Audit analytical specification multiverse

Audit whether an analytical conclusion survives a prespecified multiverse of admissible windows, cohorts, metrics, and models using aligned bootstrap draws, a weighted specification curve, practical-effect support gates, and descriptive choice-influence diagnostics.

Statistical audit & measurement Audit CTOs Tech CEOs IT managers Team leads Investors

What it's for

Makes 'confident decisions' resistant to hidden researcher degrees of freedom by showing whether the answer survives reasonable metric, cohort, window, and model definitions.

What you give it

Inputs split into evidence read from your connected systems, calibration your team owns, and numerical controls that affect precision but never the result's meaning.

Field	Type	Role	Required
confidence_level	number ≥ 0.8, ≤ 0.99	Your calibration	Optional
direction	one of "higher_is_better", "lower_is_better"	Your calibration	Yes
max_detail_rows	integer ≥ 1, ≤ 500	Numerical control	Optional
minimum_practical_effect	number ≥ 0	Your calibration	Optional
minimum_specification_support	number ≥ 0, ≤ 1	Your calibration	Optional
minimum_within_specification_probability	number ≥ 0, ≤ 1	Your calibration	Optional
specifications	array of objects (5 fields) ≥ 8 items	Evidence	Yes

Each `specifications` record

Field	Type	Required
choices	object	Yes
estimate_draws	array of number (≥ 200 items)	Yes
id	string (non-empty)	Yes
resampling_plan_id	string (non-empty)	Yes
weight	number (> 0)	Optional

Example input

{
  "direction": "higher_is_better",
  "minimum_practical_effect": 0.5,
  "specifications": [
    {
      "choices": {
        "model": "ridge",
        "window": "24-weeks"
      },
      "estimate_draws": [
        1.2,
        1.2673176787846316,
        1.2727437941460544,
        1.2112896006447893,
        1.1394558003753656,
        1.1232860580269488,
        1.1776467601440859,
        1.2525589278975031,
        1.2791486597298705,
        1.2329694788193404,
        1.1564783111288504,
        1.1200007834759438,
        1.157074166559965,
        1.2336133629461312,
        1.2792485884555895,
        1.2520230272125694,
        1.1769677346667948,
        1.1230882006496354,
        1.139921020258266,
        1.211990176773036,
        1.2730356200582102,
        1.2669324510828845,
        1.1992918952567677,
        1.1323023676659862,
        1.12755373103947,
        1.1894118599921781,
        1.2610046760383682,
        1.2765100742723603,
        1.2216724630646294,
        1.1469092892629627,
        1.120957470072571,
        1.1676769883741547,
        1.2441141344993352,
        1.2799929488085813,

Truncated for display — the full payload is 1686 lines.

What you get back

This is the actual output of running the example above — computed by the same function the platform calls, not an illustration.

Example output

{
  "assumptions": [
    "Every specification was declared admissible before results were inspected and differs only along the recorded choice axes.",
    "Draw index identifies the same resampled organization units or time blocks across every specification, preserving cross-specification dependence.",
    "Choice influence is a descriptive main-effect variance share; correlated analytical choices prevent causal attribution to one choice axis."
  ],
  "choice_influence": [
    {
      "choice": "window",
      "levels": [
        {
          "level": "12-weeks",
          "specifications": 4,
          "weighted_mean_directed_effect": 1.4
        },
        {
          "level": "24-weeks",
          "specifications": 4,
          "weighted_mean_directed_effect": 1.3501
        }
      ],
      "variance_share": 0.0474
    },
    {
      "choice": "model",
      "levels": [
        {
          "level": "fixed-effects",
          "specifications": 5,
          "weighted_mean_directed_effect": 1.39
        },
        {
          "level": "ridge",
          "specifications": 3,
          "weighted_mean_directed_effect": 1.3501
        }
      ],
      "variance_share": 0.0284
    }
  ],
  "configuration": {
    "aligned_draws": 200,
    "confidence_level": 0.9,
    "direction": "higher_is_better",

Truncated for display — the full payload is 188 lines.

How it works

Statistical audit & measurement — Check whether a number is fit to decide on: coverage, timing, reconciliation, and the gaps a dashboard hides.

1 Require every admissible specification to use the same resampling plan and aligned draw index so shared teams or time blocks preserve cross-specification dependence.
2 Orient effects to the governed favorable direction, calculate within-specification probability of clearing the practical threshold, and build the weighted specification-support rate.
3 Take a weighted median across specifications inside every aligned draw to obtain a robust aggregate interval and support probability without pretending specifications are independent studies.
4 Measure how observed effect variation aligns with each declared analytical-choice axis and retain the full specification curve instead of publishing only the preferred pipeline.

