Public executive report. Detailed evidence and methods remain in a separate confidential annex, available on request after confidentiality review and a signed NDA.
Executive summary / 28 September 2026
BGC87 / FOUR-SYSTEM COMPARISON
A cross-field assessment of reference agreement, recorded delivery time and observed repeatability.
Linur Engine / Opus / Sol / Astra
Reference agreement
Delivery time
Observed repeatability
Conceptual illustration, not measurement data or a model ranking.
Linur Engine, Opus, Sol and Astra. Reference agreement, delivery time and repeat observations remain separate. All cases retained.
Numerical reference tasks
8 planned casesElectrochemical storage
6 planned casesElectronic materials
5 planned casesEnergy and catalysis
8 planned casesIsotopes and nuclear energy
7 planned casesPlasma-facing materials
6 planned casesQuantum computing and sensing
3 planned casesAI and computing materials
11 planned casesAutonomous discovery
5 planned casesStructural materials
5 planned casesBiotechnology and medicine
11 planned casesSpectra and reaction networks
7 planned casesCrystal structure prediction
5 planned casesMaterials properties, calibration, fits, spectra and trends connect calculations with devices and decisions.
Accuracy, model assumptions, convergence, computing time and resources all matter. No separate conventional-solver comparison was run.
Declared data and contracts
One quantity is one requested output, not one task.
An output-specific allowed difference; not universal scientific truth.
Absolute error / absolute reference. Median within, then across cases.
Global field charts and task-specific evidence retain every produced result, regardless of protocol verdict. Reference differences, criterion margins and delivery time/repeatability are assessed separately. Field captions describe scope, not whole-field validation.
Reference agreement, recorded time and missing evidence remain distinct.
A requested output can be a number, category, list or matrix. Scalar summaries use the same output IDs for all four systems; selection, rank and vector evidence uses its own metrics.
8 planned tasks
Quantum many-body reference; Main-group thermochemistry; Noncovalent interactions; Solid-state reproducibility; Lattice dynamics & thermal transport; Criticality & neutron transport; Liquid water structure & dynamics
1 matched timed tasks.
6 planned tasks
Lithium-metal & solid-state batteries; Solid-state electrolytes; Cathode degradation & thermal runaway; Fast charging & lithium plating
2 matched timed tasks.
Scalar summaries are medians within, then across cases. Missing a common cohort does not mean no results were assessed. Timing boundaries differ; recovery never inherits the original unsuccessful interval.
Reference agreement, recorded time and missing evidence remain distinct.
A requested output can be a number, category, list or matrix. Scalar summaries use the same output IDs for all four systems; selection, rank and vector evidence uses its own metrics.
5 planned tasks
Twistronics & quantum materials; 2D semiconductors for quantum technology; Graphene-based sensors
5 matched timed tasks.
8 planned tasks
Perovskite photovoltaics; Electrocatalysis; Thermoelectrics; Hydrogen storage
8 matched timed tasks.
Scalar summaries are medians within, then across cases. Missing a common cohort does not mean no results were assessed. Timing boundaries differ; recovery never inherits the original unsuccessful interval.
Reference agreement, recorded time and missing evidence remain distinct.
A requested output can be a number, category, list or matrix. Scalar summaries use the same output IDs for all four systems; selection, rank and vector evidence uses its own metrics.
7 planned tasks
Advanced reactor inventory & decay heat; Radiation damage in structural materials; Medical isotope production; Hydrogen isotope separation; Actinide chemistry & reprocessing
5 matched timed tasks.
6 planned tasks
Core plasma turbulence & confinement; Plasma–material interactions; Tritium breeding blanket; High-temperature superconducting magnets
1 matched timed tasks.
Scalar summaries are medians within, then across cases. Missing a common cohort does not mean no results were assessed. Timing boundaries differ; recovery never inherits the original unsuccessful interval.
Reference agreement, recorded time and missing evidence remain distinct.
A requested output can be a number, category, list or matrix. Scalar summaries use the same output IDs for all four systems; selection, rank and vector evidence uses its own metrics.
3 planned tasks
Superconducting qubit decoherence; Colour-centre spin qubits
1 matched timed tasks.
11 planned tasks
Post-silicon transistors; Non-volatile memory; Thermal management in 3D integrated circuits; Neuromorphic computing; EUV lithography & patterning; Wide-bandgap power electronics
6 matched timed tasks.
Scalar summaries are medians within, then across cases. Missing a common cohort does not mean no results were assessed. Timing boundaries differ; recovery never inherits the original unsuccessful interval.
Reference agreement, recorded time and missing evidence remain distinct.
A requested output can be a number, category, list or matrix. Scalar summaries use the same output IDs for all four systems; selection, rank and vector evidence uses its own metrics.
5 planned tasks
AI-driven materials discovery; AI-driven protein design; Autonomous closed-loop experimentation
Recorded intervals / planned cases:
Linur 0 / 5
Opus 4 / 5
Sol 4 / 5
Astra 4 / 5
5 planned tasks
High-entropy alloys; Additive manufacturing; Corrosion & passivation; Rare-earth-lean permanent magnets
2 matched timed tasks.
Scalar summaries are medians within, then across cases. Missing a common cohort does not mean no results were assessed. Timing boundaries differ; recovery never inherits the original unsuccessful interval.
Reference agreement, recorded time and missing evidence remain distinct.
