discovered that ALL Modal Quantum ESPRESSO DFPT and HSE06 results from April 14–16, 2026 were error payloads, not usable data. The files did land on the Modal volume and were downloaded locally, but every single one had a failure status. No usable physics was extracted from the entire GPU campaign.

The 13 Failed Results

CompoundStatusImpact
BiIrSePSEUDO_ERROR — Missing pseudo for IrHSE band gap unknown; survive-vs-kill decision blocked
BaHfN2NOT_CONVERGEDHSE band gap unresolved; filed with PBE Eg=1.25 eV + caveat
Sm3InS6PSEUDO_ERRORHSE band gap unknown
Tb3InS6PSEUDO_ERRORHSE band gap unknown
La3InS6NEVER_RANPrimary lead; only PBE Eg=1.36 available
ZnSiP2NEVER_RANSolar HSE band gap unknown
AgPS3ERRORHSE band gap unknown
BaZnSOERROR (scale_sym_ops)HSE band gap unknown
BiSeClO3PARSE_ERRORBorn charges + NLO claims unvalidated
Er3GaS6PARSE_ERRORPiezo tensor unvalidated
Li2HfO3SCF_FAILEDDFPT dielectric relies on Materials Project only
BaZnOSSCF_FAILEDDFPT failed
FeBiO3DFPT_FAILEDMultiferroic DFPT failed

Root Causes

  1. Ir pseudo-potential bundling failure. add_local_file in the Modal image build likely failed silently — the pseudo-potential never actually mounted. The error fires in 0.002 seconds, before pw.x even launches.
  2. HSE06 SCF convergence failures. Typical HSE06 tuning issues: q-mesh too large, needs nosym=.true., electron_maxstep=200+, mixing_beta=0.2.
  3. DFPT parse errors. The regex pattern in _parse_dielectric didn't match the ph.x output format across QE versions.
  4. Never-produced files. Unhandled exceptions before volume.commit() wrote result JSON — 5 compounds simply never ran.

What Saved Us: The Honest-Null Pattern

The compound ledger did NOT adopt any failed values. The build script only reads kappa_300k_chosen, eg_hse, and e_* fields — all of which are null in error JSONs. The ledger correctly left these columns null for every failed compound.

This is the critical design decision that saved the rankings from corruption. Rankings were honest — they just reflected that GPU validation for ~15 compounds was still pending.

The lesson: error payloads will masquerade as results. If your pipeline reads fields from JSON files without checking the status code first, failed compute will silently corrupt downstream rankings. The honest-null pattern — leaving columns null when compute fails, rather than writing default or partial values — is what maintains data integrity when the compute layer is unreliable.

What Changed Afterward

  • Docs that said "HSE06 queued" were reclassified as "HSE06 attempted and failed; retry required."
  • Modal pseudo-potential delivery switched from add_local_file to a dedicated Modal Volume mount.
  • All provisional filings proceeded with PBE band gaps + an explicit HSE06 caveat.
  • The DFPT parser was rewritten to handle QE version-specific output format variations.

See what survived our screening

Lattice Graph computationally validates materials before you commit lab time. Explore the candidates that passed stability, cost, and cycle-life gates.

Explore the Platform