{
  "artifact": "spend_ledger",
  "sprint": "meta / S52",
  "claim": "This program's GPU spend, re-derived from artifact wall-clock time and the GPU each run actually got, rather than recalled.",
  "method": "For every certificate carrying an observed gpu_name_ground_truth and a top-level wall_s, cost = wall_s/3600 * the Modal on-demand rate for that SKU. Serverless containers bill by wall-clock second, so container wall time is the basis, not the sum of per-request gpu_seconds. Artifacts that show a GPU and cannot be priced are listed, never silently dropped.",
  "pricing_table": "pricing/2026-08.yaml",
  "pricing_verified": false,
  "rates_used": {
    "T4": 0.59,
    "L4": 0.8,
    "A10G": 1.1,
    "L40S": 1.95,
    "A100-40GB": 2.78,
    "A100-80GB": 3.72,
    "H100": 4.56
  },
  "n_artifacts_priced": 24,
  "n_artifacts_with_declared_cost": 3,
  "n_artifacts_gpu_seen_but_unpriceable": 3,
  "unpriceable": [
    {
      "artifact": "s01_variance_levels",
      "gpus": [
        "NVIDIA A10"
      ],
      "why": "no rate for ['NVIDIA A10']"
    },
    {
      "artifact": "s01_variance_levels_L1",
      "gpus": [
        "NVIDIA A10"
      ],
      "why": "no rate for ['NVIDIA A10']"
    },
    {
      "artifact": "s48_connector_smoke",
      "gpus": [
        "NVIDIA L40S"
      ],
      "why": "GPU observed but no top-level wall_s to price"
    }
  ],
  "priced": [
    {
      "artifact": "d1_engine_hit_rate.attempt1_no_counters",
      "gpu": "NVIDIA L40S",
      "sku": "L40S",
      "rate_usd_hr": 1.95,
      "wall_s": 1587.8,
      "cost_usd_lower_bound": 0.8601
    },
    {
      "artifact": "d1_engine_hit_rate.attempt2_timestamp_as_counter",
      "gpu": "NVIDIA L40S",
      "sku": "L40S",
      "rate_usd_hr": 1.95,
      "wall_s": 1844.9,
      "cost_usd_lower_bound": 0.9993
    },
    {
      "artifact": "d3_determinism_2x2",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 895.1,
      "cost_usd_lower_bound": 0.1989
    },
    {
      "artifact": "d4_step_law_profile.attempt1_no_profiler",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 153.4,
      "cost_usd_lower_bound": 0.0341
    },
    {
      "artifact": "d4_step_law_profile.attempt3_zero_events",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 1052.9,
      "cost_usd_lower_bound": 0.234
    },
    {
      "artifact": "d4_step_law_profile",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 958.9,
      "cost_usd_lower_bound": 0.2131
    },
    {
      "artifact": "d8_expert_rank",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 24.0,
      "cost_usd_lower_bound": 0.0053
    },
    {
      "artifact": "s01_aa_null_sharded",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 412.1,
      "cost_usd_lower_bound": 0.0916
    },
    {
      "artifact": "s01_aa_null_tiny",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 505.7,
      "cost_usd_lower_bound": 0.1124
    },
    {
      "artifact": "s01_positive_control",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 609.9,
      "cost_usd_lower_bound": 0.1355
    },
    {
      "artifact": "s01_positive_control_eager",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 469.4,
      "cost_usd_lower_bound": 0.1043
    },
    {
      "artifact": "s01_variance_levels_L2",
      "gpu": "NVIDIA A10G",
      "sku": "A10G",
      "rate_usd_hr": 1.1,
      "wall_s": 1779.2,
      "cost_usd_lower_bound": 0.5436
    },
    {
      "artifact": "s01_variance_levels_L3",
      "gpu": "NVIDIA A10",
      "sku": "A10G",
      "rate_usd_hr": 1.1,
      "wall_s": 589.3,
      "cost_usd_lower_bound": 0.1801
    },
    {
      "artifact": "s02_paired_knee",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 1130.4,
      "cost_usd_lower_bound": 0.2512
    },
    {
      "artifact": "s02_sweep_sharded",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 571.0,
      "cost_usd_lower_bound": 0.1269
    },
    {
      "artifact": "s02_sweep_tiny",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 897.2,
      "cost_usd_lower_bound": 0.1994
    },
    {
      "artifact": "s02_sweep_tiny_cpu32",
      "gpu": "NVIDIA L4",
      "sku": "L4",
      "rate_usd_hr": 0.8,
      "wall_s": 994.8,
      "cost_usd_lower_bound": 0.2211
    },
    {
      "artifact": "s28_crossgpu_A10G",
      "gpu": "NVIDIA A10G",
      "sku": "A10G",
      "rate_usd_hr": 1.1,
      "wall_s": 427.1,
      "cost_usd_lower_bound": 0.1305
    },
    {
      "artifact": "s28_crossgpu_L40S",
      "gpu": "NVIDIA L40S",
      "sku": "L40S",
      "rate_usd_hr": 1.95,
      "wall_s": 313.6,
      "cost_usd_lower_bound": 0.1699
    },
    {
      "artifact": "s46_apc_real_trace",
      "gpu": "NVIDIA L40S",
      "sku": "L40S",
      "rate_usd_hr": 1.95,
      "wall_s": 1742.5,
      "cost_usd_lower_bound": 0.9439
    },
    {
      "artifact": "s46b_apc_fixed_bridge",
      "gpu": "NVIDIA L40S",
      "sku": "L40S",
      "rate_usd_hr": 1.95,
      "wall_s": 1872.6,
      "cost_usd_lower_bound": 1.0143
    },
    {
      "artifact": "s46c_apc_fixed_bridge_n8",