A requested output can be a number, category, list or matrix. Scalar summaries use the same output IDs for all four systems; selection, rank and vector evidence uses its own metrics.
11 planned tasks
Structure-based drug design; Drug resistance; Enzyme design; Antimicrobial resistance; Biopharmaceutical formulation; Nanomedicine & drug delivery
Recorded intervals / planned cases:
Linur 0 / 11
Opus 11 / 11
Sol 11 / 11
Astra 9 / 11
7 planned tasks
Homogeneous catalysis; Combustion & sustainable fuels; Excited-state dynamics; Spectroscopic prediction
2 matched timed tasks.
Scalar summaries are medians within, then across cases. Missing a common cohort does not mean no results were assessed. Timing boundaries differ; recovery never inherits the original unsuccessful interval.
Reference agreement, recorded time and missing evidence remain distinct.
A requested output can be a number, category, list or matrix. Scalar summaries use the same output IDs for all four systems; selection, rank and vector evidence uses its own metrics.
5 planned tasks
High-temperature superconductors; Gas separation & carbon capture; Metastable phase discovery
2 matched timed tasks.
Scalar summaries are medians within, then across cases. Missing a common cohort does not mean no results were assessed. Timing boundaries differ; recovery never inherits the original unsuccessful interval.
Produced / requested primary outputs; coverage is not correctness.
Bars show produced / requested primary outputs as a percentage. An output can contain many matrix elements; coverage is not scientific correctness. Exact values, cohorts and definitions: Annex.
Median seconds on the same complete-output cases for all four systems.
Median seconds on shared complete-output cases. Recorded timing boundaries differ: not equal-hardware compute time. Missing Linur intervals are never inferred. Exact values, cohorts and definitions: Annex.
Scalar reference difference; lower is closer to the declared reference.
Absolute error / absolute reference; median within each case, then across cases. Not accuracy percentages. Medians can hide large errors; censored and undefined ratios stay separate. Exact zero is written as 0 instead of a log-scaled bar. Exact values, cohorts and definitions: Annex.
Error / allowance; below 1 is inside the numerical allowance.
Median error / allowance within each case, then across cases. The bold line at 1 marks the original allowance, not a new acceptance rule. Exact zero is written as 0 instead of a log-scaled bar. Exact values, cohorts and definitions: Annex.
Exact agreement across three recorded trials on the same shared output series.
Percentage of shared complete three-trial series with exact agreement. No shared series and no planned repeats are different gaps; agreement is not correctness or controlled reproducibility. Exact values, cohorts and definitions: Annex.
FAIL and schema mismatch do not exclude a produced numerical observation.
Less is better. 35 matched timed tasks; median seconds.
Different boundaries; not common-hardware compute time.
Less is better. 57 cases / 668 shared scalar outputs.
Median of case-median relative errors. Log ratio axis, not accuracy %.
Less is better. 42 cases / 366 shared tolerance outputs.
Median of case-median error/allowance ratios. Limit 1 lies beyond this log axis.
More is better. 34 shared complete three-value series.
Exact observed agreement; not correctness or controlled reproducibility.
Each panel has its own population and scale. Categories, vectors and eight censored conductivity bounds use their own metrics in the annex. Medians can hide large individual errors; no universal score or composite system ranking is created.
4,724 produced result cells across primaries and repeats. 4,715 support a reference-based numerical or categorical assessment; nine have no supplied reference.
PASS / FAIL remains the historical protocol label. Signed errors and distance from the original allowance are shown even for FAIL. A negative margin is inside the allowance; positive is outside.
All 121 recovered conductivity values are assessed numerically: 113 exact references and eight censored bounds. Their frozen exact-text score remains unresolved, not rewritten.
Closer agreement with a specified reference is measurable evidence, not independently established scientific truth. Reference uncertainty, source resolution, physical interpretation and model limitations still matter. No threshold is invented for an originally descriptive rule.
Timing uses its recorded boundaries, not a PASS filter. 780 of 1,012 planned output-repeat series have at least two produced observations; partial series and runtime identity limits remain explicit.
No missing output, time, resource measurement or repeat has been synthesized. This is not a common-hardware efficiency or universal superiority claim.
Original and recovery packages, all 492 trial records, and historical scores remain intact. Internal recovery is first-attempt evidence by user instruction, not witnessed execution attestation.
Source-declared log, vector, selection and ranking metrics are identified. Additional signed differences, criterion margins and descriptive summaries are labelled analyst calculations.
Nine saved B-minus-A values are assessed against explicit derived A/B reference contrasts. They do not increase the 881 requested outputs per system.
Upper-bound consistency is not exact-reference agreement. Numeric text is parsed only for descriptive arithmetic; no response is converted or categorical rule changed. Sentinel mismatches, ordering, status provenance and unresolved cutoff ties remain explicit.
The S032 temperature interpretation/equality discrepancy and source-resolution caveats remain. Original Linur accepted-attempt IDs stay null. The historical 45-case / 420-check protocol cohort is preserved as historical evidence, not used as the new assessment filter.
Thirty-five cases have common complete-output timing. Recovery component intervals are reported separately, without replacing missing end-to-end time. Agreement and recorded repeat stability do not establish architecture, mechanism, universal capability or independently validated correctness.
FULL EVIDENCE EDITION. Reference-sensitive material. Not independent scientific peer review.