How much should you trust
this number?
A single model can be calibrated. TerraGuard goes a level above: it supervises an ensemble of heterogeneous models, time-series forecasters, vision classifiers, sensor regressions, and forces them all to express belief in one common currency. Calibrated confidence. This is the entire intellectual signature of the platform.
Push the three trust signals and watch the Core decide. This runs the exact classifyTrust() function that governs every sphere on the platform.
Inputs in-distribution, members in agreement, interval tight — the Trust Core stands behind this number.
Every sphere model is wrapped and governed
Calibrated uncertainty
Each model's held-out residuals are ranked; the ⌈(n+1)(1−α)⌉-th gives a quantile q̂ that wraps every prediction in an interval [ŷ−q̂, ŷ+q̂] with marginal coverage ≥ 1−α, and no distributional assumptions. It is the exact technique from the F10.7 research, promoted to a universal interface every sphere shares.
Explainability
Each forecast decomposes into named, signed drivers, recent flare activity, fire-season phase, monsoon dilution, canopy decline. A reader sees not just the number but the reasoning, in plain language. The four-layer XAI suite from the solar research becomes the platform standard.
Disagreement & OOD
Five heterogeneous members forecast each value. When their spread exceeds the error we already expect, or when the latest input drifts far from the training distribution, the Core registers it, before it ever reaches a human.
Abstention
When disagreement, OOD, or interval width crosses threshold, the Core refuses to assert. It flags its own output as untrustworthy and drops out of the fusion rather than poison the index. The direct descendant of AI Guard's refuse-or-warn behavior.
The interval is a promise, not a guess
Most student projects report a single number. A few add error bars pulled from a Gaussian assumption that rarely holds. Split conformal prediction makes a different kind of claim: across many predictions, the true value lands inside the band at least 90% of the time, guaranteed by the data itself, not by a distribution we hoped for.
The width of that band is the honesty. A tight band is a confident model; a band that blows wide is the model telling you it does not know, and the moment the Trust Core considers abstaining.
scores = |y_cal − ŷ_cal|
q̂ = quantile(scores, ⌈(n+1)(1−α)⌉ / n)
interval = [ŷ − q̂, ŷ + q̂]
Most strong applicants present a model. TerraGuard presents a system that governs models, AI Guard's architecture, generalized from supervising language models to supervising a model of the world.
That is a graduate research posture, uncertainty quantification, model governance, and explainability, expressed as a working product that audits itself in public. The Trust Core is not bolted on. It is the layer that gives the platform its name.