Mirror of github.com/dimhold/llm-arithmetic data (history squashed after redaction)

.gitattributes

@@ -0,0 +1,60 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,63 @@
+---
+license: mit
+language:
+- en
+pretty_name: llm-arithmetic
+size_categories:
+- n<1K
+task_categories:
+- question-answering
+tags:
+- llm-evaluation
+- arithmetic
+- finance
+- claude
+configs:
+- config_name: rows
+  data_files:
+  - split: train
+    path: results.json
+  field: rows
+---
+
+# llm-arithmetic
+
+Author: Dmitriy Semenkevich (dimhold)
+
+This dataset measures whether 4 Claude models can do 10 pieces of SaaS arithmetic (MRR, ending cash, LTV, runway) with every tool taken away.
+
+Headline, from the repository: exact answers were 10 / 10 for claude-opus-5, 9 / 10 for claude-sonnet-5, 3 / 10 for claude-haiku-4-5 and 9 / 10 for claude-fable-5; all 4 models were within 1% on 10 / 10 (RESULTS.md).
+
+This is a mirror. The canonical home is the Zenodo record and the GitHub repository; criteria, limitations and full tables live there.
+
+- Paper: https://dimhold.by/papers/llm-arithmetic/
+- Code and full write up: https://github.com/dimhold/llm-arithmetic
+- DOI: https://doi.org/10.5281/zenodo.22128841
+
+## Files
+
+| path | what it is |
+|---|---|
+| `results.json` | every call: tasks with expected values computed in code, and one row per model and task under `rows` |
+| `transcript.md` | every question and every full reply |
+| `RESULTS.md` | the result tables |
+
+## How it was produced
+
+Run on 2026-08-16 through the Claude Code CLI with built in tools and MCP servers disabled. Tasks are generated from seed 20260812, ground truth is computed in `run.ts` and never by a model, and the answering model is verified against `modelUsage`. The scripts that turn these files into every number are in the GitHub repository.
+
+## Licence
+
+MIT, the licence of the repository, covering code and data.
+
+## Citation
+
+```bibtex
+@misc{semenkevich2026llmarithmetic,
+  author = {Semenkevich, Dmitriy},
+  title  = {llm-arithmetic: can Anthropic's models do SaaS arithmetic without a calculator?},
+  year   = {2026},
+  doi    = {10.5281/zenodo.22128841},
+  url    = {https://doi.org/10.5281/zenodo.22128841}
+}
+```

RESULTS.md

@@ -0,0 +1,70 @@
+# Results — 2026-08-16 (v2, four models)
+
+Ten arithmetic problems, four Anthropic models, no tools, no MCP, single-shot
+answers. Expected values computed in code, never by a model.
+
+v1 (2026-08-12) ran only the two ends of the range, claude-opus-5 and
+claude-haiku-4-5. This run adds claude-sonnet-5 and claude-fable-5 and repeats
+all four in one sitting, so "same questions, same prompt, same isolation, only
+the model changed" still holds across the whole table. The v1 headline numbers
+reproduced: opus 10/10, haiku 3/10.
+
+**Note on claude-fable-5:** it is the model that ran this session. That does not
+affect the score, because the ground truth is computed in `run.ts` and the model
+never grades itself; it only submits an answer that the script checks.
+
+| | opus-5 | sonnet-5 | haiku-4-5 | fable-5 |
+|---|---|---|---|---|
+| answered | 10 / 10 | 10 / 10 | 10 / 10 | 10 / 10 |
+| exact | **10 / 10** | 9 / 10 | **3 / 10** | 9 / 10 |
+| within 1% | 10 / 10 | 10 / 10 | 10 / 10 | 10 / 10 |
+| worst error | 0.00% | 0.02% | **0.145%** | 0.00% |
+| reasoning tokens (total) | 7,771 | 34,520 | **58,027** | 8,274 |
+| reasoning tokens (median) | **936** | 3,281 | 7,051 | 1,148 |
+
+"Exact" means the answer differs from the computed value by no more than display
+rounding (relative error ≤ 1e-6).
