Inspect a retained run

Each non-dry runner invocation reserves a new timestamped directory under run.output_dir. Read run_manifest.json first: it records stage states, artefact paths, warnings and execution failures. Directory existence alone does not show that a stage ran or produced valid evidence.

Directory	Role
`00_run_metadata`	Resolved configuration, provenance and CLI log
`05_prior_sensitivity`	Prepared alternative configurations
`08_ai_advisor`	Local configuration evidence and rules
`10_pre_diagnostics`	Prior predictive and pre-fit evidence
`20_model_fit`	Fitted `model.nc` and fit evidence
`30_model_assessment`	Posterior predictive assessment
`35_holdout_validation`	Optional fresh-fit predictive evidence
`40_decomposition`	Model component summaries
`50_diagnostics`	Diagnostic reports and resolved gate policy
`60_response_curves`	Adstock and saturation outputs
`70_optimisation`	Reserved stage; no validated YAML optimisation block
`75_prior_sensitivity_fits`	Optional fitted-scenario comparisons
`80_interpretation`	Interpretation inventory
`90_ai_advisor`	Final local or provider-backed evidence review

The CLI retains 00_run_metadata/run.log; failures can also retain 00_run_metadata/error.traceback.txt. The resolved configuration is 00_run_metadata/config.resolved.yaml. Use the manifest’s recorded paths rather than treating a list of filenames from another release as exhaustive.

Stage states distinguish completed, skipped, failed and not-reached work. A run can complete while its diagnostic floor is fail; inspect 50_diagnostics/diagnostics_report.csv separately. Failed runs retain completed stages and error information. A fitted file in a failed run is not equivalent to a completed run accepted by scenario mode.

Optional outputs have explicit limits. Stage 05 alone does not establish prior robustness; stage 75 runs only when fit_scenarios: true. An advisor response is an interpretation of retained evidence, and the optimisation stage is currently skipped because no validated YAML block exists. See prior sensitivity, advisor policy and scenario recipes.

Manifest schema and provenance

run_manifest.json has schema version 2. The fields below describe execution identity; they do not constitute model approval (src/ammm/pipeline/manifest.py:40, src/ammm/pipeline/manifest.py:120).

Fields	Meaning
`run_name`, `timestamp`, `output_dir`	Name, UTC directory timestamp, absolute retained directory
`config_path`, `config_sha256`, `model_class`	Source config identity and resolved class
`data`	Input path/hash entries where configured; preserve actual inputs separately
`overrides`	Applied sampler, sample-count, stage and quick-mode changes
`status`, `finished_at`, `warnings`, `error`	Run lifecycle; error includes failing stage, type and message
`stages`	Mapping from stage key to the record below

Each stage record has directory, status, started_at, finished_at, artifacts, warnings and error. artifacts maps semantic keys to retained paths; resolve those paths from the run directory. Stage states are pending, running, completed, skipped, failed and not_reached. Run states are pending, running, completed and failed (src/ammm/pipeline/manifest.py:10). A skipped directory can exist without usable evidence, and a later exception leaves earlier completed outputs intact.

Retain source/config/data identity, the lockfile and effective environment, all random seeds, study evidence and the analysis rationale for reproduction. Stage 00 records selected package versions and a Git SHA, but does not archive every dependency, data file or uncommitted source change. Its release_status: supported is a static metadata label, not estimator qualification (src/ammm/pipeline/stages/core.py:243, src/ammm/pipeline/stages/core.py:274). The runner uses an integer sampler seed for its predictive/curve context, otherwise 42; fit settings themselves retain their configured sampler semantics (src/ammm/pipeline/runner.py:405).

Files and conditional presence

These filenames are written by the listed stages at the reviewed commit. Read the manifest before opening them because disabled, failed or not-reached stages may leave them absent. Plots are presentation artefacts; retain the numeric source and model for reanalysis.

