Directorate subtracts every known cause; what remains is filed as the finding

1. STATIONS. Five collectors: Vast.ai (REST), RunPod (GraphQL), AWS EC2 Spot (boto3), Azure retail catalog, GCP catalog. Timer fires 08:00 and 20:00 UTC, persistent, with a randomized delay. A collector that fails returns an empty list; one outage cannot fail a cycle.

2. THE DUAL QUERY. Vast is asked twice per run: once for on-demand, once for interruptible bids. The payload's is_bid field reads FALSE on every offer of BOTH queries (verified across 315,685 stored rows), so commitment type is derived from WHICH ENDPOINT ANSWERED, not from the data. A machine listed by both queries yields two observations at two prices: two distinct products, not duplicates.

3. THE CAP. Since 2026-06-23 unauthenticated Vast responses are silently capped at the 64 cheapest offers, structurally hiding the premium tier. Collection is authenticated.

4. CUSTODY. raw_observations stores the provider's complete response (~95 fields per Vast offer) as write-once JSONB plus five extracted columns. Derived tables (canonical_offers onward) are deleted and regenerated whenever the rules change.

5. POPULATION RULE (ADR-0007). Analytics reads only cron-collected rows; rows flagged as backfill are excluded. Precedent: a 90-day AWS backfill made pre-boundary days carry ~57 rows and post-boundary days ~30; pooled, this produced an apparent decline that was purely a change in who was in the sample. Caught before publication.

1. GPU SKU. Exact-string dictionary, ~190 hand-written entries → 96 canonical SKUs, format {model}_{form_factor}_{vram}gb. SXM / PCIe / NVL are different SKUs; 40GB / 80GB are different SKUs. Unmatched → skipped and counted (skipped_unknown_gpu). No fuzzy matching, no case folding.

2. COMMITMENT. Eight canonical values: on_demand, spot, reserved_1w/1m/3m/6m/1y/3y. Alternates map (interruptible, bid, preemptible → spot). Unknown non-null strings pass through unchanged as their own group. DISCLOSURE: a null commitment is coerced to on_demand, the one inference in the layer.

3. REGION. Per-provider lookup maps → (country, state, city). ONLY COUNTRY enters the model; state and city are stored, unused. RunPod reports no region → UNKNOWN.

4. BUNDLE. vCPU / RAM / storage extracted for marketplaces only (Vast, TensorDock); hyperscalers and RunPod contribute none → UNKNOWN. Bundle sits LAST in the fixed order, so its provider-shaped missingness is absorbed by the provider term before bundle takes its turn.

5. PROVENANCE. Each module carries a parallel explain_* trail (matched key, transformation, source value). An AST-level test fails the build if the write path references any explain identifier. Endpoint: /api/raw-observation/{id}/explain.

1. Grouped twice: per (date, sku) across all providers, and per (date, sku, provider). Outputs: observation_count, median_price, p25_price, p75_price.

2. MINIMUM 3 OBSERVATIONS per bucket, else the bucket is skipped entirely.

3. Percentiles, never mean ± sd: quote distributions are asymmetric, and a single mispriced line would average the price somewhere no buyer could transact.

Per (date, gpu_sku) with n ≥ 5:
    y        = ln(price_usd_per_hour)
    total_ss = Σ (yᵢ − ȳ)²
    for factor in (region, commitment, provider, bundle):  # fixed order
        keys = factor values, NULL → "UNKNOWN"
        if distinct(keys) < 2: attribute 0; continue
        ss_within = Σ (yᵢ − mean within CROSS-PRODUCT of prior keys)²
        attributed[factor] = max(0, previous_ss_within − ss_within)
    residual = max(0, total_ss − Σ attributed)
    # each component / (n − 1) to become a variance

SEVEN STANDING DECISIONS: (1) LOG PRICES: ratios, not dollar gaps. (2) SEQUENTIAL, NOT SIMULTANEOUS: conditioning is on the cross-product of all prior factors. (3) FIXED ORDER, most-exogenous → most-internal: region → commitment → provider → bundle. (4) ORDER CHANGES THE FACTORS, NEVER THE RESIDUAL. (5) NULL IS A GROUP: UNKNOWN participates as a real level. (6) A factor with fewer than 2 distinct values contributes exactly zero. (7) max(0,·) flooring: numerically safe.

The exhibit above runs this exact arithmetic, in the reader's browser, on the day's simulated panel. In production it runs on the live panel and every share can be walked back to the contributing offers.

1. MARKET-PRICED is defined as the panel excluding the administered catalogs (azure, gcp). The exclusion RECOMPUTES the decomposition through the same production function the nightly job uses; it does not filter a cached result. Endpoint: /api/basis/{sku}/timeseries?exclude_providers=azure,gcp.

2. THE BOUND. XGBoost on ~45 observable features, day-based expanding-window CV, a 10-day untouched holdout, a hard-failing permuted-target leakage guard, and day-demeaned R² for parity with the ANOVA. Result: negative gap: the rich model, tested out-of-sample, explains less than four coarse factors claimed in-sample.

3. HOST ICC ≈ 0.55, Vast on-demand panel only, ≥10-day tenure primary, recomputed at ≥5 and ≥20 with a 0.15 swing alarm. It says the effect is persistent; it does not say why.