+
+## Per task
+
+| task | expected | opus-5 | sonnet-5 | haiku-4-5 | fable-5 |
+|---|---|---|---|---|---|
+| `mrr-0` growth then churn, 36 mo | 38,115.31 | 38,115.31 | 38,115.31 | 38,115 (0.001%) | 38,115.34 |
+| `cash-1` revenue vs fixed burn, 60 mo | 4,864,506.92 | 4,864,506.92 | 4,864,506.96 | 4,864,544.82 (0.001%) | 4,864,506.92 |
+| `ltv-2` one division | 7,800.00 | 7,800 | 7,800 | 7,800 | 7,800 |
+| `two-phase-3` growth switches at m6, 60 mo | 18,578.75 | 18,578.75 | 18,578.73 | 18,553 (0.139%) | 18,578.75 |
+| `runway-4` stepped burn, month number | 5 | 5 | 5 | 5 | 5 |
+| `mrr-5` growth then churn, 12 mo | 14,887.45 | 14,887.45 | 14,887.45 | 14,887.54 (0.001%) | 14,887.46 |
+| `cash-6` revenue vs fixed burn, 60 mo | 5,822,959.30 | 5,822,959 | 5,822,959.31 | 5,831,410 (0.145%) | 5,822,959 |
+| `ltv-7` one division | 2,346.84 | 2,346.84 | 2,347.37 (0.022%) | 2,346.32 (0.022%) | 2,346.84 |
+| `two-phase-8` growth switches at m6, 60 mo | 21,262.90 | 21,262.90 | 21,262.91 | 21,271 (0.038%) | 21,263 |
+| `runway-9` stepped burn, month number | 11 | 11 | 11 | 11 | 11 |
+
+## What the numbers say
+
+**Nobody refuses and nobody collapses.** All four answered every question and
+every answer landed within 0.145% of the truth, which is already at odds with
+the shorthand that language models cannot do arithmetic.
+
+**"To the cent" is where they separate.** Exact matches: opus 10, sonnet 9,
+fable 9, haiku 3. The difference is drift, not failure. haiku was never wildly
+wrong, it was slightly wrong seven times out of ten. On `cash-6` that slight
+wrongness is **$8,451 off** a five-year cash forecast, and nothing in the answer
+signals it happened. The number simply looks right.
+
+**More thinking did not buy accuracy, it went the other way.** haiku spent the
+most reasoning of any model, 58,027 tokens, roughly 7.5x opus's 7,771, and
+finished last at 3/10. opus was both the most accurate and the cheapest to
+think. fable landed 9/10 on 8,274 tokens; sonnet landed the same 9/10 but spent
+34,520 getting there.
+
+**Difficulty concentrates where the rule changes.** The growth-switches-at-
+month-6 problems and the long cash forecasts produced haiku's largest errors.
+Straight compounding and single divisions were handled cleanly by all four.
+
+## Scope
+
+Ten problems, four models, one sitting, single-shot answers. This compares four
+models on identical tasks. It does not establish a failure rate for any of them,
+and says nothing about behaviour with tools enabled, which is how anyone would
+actually run these in production. It measures the model alone, on purpose.

results.json

@@ -0,0 +1,617 @@
+{
+  "date": "2026-08-16",
+  "seed": 20260812,
+  "models": [
+    "claude-opus-5",
+    "claude-sonnet-5",
+    "claude-haiku-4-5",
+    "claude-fable-5"
+  ],
+  "prompt": "Answer the question below. Do not show any working, do not explain. Reply with exactly one line in the form:\nANSWER: <number>",
+  "cliArgs": [
+    "-p",
+    "--output-format",
+    "json",
+    "--model",
+    "<model>",
+    "--tools",
+    "\"\"",
+    "--strict-mcp-config",
+    "--mcp-config",
+    "<private-project>\\research\\llm-arithmetic\\no-mcp.json"
+  ],
+  "tasks": [
+    {
+      "id": "mrr-0",
+      "kind": "mrr",
+      "months": 36,
+      "want": 38115.30678167854,
+      "question": "A SaaS starts at $8,000 MRR. Each month, in this exact order: MRR first grows by 7.0%, then 2.4% of the resulting MRR churns away. What is MRR after 36 months?"
+    },
+    {
+      "id": "cash-1",
+      "kind": "cash",
+      "months": 60,
+      "want": 4864506.921884278,
+      "question": "A company starts with $900,000 in cash. It spends a fixed $40,000 every month. Revenue is $18,000 in month 1 and grows 5.0% every month after that (month 2 revenue = month 1 revenue x 1.05, and so on). Cash after 60 months = starting cash + all revenue collected - all spend. What is it?"
+    },
+    {
+      "id": "ltv-2",
+      "kind": "ltv",
+      "months": null,
+      "want": 7800.000000000001,
+      "question": "ARPU is $240/month, gross margin is 78%, monthly customer churn is 2.4%. Using LTV = ARPU x gross margin / monthly churn, what is LTV?"
