USAGE:
  sediment report model [-h] [--org ORG] [--since-days SINCE_DAYS] [--json]
                        [--trend] [--target-mde TARGET_MDE]
                        [--completions-per-day MODEL_A_PER_DAY MODEL_B_PER_DAY]
                        [--bootstrap-check]
                        [--bootstrap-iterations BOOTSTRAP_ITERATIONS]
                        [--bootstrap-seed BOOTSTRAP_SEED]
                        [--non-inferiority-margin NON_INFERIORITY_MARGIN]
                        [--peek-fraction PEEK_FRACTION]
                        [--ci-grain {commit,attributed_completion}]
                        [--compare MODEL_A MODEL_B]
                        [--compare-all MODEL [MODEL ...]] [--effect-decay]
                        [--regret] [--database-url DATABASE_URL]
                        [--mirror-path MIRROR_PATH]

  Per-model outcome report — attribution, CI, and acceptance by model.

OPTIONS:
  -h, --help            show this help message and exit
  --org ORG             org id (default: $SEDIMENT_ORG_ID)
  --since-days SINCE_DAYS
                        window inference calls (and their attributed
                        completions) to the last N days by capture time;
                        default: all history
  --json                emit JSON rows instead of a table
  --trend               also bucket attribution_rate and ci_pass_rate into 7-day
                        captured-at windows and run a Mann-Kendall monotonic
                        trend test
  --target-mde TARGET_MDE
                        with --compare, also report the equal-allocation n per
                        arm needed to detect this target effect, expressed as
                        raw proportion points (0.05 means five percentage
                        points)
  --completions-per-day MODEL_A_PER_DAY MODEL_B_PER_DAY
                        with --target-mde, also estimate calendar days using
                        per-arm completions/day in the same order as --compare;
                        zero or negative rates omit the duration line
  --bootstrap-check     with --compare, also run a fixed-seed bootstrap
                        percentile CI diagnostic against the analytic two-
                        proportion CI
  --bootstrap-iterations BOOTSTRAP_ITERATIONS
                        resamples for --bootstrap-check (default: 2000)
  --bootstrap-seed BOOTSTRAP_SEED
                        deterministic RNG seed for --bootstrap-check (default:
                        151)
  --non-inferiority-margin NON_INFERIORITY_MARGIN
                        with --compare, test whether MODEL_B is not worse than
                        MODEL_A by more than this raw proportion-point margin
                        (0.05 means five percentage points)
  --peek-fraction PEEK_FRACTION
                        with --compare, also report an O'Brien-Fleming
                        sequential boundary read for this information fraction
                        (0 < t <= 1)
  --ci-grain {commit,attributed_completion}
                        unit of analysis for ci_linked/ci_passed: 'commit'
                        (default, correct) dedups a commit's file-grained
                        attributed completions to one trial before counting;
                        'attributed_completion' is the legacy per-attributed-
                        completion count, kept only as an escape hatch for
                        explicit comparison — it inflates sample size and is not
                        recommended for inference
  --compare MODEL_A MODEL_B
                        two-proportion z-test of MODEL_A vs MODEL_B on CI pass
                        rate and attribution rate, over the same --org/--since-
                        days window; a model with no data in this window
                        compares as insufficient data, not an error
  --compare-all MODEL [MODEL ...]
                        pairwise two-proportion z-tests for 2+ models with
                        Benjamini-Hochberg FDR control across both metrics and
                        all pairs; a model with no data in this window compares
                        as insufficient data, not an error
  --effect-decay        with --compare, print a regression-to-the-mean
                        diagnostic: Cohen's h at 25/50/75/100% prefixes for CI
                        pass rate and attribution rate, using natural arrival
                        order
  --regret              with --compare-all, add expected regret for the
                        apparent-best model on each metric
  --database-url DATABASE_URL
                        PostgreSQL URL (default: $SEDIMENT_DATABASE_URL)
  --mirror-path MIRROR_PATH
                        base dir for git mirrors (default: $SEDIMENT_MIRROR_PATH
                        or './mirrors')
