Skip to content

Supplementary information

What each workspace does, and how it was checked

The five workspaces behind the article, the recorded runs that test them, and what is not built yet.

S1Engineering

Finite-element analysis on dedicated cloud compute. Geometry is meshed with gmsh into quadratic tetrahedra and solved with CalculiX; the result is checked against the closed-form case the run declares, and its outer surface is extracted for inspection in the browser.

The benchmark is a cantilever chosen because its answer is known. It is the run behind the article’s structural results.

Table S1 The benchmark run, 0b63e3eb.
ItemRecorded
Modelcantilever 1.0 × 0.05 × 0.05 m, carbon fibre / epoxy, E = 70 GPa, ν = 0.30
Load1000 N at the free end
MeshC3D10 at 0.005 m — 146,375 nodes, 95,558 elements
Solve103 s on c6i.2xlarge; result file 56.3 MB
Tip deflection9.1350 mm against 9.1429 mm closed-form (−0.086 %)
Peak von Mises62.14 MPa against 48.00 MPa nominal (+29.5 %, lower bound)
Safety factor9.6553 against a 600 MPa yield strength
Repeatretried as 81c9ace7 — identical metrics
From the interface8b5a6280, f37401d7 and faa58e92, 5–7 Oct 2026: the same figures, 164.4 to 169.3 s from launch to result
Viewer19,716 triangles on 9,860 nodes, 0.34 MB (run cd6f54b4)

Checks. Every completed solve is read against four fixed rules, by code rather than by a language model: peak stress above yield (a safety factor below 1); the peak within one element length of a fully fixed face, where the restraint raises the stress and the value keeps rising as the mesh is refined; fewer than four elements across the part's thinnest side, a rule of thumb and labelled as one; and a nonlinear solve that did not converge.

Explanation. When an owner switches the assistant on, a run can be described in plain English by a language model. The model writes words and nothing else: the server inserts every figure from the run's record, and withholds any sentence that states a number of its own or says which way is up, which no record states. The model computes nothing and decides nothing.

Table S1b The first explanation, of run faa58e92 (7 Oct 2026).
ItemRecorded
Modelopenai/gpt-oss-120b on Groq, zero data retention on Sojist's account
Sentthe run's record: its name, setup, material data and results, and the geometry file's name; not the file itself, and no person's name or e-mail address
Answer4 sentences in 1,876 ms; every figure in them inserted by the server
Withheld1 of the 4 under the rule added that evening: it called the load “downward”, and the record gives the direction only as an axis, −Z
Auditthe request, the answer and the check's verdict are stored with who asked and when
Defaultoff in every workspace until an owner switches it on

S2Physics

Closed-form verification. A run declares which analytical case it models; the platform computes the reference values and compares each measured quantity in the mode the case permits — directly, or as a lower bound where a support can only raise the stress. A quantity the run did not report joins no verdict, and a run with nothing measured is inconclusive.

Table S2 The five analytical cases.
CaseSourceDeflectionStress
Cantilever, end loadRoark, Table 8.1, 1aPL³/3EI · directPLc/I · lower bound
Cantilever, uniform loadRoark, Table 8.1, 1cqL⁴/8EI · direct(qL²/2)c/I · lower bound
Simply supported, central loadRoark, Table 8.1, 2aPL³/48EI · direct(PL/4)c/I · lower bound
Simply supported, uniform loadRoark, Table 8.1, 2c5qL⁴/384EI · direct(qL²/8)c/I · lower bound
Bar in tensionHooke’s lawPL/AE · directP/A · direct

Try the first case with the platform’s own rules in Box 1.

S3Biology

Sequence analysis runs on the server, from the stored bytes: length, base composition, GC content over unambiguous bases, GC in up to 200 windows, and open reading frames of at least 100 aa in all six frames. Inputs above 20 million bases are refused rather than truncated, so a partial analysis is never reported as complete.

Read alignment runs on a dedicated instance with minimap2. The output is parsed by read, not by line — minimap2 writes supplementary alignments, and counting lines would push the mapping rate above 100 %.