+    },
+    {
+      "id": "two-phase-3",
+      "kind": "two-phase",
+      "months": 60,
+      "want": 18578.74834305274,
+      "question": "A SaaS starts at $26,000 MRR. Each month, in this exact order: MRR grows, then churn is applied to the grown figure. Growth is 9.0% for months 1 through 6, and 3.0% from month 7 onwards. Churn is 4.0% every month. What is MRR after 60 months?"
+    },
+    {
+      "id": "runway-4",
+      "kind": "runway",
+      "months": null,
+      "want": 5,
+      "question": "A company has $300,000 in cash and no revenue. Monthly burn is $60,000 in months 1, 2 and 3, then increases by $10,000 every 3 months (so it steps up in month 4, again in month 7, and so on). In which month number does the cash balance first go below zero?"
+    },
+    {
+      "id": "mrr-5",
+      "kind": "mrr",
+      "months": 60,
+      "want": 14887.453040647115,
+      "question": "A SaaS starts at $12,000 MRR. Each month, in this exact order: MRR first grows by 4.0%, then 3.5% of the resulting MRR churns away. What is MRR after 60 months?"
+    },
+    {
+      "id": "cash-6",
+      "kind": "cash",
+      "months": 36,
+      "want": 5822959.304807226,
+      "question": "A company starts with $420,000 in cash. It spends a fixed $40,000 every month. Revenue is $18,000 in month 1 and grows 11.0% every month after that (month 2 revenue = month 1 revenue x 1.11, and so on). Cash after 36 months = starting cash + all revenue collected - all spend. What is it?"
+    },
+    {
+      "id": "ltv-7",
+      "kind": "ltv",
+      "months": null,
+      "want": 2346.842105263158,
+      "question": "ARPU is $49/month, gross margin is 91%, monthly customer churn is 1.9%. Using LTV = ARPU x gross margin / monthly churn, what is LTV?"
+    },
+    {
+      "id": "two-phase-8",
+      "kind": "two-phase",
+      "months": 36,
+      "want": 21262.901931922686,
+      "question": "A SaaS starts at $9,000 MRR. Each month, in this exact order: MRR grows, then churn is applied to the grown figure. Growth is 9.0% for months 1 through 6, and 3.0% from month 7 onwards. Churn is 1.5% every month. What is MRR after 36 months?"
+    },
+    {
+      "id": "runway-9",
+      "kind": "runway",
+      "months": null,
+      "want": 11,
+      "question": "A company has $750,000 in cash and no revenue. Monthly burn is $60,000 in months 1, 2 and 3, then increases by $10,000 every 3 months (so it steps up in month 4, again in month 7, and so on). In which month number does the cash balance first go below zero?"
+    }
+  ],
+  "rows": [
+    {
+      "id": "mrr-0",
+      "kind": "mrr",
+      "months": 36,
+      "model": "claude-opus-5",
+      "want": 38115.30678167854,
+      "answeredBy": "claude-opus-5",
+      "got": 38115.31,
+      "relError": 8.443645686841937e-8,
+      "thinkingTokens": 936,
+      "outputTokens": 949,
+      "raw": "ANSWER: 38115.31"
+    },
+    {
+      "id": "cash-1",
+      "kind": "cash",
+      "months": 60,
+      "model": "claude-opus-5",
+      "want": 4864506.921884278,
+      "answeredBy": "claude-opus-5",
+      "got": 4864506.92,
+      "relError": 3.873522928947228e-10,
+      "thinkingTokens": 1018,
+      "outputTokens": 1032,
+      "raw": "ANSWER: 4864506.92"
+    },
+    {
+      "id": "ltv-2",
+      "kind": "ltv",
+      "months": null,
+      "model": "claude-opus-5",
+      "want": 7800.000000000001,
+      "answeredBy": "claude-opus-5",
+      "got": 7800,
+      "relError": 1.1660188484268308e-16,
+      "thinkingTokens": 27,
+      "outputTokens": 38,
+      "raw": "ANSWER: 7800"
+    },
+    {
+      "id": "two-phase-3",
+      "kind": "two-phase",
+      "months": 60,
+      "model": "claude-opus-5",
+      "want": 18578.74834305274,
+      "answeredBy": "claude-opus-5",
