SodiumSpan
Choose the next revealing sodium-ion battery experiment—not simply the next available test.
Which test separates time-related loss from cycling-related loss?
When elapsed time and charge throughput move together, more cycling cannot distinguish the explanations. Give the planner a finite budget and let it choose an orthogonal test.
Why the test matters
The starting observations follow one diagonal path: time and throughput rise together. Storage/reference and changed-duty tests move away from that path, making the assumed effects distinguishable inside this model.
What the proof does not claim
No electrolyte discovery, lifetime prediction, laboratory validation, human-speed advantage, or calibrated battery-failure probability.
Six scientific decisions, one reconstructable loop
OmniGent owns orchestration. Each specialist owns a bounded scientific decision and returns structured output.
Literature
Find applicable evidence, contradictions, and gaps.
Insight
Propose competing, falsifiable explanations.
Planner
Choose the most revealing affordable test.
Runner
Execute the exact experiment specification.
Analysis
Interpret uncertainty and reopen assumptions.
Knowledge graph
Preserve evidence, revisions, and lineage.
Seneca
A bounded outside-opinion signal for alternative hypotheses and counterarguments. It cannot create measurements, citations, or scientific authority.
FableGauge
The operator and explanation layer: configure the mission, watch real handoffs, inspect proof, reconstruct decisions, and communicate what changed.
KIT NVP/C–hard-carbon formation record
Source-specific parsing consistency only. It does not establish instrument calibration, electrolyte ranking, lifetime, or degradation mechanism.
Trace contract
EvidenceRecordSource, exact context, applicability, and limitationsExperimentSpecSelected test, cost, alternatives, and stop ruleResultRecordInputs, code hash, observations, units, and statusAnalysisRecordUncertainty, mismatch, support changes, and next-test needsDefinition of done
- Six specialist roles return actual structured outputs.
- Independent work runs in parallel through real handles.
- A numerical budget forces a choice between tests.
- A computed result changes the next scientific decision.
- The knowledge graph retains the original and revised reasoning.
- Seneca opinion stays distinct from measured evidence.
- FableGauge exposes both the judge story and raw proof.
- The complete run repeats from a clean state.
Current boundary
Implemented locally
Public-data parser, synthetic reference experiment, strict scientific helper, bounded native configuration, and this explanatory surface.
Still open
Completed native OmniGent handoffs, live Seneca input, result-dependent graph persistence, measured workflow comparison, and final judge-ready replay.