Directory	Numeric and text evidence	Writer
`00_run_metadata`	`config.original.yaml`, `config.resolved.yaml`, `config.txt`, `config.yml`, `data_dictionary.csv`, `dataset_metadata.json`, `design_matrix_manifest.csv`, `estimator_manifest.yaml`, `estimator_summary.txt`, `git_sha.txt`, `model_metadata.json`, `output.txt`, `session_info.txt`, `spec_summary.csv`; conditional holiday manifest; CLI log/traceback	`src/ammm/pipeline/stages/core.py:148`
`05_prior_sensitivity`	Scenario directories containing `config.resolved.yaml`; plan metadata as recorded in manifest	`src/ammm/prior_sensitivity/scenarios.py:177`
`08_ai_advisor`	`evidence.json`, `rules_summary.json`, `advisor_status.json` when enabled and writing outputs	`src/ammm/pipeline/stages/ai_advisor.py:41`
`10_pre_diagnostics`	Prior-predictive plot, `stationarity_summary.csv`, `transfer_entropy_summary.csv`	`src/ammm/pipeline/stages/core.py:356`
`20_model_fit`	`model.nc`, `posterior_summary.csv`, `trace.png`	`src/ammm/pipeline/stages/core.py:425`
`30_model_assessment`	`posterior_predictive.nc`, `posterior_predictive_summary.csv`, `fitted.csv`, `observed.csv`, `residuals.csv`, `fit_metrics.json` and diagnostic plots	`src/ammm/pipeline/stages/core.py:455`
`35_holdout_validation`	`validation_metadata.json`, `holdout_predictive_report.json`, `holdout_predictive_summary.csv`, `holdout_posterior_predictive.nc`, `holdout_fitted.csv`, `holdout_observed.csv`, `holdout_residuals.csv`, plots	`src/ammm/pipeline/stages/validation.py:85`
`40_decomposition`	`channel_contributions.csv`, `baseline_contributions.csv`, `mean_contributions_over_time.csv`, decomposition plots	`src/ammm/pipeline/stages/core.py:539`
`50_diagnostics`	`diagnostics_report.csv`, `diagnostics_summary.txt`, `diagnostic_gates.resolved.yaml`; design, MCMC, predictive, Bayesian-criteria and calibration reports; VIF/residual summaries	`src/ammm/pipeline/stages/core.py:659`
`60_response_curves`	`adstock_curve`, `saturation_curve`, `forward_pass_contribution_curve`: each `.nc`, `_summary.csv` and `.png`; `all_response_curves.csv`, `current_input_response.csv`, `response_curves.png`	`src/ammm/pipeline/stages/core.py:819`
`70_optimisation`	No optimisation result; reserved stage is skipped	`src/ammm/pipeline/stages/core.py:969`
`75_prior_sensitivity_fits`	`scenario_fit_diagnostics.csv`, `sensitivity_comparison.csv`, `channel_robustness.csv`, `roas_sensitivity.png`; no alternative posterior files	`src/ammm/pipeline/stages/core.py:584`
`80_interpretation`	`interpretation_report.md`, `evidence_inventory.md`	`src/ammm/pipeline/stages/core.py:979`
`90_ai_advisor`	Local evidence/rules/status and `advisor_review.md`; parameter-identification report and local lookup; conditional accepted response and patch/approval records	`src/ammm/pipeline/stages/ai_advisor.py:88`

The local parameter lookup is not a provider-safe evidence bundle. See advisor operations for exact invocation flags and conditional proposal files. A manifest contains what was actually retained, so do not infer a live provider call from the directory name.

Table keys, units and intervals

Dimension columns are coordinate keys: commonly date and channel, with configured panel axes such as geo added where retained. Join by named keys, never by row position. The following schemas describe this writer; optional parameter columns from ArviZ depend on the installed package settings.

