The platform turns raw acquisition data into a single engineering record per test — every channel visualized, every derived parameter computed, every deviation flagged, with a plain-language explanation on top. Deterministic calculation produces the facts; the language model explains and assembles them, and never states an engineering conclusion without a computed metric behind it. The result is not a black box — it is a record you can defend in a review.
A firing does not return a summary page. It returns a descriptive report of sixteen tabs, a diagnostic report of ten, and eight campaign sections that place the run inside the series — the same structure every time, so the tenth firing is read exactly the way the first one was.
Every tab is generated from the same run, by the same core, in one pass. Nothing here is assembled by hand after the test.
The diagnostic report adds ten tabs of its own and the campaign report eight sections. Nothing on this list is optional: where the data for a tab is missing, the tab states the reason instead of disappearing.
Together these cover what a firing can be asked: what happened, how well the engine performed, how it started and held, and what the series says. Each area is a section of the report, not a feature on a list.
Every acquired channel on one time base — thrust, chamber and feed pressures, propellant temperatures, mass flow, valve states — with statistics computed separately for the ignition transient, the steady-state window and shutdown, never averaged across them. An event timeline marks each command and each flag, and stand video is locked to the same clock, so a frame can be put next to the moment it belongs to.
Before any conclusion, the report states what was measured reliably: which channels were present for the whole firing, where samples are missing, and where a calibration record could not be confirmed. A parameter that depends on an unverified input is marked unavailable rather than estimated — the boundary of the measurement is part of the result.
Inspection findings are attached to the run they belong to: what the hardware looked like after the firing, beside the data that preceded it. Erosion, deposits or a changed throat dimension become part of the record instead of living in a separate photo folder that nobody opens again.
Characteristic velocity and its efficiency against a thermochemical baseline computed for your propellant pair and chamber conditions; thrust coefficient, specific impulse and mixture ratio tracked through the burn. Geometry, units and assumptions are printed next to every result, so a reviewer can check the derivation rather than trust the figure.
Line pressures and mass flows on both sides of the injector, and the valve sequence as commanded against the sequence as executed. A late or slow valve appears as a timing difference with a number on it, not as an unexplained step in a pressure trace.
The start is classified from the chamber-pressure trace: delay from valve command, peak-to-steady overshoot, and the ignition class that follows. Every flag carries the channel and the moment that raised it, so the classification can be argued with instead of accepted.
Temperature gradients in the oxidizer line before the valve opens, read as a precursor rather than a footnote. A line that is not cold enough at the command is one of the few things a hard start announces in advance — if someone is looking at the right window.
Thrust variation inside the steady-state window, its spectrum, and its coherence with feed pressure — which separates feed-coupled oscillation from combustion-driven behaviour. The resolved band is bounded by your acquisition rate and stated as such, never implied.
Wall and nozzle temperatures across the firing, with rates and per-phase maxima. Thermal behaviour is reported by phase, so a slow soak-back after shutdown is not hidden inside an average taken over the whole burn.
Structural response from the accelerometers on the stand and the article, in the same time base and the same windows as thrust — so a structural mode is not read as combustion instability, and the reverse.
Thrust tail-off, how the pressures come down, and the purge that follows. The shutdown window is measured rather than trimmed off, because a meaningful part of hardware life is spent in it.
Run-to-run trends, repeatability across the series and exit criteria for the campaign; and two runs, or two hardware configurations, compared over matched windows. The comparison is computed, not eyeballed across two charts drawn on different scales.
Every parameter below is computed from your own measurements by fixed formulas, with the geometry, units and assumptions shown next to the result. Where an input is missing — an uncalibrated channel, an absent flow measurement — the parameter is marked unavailable rather than estimated.
Common acquisition exports are accepted as written. Channel names are mapped to a schema once per stand, then every later test flows through without manual preparation.
Derived parameters come from fixed formulas with stated geometry, assumptions and units. The same input always produces the same number.
Narrative is assembled from computed values. The language model never produces an engineering number, and no conclusion appears without a metric behind it.
The analyst writes the read of a firing in plain language. It does not compute it. That division is the whole design: the numbers come from the deterministic core, the language comes from the model, and the two are never allowed to swap roles — which is why an answer can be checked instead of believed.
Any figure in an answer resolves this way: metric, the window it was computed over, the channels behind it, and the baseline it was measured against.
A tool that hides its boundaries is harder to trust than one that names them. These are ours.
Reports are engineering analysis. They support a decision made by your team; they are not a compliance statement or a qualification credential.
Measurement quality is bounded by your instrumentation and its calibration. Where calibration data is missing, the affected parameter is flagged — not quietly assumed.
High-frequency combustion behaviour requires fast recording. If your data cannot resolve it, the report says so instead of implying coverage.
The platform compresses the time from firing to understanding. The judgement — fire again, change the sequence, open the hardware — stays with your team.
Send one firing's data — we return a full report.