Before you trust it

Every tool in the catalog ships with the conditions under which its answer is meaningful — and the conditions under which it should abstain instead of guessing.

Assumptions & guardrails

Metric definitions, weights, aggregate grain, sampling, missingness, dependence, and comparison windows correspond to the management claim being audited.
All specifications were declared scientifically admissible before inspecting their outcomes; including intentionally bad models or omitting inconvenient reasonable models invalidates the robustness claim.
Each draw index resamples the same organization units or time blocks across all specifications.
Association, instability, or measurement quality is not a causal effect and must not be converted directly into an individual employment decision.
Robustness across analyst choices does not create causal identification, repair a biased dataset, or prove the outcome matters economically.

Minimum evidence

specifications: at least 8 rows/items
direction: required and organization-defined

How to validate it

Validate on future periods or held-out aggregate units, compare with a simple baseline, and require stability across plausible metric definitions and decision thresholds.

Calibrating it to your org

Same for everyone

The mathematical kernel, validation rules, method version, and JSON output semantics are organization-independent; no tenant-trained coefficients or company benchmark is embedded in the function.

Specific to you

one effect draw per specification and aligned resample index
complete declared analytical-choice vector per specification
admissible-specification manifest and optional prespecified weights
favorable direction and minimum practical effect
within-specification and across-specification support thresholds
scientific admissibility policy fixed before outcomes are inspected

Calibration workflow

1 Define the management decision, target outcome, aggregate unit, privacy boundary, cadence, and prediction/intervention horizon for this organization.
2 Build a tenant-scoped historical cohort using only information available before each prediction or decision; preserve zero periods, censoring, assignment probabilities, and unresolved outcomes when the method requires them.
3 Estimate statistical parameters on training history, but obtain costs, utilities, risk tolerance, practical-effect thresholds, capacity, and policy constraints from accountable decision owners.
4 Validate on later time windows or held-out aggregate units at the deployment grain, against a simple baseline and the function-specific validation strategy.
5 Deploy only if the returned decision clears evidence, overlap, calibration, robustness, and guardrail checks; warning, unsupported, schema-gap, and fallback decisions are abstentions.
6 Monitor realized outcomes, data drift, coverage, and decision regret; recalibrate at a governed cadence or after a detected regime/definition change, never merely because a stakeholder dislikes the result.

Call it from your AI

You don't wire up 388 tools in your MCP client. The GitRevio MCP server exposes 18 tools, three of which let an agent search the catalog, read a tool's schema, and run it — so the assistant finds this one on its own.

gitrevio_capabilities_search
  { "q": "audit whether an analytical conclusion survives" }
  → finds "audit_analytical_specification_multiverse"

gitrevio_capability_describe
  { "capability_id": "audit_analytical_specification_multiverse" }
  → returns the input schema and agent guidance shown on this page

gitrevio_capability_run
  { "capability_id": "audit_analytical_specification_multiverse", "arguments": { ... } }
  → returns the result shown above

Works in Claude Desktop, Claude Code, Cursor, Cline, Continue.dev, Goose and Aider. See the MCP server.

Related tools

Solve bayesian influence diagram

Solve an exact discrete Bayesian influence diagram over actions, chance-node DAGs, action-dependent conditional probabilities, pre-decision evidence, and additive utility tables, then quantify action regret and the expected value of perfect information for observable exogenous nodes.

Sequential Bayesian & bandits

Value AI assistant rollout ROI

Value an aggregate AI-assistant rollout from aligned joint causal-effect draws, preserving delivery/time/defect/incident dependence while enforcing identification, out-of-time, overlap, metric-integrity, effective-sample, quality-harm, NPV, ROI, and payback gates.

Causal inference & experiment design

Audit analytics challenger independence

Audit whether an analytical challenger supplies genuinely independent error information: use paired temporal moving-block bootstrap bounds on error correlation, incumbent-failure catch rate and common-mode joint failure, with simultaneous Bonferroni control across every screened challenger and explicit evidence gates.

Statistical audit & measurement

Audit analytics function calibration readiness

Gate analytical functions on paired out-of-time decision loss against a frozen baseline using temporal moving-block bootstrap, autocorrelation- and weight-adjusted effective sample size, evidence coverage, lower confidence bounds, improvement probability and recent degradation rather than declaring a model calibrated from training fit.