File or family	Keys and values	Units / interpretation
`data_dictionary.csv`	`column, role, dtype, missing_count, unique_count`	Input schema audit
`design_matrix_manifest.csv`	`column, role, dtype, variance, nonzero_count`	Raw-input screening, not a complete transformed estimability proof
`spec_summary.csv`	`setting, value`	Model specification summary
`stationarity_summary.csv`	`variable, n, adf_statistic, adf_p_value`	Univariate screening; unavailable tests can have missing values
`transfer_entropy_summary.csv`	`channel, screening_metric, value, interpretation`	Despite its legacy filename, this is lag-zero Pearson correlation screening, not transfer entropy or causality
`posterior_summary.csv`, `mcmc_summary.csv`	`parameter` plus ArviZ statistics	Parameter units depend on scaling; interval names/probability follow the actual ArviZ output, not a guaranteed 94% HDI
`fitted.csv`, `observed.csv`, `residuals.csv`	Observation keys; `fitted_mean`, `fitted_median`, `observed`, `residual` as applicable	Original outcome units; residual = observed minus fitted mean
`posterior_predictive_summary.csv`	Observation keys; `mean, median, abs_error_94_lower, abs_error_94_upper, observed`	Original outcome draws, including observation variation
`channel_contributions.csv`	Observation/channel keys; `mean, median, abs_error_94_lower, abs_error_94_upper`	Original outcome component units
`mean_contributions_over_time.csv`	`component_type`, observation keys, `component` and summary values	Original outcome component units
`baseline_contributions.csv`	`component, mean` plus panel keys where present	Mean per-period baseline contribution, not automatically a horizon total
`diagnostics_report.csv`	`check_id, phase, severity, status, metric, value, threshold, message, overall_status`	Policy outcomes; `skipped` is not `pass`
Metric summaries	`metric, value`	Consult metric definition; null is not zero
Holdout row tables	`date, observed, posterior_mean, posterior_median, residual, lower_50, upper_50, lower_80, upper_80, lower_94, upper_94` or their selected subsets	Original outcome units; equal-tailed predictive intervals
`holdout_predictive_report.json`	`schema_version: 1`, `evidence_type`, `fresh_fit`, `metrics`, `sampling_diagnostics`	Fresh conditional evaluation; review calibration leakage boundary
Stage 75 comparisons	Scenario and coordinate keys, `metric`, reference/scenario means and HDI bounds, `relative_change`, `hdi_overlap`, `robust`	Contribution share or all-time incremental ROAS; flags are declared tolerance rules

Schemas come from src/ammm/pipeline/stages/core.py:196, src/ammm/pipeline/stages/core.py:356, src/ammm/pipeline/stages/core.py:455, src/ammm/pipeline/stages/core.py:539, src/ammm/pipeline/stages/core.py:684, src/ammm/mmm/blocked_holdout.py:178 and src/ammm/prior_sensitivity/comparison.py:74.

The abs_error_* columns contain interval endpoints, not errors or distances from the mean. The summary facade uses HDIs when chain and draw remain separate, but equal-tailed quantiles when only sample remains; retain the dimensions and interval method with exports (src/ammm/mmm/summary/factory.py:240). ArviZ’s parameter summary uses its installed defaults; in the reviewed execution it emitted eti89_lb and eti89_ub. Do not relabel those as 94% HDIs (src/ammm/pipeline/stages/core.py:440).

Stage 60 always summarises curves with mean, median, eti_94_lower and eti_94_upper, denoting 94% equal-tailed intervals. adstock_curve_summary.csv uses time since exposure; saturation uses scaled x plus original-unit input; forward-pass curves use history multiplier sweep plus total original input. all_response_curves.csv unions these columns and adds curve_family, so family-inapplicable fields are missing. current_input_response.csv retains total_input, mean_input, contribution summaries and marginal summaries at sweep=1; its marginal is a derivative with respect to total channel input, which is only monetary when the input is spend (src/ammm/pipeline/stages/core.py:858, src/ammm/pipeline/stages/core.py:948, src/ammm/pipeline/stages/_core_metrics.py:139).

NetCDF and review handoff

model.nc retains the model persistence metadata and canonical fitted inputs alongside inference groups. Predictive and curve .nc files retain draw and coordinate axes but are not standalone fitted models. Inspect each dataset’s variables/dimensions before combining it with another run; save coordinate labels and preserve joint posterior draws when aggregating uncertainty (src/ammm/mmm/persistence.py:62, src/ammm/pipeline/artifacts.py:147).

A review handoff should name the decision and estimand, source/data identities, resolved config and seeds, run/stage states, diagnostic policy/results, holdout design/results, calibration provenance, sensitivity alternatives and proposed allocation constraints. Include a statement of what was not evaluated. The retained files support that review; they do not make the acceptance decision.

Implementation reference at 7cb7f20: src/ammm/pipeline/artifacts.py:71, src/ammm/pipeline/runner.py:156, src/ammm/pipeline/stages/core.py:969.