      "gpu": "NVIDIA L40S",
      "sku": "L40S",
      "rate_usd_hr": 1.95,
      "wall_s": 2083.7,
      "cost_usd_lower_bound": 1.1287
    },
    {
      "artifact": "s47_aa_null_real_trace",
      "gpu": "NVIDIA L40S",
      "sku": "L40S",
      "rate_usd_hr": 1.95,
      "wall_s": 1796.2,
      "cost_usd_lower_bound": 0.9729
    },
    {
      "artifact": "s47b_aa_fixed_bridge_n8",
      "gpu": "NVIDIA L40S",
      "sku": "L40S",
      "rate_usd_hr": 1.95,
      "wall_s": 4110.0,
      "cost_usd_lower_bound": 2.2262
    }
  ],
  "declared": [
    {
      "artifact": "d2_positions_per_pass",
      "cost_usd": 0.0033,
      "basis": "declared on the artifact"
    },
    {
      "artifact": "sprint28_heterogeneous",
      "cost_usd": 0.3,
      "basis": "declared on the artifact"
    },
    {
      "artifact": "sprint8_connector_api_probe",
      "cost_usd": 0.01,
      "basis": "declared on the artifact"
    }
  ],
  "derived_usd": 11.0973,
  "declared_usd": 0.3133,
  "total_usd_lower_bound": 11.4106,
  "budget_authorised_usd": 15000.0,
  "fraction_of_budget": 0.000761,
  "is_lower_bound": true,
  "previously_reported_recalled_usd": 85.0,
  "recalled_over_derived_ratio": 7.45,
  "recalled_figure_is_unsupported": true,
  "verdict": "**$11.41 is a LOWER BOUND on this program's GPU spend, and it is the first such figure derived from evidence rather than recalled.** Of 84 certificates, exactly 3 carry a non-zero `cost_usd`; every paid GPU run since -- S46, S47, S48, S53, the S44p ladder -- carries `cost_usd: None`, in violation of \u00a712's per-artifact requirement. The ingredients were always present (top-level `wall_s`, per-arm `gpu_name_ground_truth`); **the multiplication was never done.** **3 artifact(s) show a GPU and are NOT priced here** -- NVIDIA A10, NVIDIA L40S has no rate in the table, and one run records a GPU with no top-level `wall_s`. They are listed rather than dropped, and their absence pushes the total DOWN, which is the safe direction for a floor. **It is a floor, not an estimate.** `wall_s` covers the measured window only: container boot, image pull and model load are billed before it starts. And runs that crashed before writing an artifact left no `wall_s` to count -- S46's first attempt saturating at the ladder floor, the OOM on the variance ladder -- **money spent on a failed run is still spent.** The real bill is larger and this does not estimate by how much. **Every figure is stamped `pricing_verified: false`.** `pricing/2026-08.yaml` carries inherited placeholder rates, not sourced quotes; its own header records that the estate held five mutually inconsistent copies of `GPU_USD_HR` differing by up to 50% for the same SKU. An unverified rate must not silently become a claim, which is why `bench/economics.py` refuses to quote one and why this artifact carries the flag rather than a footnote. **AND IT CONTRADICTS WHAT WAS BEING REPORTED.** Every status report in this session quoted a running total of roughly **$85**. That figure was never derived from anything; this floor is **$11.41**. The known omissions -- container boot, model load, and runs that crashed before writing an artifact -- do not plausibly account for an ~7x gap on runs of this length and SKU. **The recalled figure should not be used.** Settling it needs Modal's own billing, which is a one-line owner action and not a measurement this repo can make. **The useful part is the direction.** Against a $15,000 authorisation the floor is 0.08%. Underspend of that magnitude is not thrift -- it is the finding \u00a720.2 already recorded, that this repo grew an instrument where a product should be, expressed in dollars for the first time.",
  "does_not_prove": [
    "THIS IS A FLOOR. wall_s covers the measured window; container boot, image pull and model load are billed before it. Failed runs that never wrote an artifact contribute nothing here and cost real money.",
    "pricing_verified is false. The rates are inherited placeholders whose own header records five mutually inconsistent copies in the estate, differing by up to 50% for the same SKU. No dollar figure here may leave the repo as a claim.",
    "Modal's billing granularity and any minimum-duration or idle-timeout charges are not modelled. The basis is wall_s times a published on-demand rate and nothing else.",
    "The gap between the recalled ~$85 and this $4.49 floor is NOT resolved here. This artifact establishes that the recalled figure has no derivation and that the evidence in the repo supports a much smaller floor. Only Modal's billing record settles the actual number.",
    "This measures OUR cost of experimenting. It is not a customer's cost basis and must never be used as one -- the pricing file's own framing rule says a serverless retail rate is not an inference provider's cost."
  ],
  "cost_usd": 0.0,
  "generated_utc": "2026-09-03T00:33:03.543012+00:00",
  "git_rev": "093ddaf",
  "git_dirty": true,
  "run_id": "8eeedf82897c455b",
  "host_kind": "Darwin-arm64",
  "provenance_schema": "gpu-cert-provenance/1"
}