ST-E1 spherical tokamak — Tokamak Energy · 3 research iterations · 11 sources
Each source is fetched, extracted, and filtered by the pipeline before any human reads it. A decisions.json records page-level filtering — pages with garbled equations or failed table detection are flagged claude_replace and reconstructed. Click any source card to see what it contributed and why it was kept.
All sources
Iteration 1
Iteration 2
Iteration 3
After 3 iterations: 10 of 12 taxonomy columns at high or medium-high confidence.
Two gaps — thermal cycle type and Q value — are confirmed proprietary. The pipeline marks these as unlikely to resolve without direct company disclosure and stops iterating.
Iterative dossier — how each column evolved across 3 iterations
After each research iteration, a synthesis agent merges findings into the dossier using strict merge rules: higher-confidence values upgrade lower ones; conflicts are surfaced explicitly rather than silently resolved; no high-confidence value is downgraded without a direct authoritative contradiction. The changelog is the complete audit trail.
Final dossier — 12 taxonomy columns
Gap report — readiness for LCOE modelling
The gap report asks: is what we have sufficient to run a cost model? It classifies each missing parameter by gap type, criticality, and whether an analogue exists. The overall verdict here is "Mostly Ready" — all blocking gaps can be bridged with clearly-flagged analogues.
model_setup.py — parameters passed to 1costingFE
Before model_setup.py is written, the gap report must be manually reviewed and approved for each blocking gap. Items marked as unresolvable are either set to framework defaults or explicitly omitted. These inputs are then passed to 1costingFE, a standardized fusion costing framework that models the full LCOE using a bottom-up accounting methodology consistent with Nuclear industry standards.
HIGH — published valueUNCERTAIN — analogueDEFAULT — frameworkNOT MODELLED
Known model biases: blanket volume assumes full 4π coverage — overestimates CAS22 blanket cost (conservative bias).
Pulsed-operation thermal energy storage buffer is not modelled — a downward bias on total capital of unknown magnitude.
model_setup.py passes geometry, efficiency parameters, and operating requirements into 1costingFE, which closes the power balance internally before computing costs. The model runs through a sequence of computation stages: power balance (dispatched by confinement family, producing fusion power, Q_eng, and recirculating fraction), geometry (radial build to component volumes), capital costs (CAS10–60, with CAS22 disaggregated into sub-accounts for blanket, magnets, heating systems, vacuum, and others), and O&M and levelized economics (CAS70–90). Where the research pipeline could not characterize a parameter — radial build thicknesses, auxiliary power loads, cost account scaling factors — 1costingFE loads concept-specific defaults from its YAML configuration, and the corresponding parameter is tagged DEFAULT in the input file. The sliders below adjust three high-elasticity inputs and recompute LCOE using the actual JAX autodiff elasticities from the model run.