Repository navigation
Trust Atlas
The Trust Atlas is a partition of the model's parameter domain into axis-aligned cells. Each cell accumulates observations, risk statistics, evidence-channel masks, and whether any absolute fact occurred. Leaf cells receive one of three labels under a recorded policy.
- TRUSTED — sampled enough under the configured policy, with no suspicious evidence. It means no current evidence of a problem under the tested assumptions and evidence model, not proof of correctness.
- SUSPICIOUS — the cell meets the risk threshold and contains an absolute fact or sufficiently corroborated evidence.
- UNKNOWN — sampling or evidence is insufficient, or risk is elevated but lacks the required corroboration.
The default policy uses suspicious_mean = 0.45, suspicious_peak = 0.7, min_samples = 4, min_channels = 1, max_depth = 12, refine_below = 0.12, and suspicious_channels = 2. These are implementation defaults, not a universal validation standard.
The initial cell spans the declared parameter domains. A finite choice axis is represented by indices so it can participate in the same partition operations. When a cell has enough samples and is suspicious, has sufficient mean risk, or borders a cell with a different label, refinement may split it. The axis selected is the widest in normalised domain coordinates. Refinement stops at the configured depth and respects a sample floor, which prevents one isolated high-risk point from creating a long chain of unsupported cells.
Search-time measurements are useful for choosing new points. At the end of a campaign, the Atlas is remeasured and relabelled from the completed evidence record before findings are collected. This allows whole-run relations, fitted calibration, sensitivity references, and precision comparisons to inform the final map without inventing new executions.
A boundary band is derived where neighbouring labelled cells disagree. The partition is axis-aligned, so it can only approximate curved boundaries with a staircase of boxes. An unresolved or too-narrow region may correctly remain UNKNOWN; a missing band means the current partition did not produce a qualifying disagreement between labelled cells, not that no boundary exists.
Coverage reports the fractions assigned to trusted, suspicious, and unknown cells. These fractions summarise the current partition and policy, not the continuous domain's true correctness. For the measured boundary performance and examples of resolution limits, see Experiments & Results and Limitations.
Home · Evidence Model · Adaptive Search · Archives and Replay