Open reading frames are labelled as open reading frames, not genes: an ORF finder locates a stretch between a start and a stop codon, and a gene is an annotation.

Sequence files, FASTA and FASTQ, are uploaded from the browser and counted as they are read. The λ genome uploaded that way on 7 Oct 2026 was read as one record of 48,502 bases, and its analysis gave the August figures again: GC 49.86 %, longest open reading frame 1,132 aa.

Table S3 Alignment run 293c3e5c, c6i.2xlarge.
ItemRecorded
Reads3,370 Oxford Nanopore reads, SRR12602365
ReferenceNC_001416.1, phage λ, 48,502 bp
Mapped99.7626 % of reads
Covered99.9526 % of the genome
Mean MAPQ59.95 of 60
Peak depth202
Supplementary11 alignments
Cross-checka local run agreed on every rate above to four decimals
Instanceterminated after the run — verified, not assumed

S4Science

Publication metadata by DOI, from Crossref: title, authors, venue and date, parsed from the live record.

An abstract is shown only when the publisher deposited one — often it did not — and is otherwise absent, not an empty string. Relevance is never filled in by a look-up: whether a paper matters to your work is a judgement no external service can make.

DOI
10.1103/PhysRevLett.116.061102
Title
Observation of Gravitational Waves from a Binary Black Hole Merger
Venue
Physical Review Letters 116, 061102 (2016)
Authors
1012 — including collaboration entries, which carry a name rather than a family name
Abstract
present, as JATS — deposited by the publisher
Relevance
— never set by a look-up
Fig. S1One DOI, resolved. The record as the science workspace stores it after a look-up. The parser was written against a live Crossref response, not from memory: titles arrive as arrays, collaboration authors carry a name instead of a family name, and dates as nested date-parts.

S5Aerospace

Rocket flight, by Sojist's own solver: written from the physics with Python's standard library and no third-party code. Thrust curve, burning mass, drag against Mach number, the U.S. Standard Atmosphere 1976, gravity falling with height, wind by height, the launch rail and parachutes with their triggers and opening lag. Each flight runs on a machine of its own.

Campaigns fly the same rocket up to 10,000 times, with chosen inputs — thrust, masses, drag, the rail, the wind, each parachute — drawn afresh from a stated distribution. Every scenario keeps the values it was given, so any one can be flown again from the record; one the solver cannot fly is recorded with the reason, not dropped.

The solver has three degrees of freedom: the rocket flies nose into the airflow, as a perfectly stable rocket would. Rotation and stability, the Earth's rotation, computed aerodynamics and parachute inflation are not modelled, and no flight has yet been compared with a measured one. Table S5 is how it was checked instead.

Review decisions on engineering packages are made by a person and attributed: each is signed and time-stamped by the server, not the browser. Automated package checks are not built, so reviews are labelled review only. Sojist never commands flight or actuation.

Table S5 The flight solver against closed-form solutions of its own equations.
CaseRelative difference
Coast in vacuum, apogee1.6 × 10⁻¹⁴
Powered ascent in vacuum, apogee (rocket equation with gravity loss)6.3 × 10⁻¹³
Inclined rail, exit speed1.7 × 10⁻¹⁰
Coast with quadratic drag, apogee9.6 × 10⁻¹⁵
Descent under a parachute, landing time2.7 × 10⁻⁸
Step a quarter as long: apogee, landing time5.5 × 10⁻¹¹, 3.2 × 10⁻⁷
Atmosphere: the 1976 standard's layer-base pressureswithin 10⁻⁵ of the printed values

S6Not built yet

  • Automated aerospace package checks

    Decisions are recorded and attributed; nothing checks the package itself.

  • Modal and buckling analysis

    The solver machine has no frequency or buckling step, so the platform refuses both rather than solving something else under their names.

  • Flight in six degrees of freedom

    Rotation, stability and angle of attack: the flight solver flies the rocket nose into the airflow.

  • A flight compared with a measured one

    The flight solver is checked against closed-form solutions only (Table S5).

  • Payment

    Plans, quotas and usage tracking exist; a plan can be chosen and not yet paid for.