Skip to content

SKILL.md

Latest commit

 

History

History
54 lines (44 loc) · 2.56 KB

SKILL.md

File metadata and controls

54 lines (44 loc) · 2.56 KB
Raw
name simba-prior-conventions
description Simba prior-override payload conventions for create_model — smart-default merging, strict field rejection, the half-saturation/half-marginal/half-life anchor families and which combinations are invalid. Use before constructing any priors[] override through the Simba MCP tools.

Simba prior payload conventions

The merge model

Priors are smart defaults (cost shares, industry benchmarks via total_media_effect, channel-type detection) with per-channel overrides merged ON TOP. Each priors[] entry names a channel (matching channels[].name) plus ONLY the fields to override — everything else keeps its smart default.

Strict rejection (#630)

Unknown keys in a priors[] entry are rejected with a 400 naming the field — they used to be dropped silently, fitting a hybrid model. Common misses: beta/beta_mean → mean, beta_sd → sd, sat_shape → sat_shape_mean. name and parameter are rejected too. The same strictness applies to config.sampler and config.var_priors — but NOT to the request root or config itself, where unknown/misplaced keys are silently ignored (so placement mistakes fit a wrong model without error).

Anchor families — pick ONE per concern

Carryover (adstock):

  • Preferred: half_life_lower/half_life_upper (periods until effect halves) over the legacy decay_lower/decay_upper.
  • theta_mean/theta_sd only with adstock_type="delayed"; dual_weight_mean/dual_weight_sd only with "dual_geometric".
  • Adstock types are geometric, delayed, dual_geometric — there is no power-law adstock.

Saturation:

  • Preferred: half_saturation_mean/half_saturation_sd — the 50%-of-maximum-response point in the channel's ACTIVITY units. Cannot be combined with the legacy alpha_sd/scalars pair in the same override.
  • sat_shape_mean/sat_shape_sd only with saturation_type="generalized_log" (small = near-logarithmic, 1.0 ≈ michaelis_menten).
  • half_marginal_mean/half_marginal_sd (generalized_log ONLY): the activity level where MARGINAL returns have halved. Use it instead of half_saturation at near-logarithmic curvature — the 50% point overflows below sat_shape ≈ 0.00098 and 400s, while the half-marginal point is finite at every shape. Cannot be combined with the other two anchors.

Reading posteriors back

get_model_results sections posterior_transforms (the importable transform-parameter grid, keyed by activity column — join via channel_map) and posterior (94% HDIs) close the prior → posterior loop for the next fit.