+      "got": 18578.75,
+      "relError": 8.918508557902054e-8,
+      "thinkingTokens": 1665,
+      "outputTokens": 1678,
+      "raw": "ANSWER: 18578.75"
+    },
+    {
+      "id": "runway-4",
+      "kind": "runway",
+      "months": null,
+      "model": "claude-opus-5",
+      "want": 5,
+      "answeredBy": "claude-opus-5",
+      "got": 5,
+      "relError": 0,
+      "thinkingTokens": 90,
+      "outputTokens": 100,
+      "raw": "ANSWER: 5"
+    },
+    {
+      "id": "mrr-5",
+      "kind": "mrr",
+      "months": 60,
+      "model": "claude-opus-5",
+      "want": 14887.453040647115,
+      "answeredBy": "claude-opus-5",
+      "got": 14887.45,
+      "relError": 2.0424226399734698e-7,
+      "thinkingTokens": 338,
+      "outputTokens": 351,
+      "raw": "ANSWER: 14887.45"
+    },
+    {
+      "id": "cash-6",
+      "kind": "cash",
+      "months": 36,
+      "model": "claude-opus-5",
+      "want": 5822959.304807226,
+      "answeredBy": "claude-opus-5",
+      "got": 5822959,
+      "relError": 5.2345759298359134e-8,
+      "thinkingTokens": 1440,
+      "outputTokens": 1452,
+      "raw": "ANSWER: 5822959"
+    },
+    {
+      "id": "ltv-7",
+      "kind": "ltv",
+      "months": null,
+      "model": "claude-opus-5",
+      "want": 2346.842105263158,
+      "answeredBy": "claude-opus-5",
+      "got": 2346.84,
+      "relError": 8.970621215001095e-7,
+      "thinkingTokens": 45,
+      "outputTokens": 58,
+      "raw": "ANSWER: 2346.84"
+    },
+    {
+      "id": "two-phase-8",
+      "kind": "two-phase",
+      "months": 36,
+      "model": "claude-opus-5",
+      "want": 21262.901931922686,
+      "answeredBy": "claude-opus-5",
+      "got": 21262.9,
+      "relError": 9.085884377751978e-8,
+      "thinkingTokens": 2055,
+      "outputTokens": 2068,
+      "raw": "ANSWER: 21262.9"
+    },
+    {
+      "id": "runway-9",
+      "kind": "runway",
+      "months": null,
+      "model": "claude-opus-5",
+      "want": 11,
+      "answeredBy": "claude-opus-5",
+      "got": 11,
+      "relError": 0,
+      "thinkingTokens": 157,
+      "outputTokens": 167,
+      "raw": "ANSWER: 11"
+    },
+    {
+      "id": "mrr-0",
+      "kind": "mrr",
+      "months": 36,
+      "model": "claude-sonnet-5",
+      "want": 38115.30678167854,
+      "answeredBy": "claude-sonnet-5",
+      "got": 38115.31,
+      "relError": 8.443645686841937e-8,
+      "thinkingTokens": 5896,
+      "outputTokens": 5909,
+      "raw": "ANSWER: 38115.31"
+    },
+    {
+      "id": "cash-1",
+      "kind": "cash",
+      "months": 60,
+      "model": "claude-sonnet-5",
+      "want": 4864506.921884278,
+      "answeredBy": "claude-sonnet-5",
+      "got": 4864506.96,
+      "relError": 7.835474949328383e-9,
+      "thinkingTokens": 3281,
+      "outputTokens": 3295,
+      "raw": "ANSWER: 4864506.96"
+    },
+    {
+      "id": "ltv-2",
+      "kind": "ltv",
+      "months": null,
+      "model": "claude-sonnet-5",
+      "want": 7800.000000000001,
+      "answeredBy": "claude-sonnet-5",
+      "got": 7800,
+      "relError": 1.1660188484268308e-16,
+      "thinkingTokens": 0,
+      "outputTokens": 11,
+      "raw": "ANSWER: 7800"
+    },
+    {
+      "id": "two-phase-3",
+      "kind": "two-phase",
+      "months": 60,
+      "model": "claude-sonnet-5",
+      "want": 18578.74834305274,
+      "answeredBy": "claude-sonnet-5",
+      "got": 18578.73,
+      "relError": 9.873136984693287e-7,
+      "thinkingTokens": 10793,
+      "outputTokens": 10807,
+      "raw": "ANSWER: 18578.73"
+    },
+    {
+      "id": "runway-4",
+      "kind": "runway",
+      "months": null,
+      "model": "claude-sonnet-5",
+      "want": 5,
+      "answeredBy": "claude-sonnet-5",
+      "got": 5,
+      "relError": 0,
+      "thinkingTokens": 133,
+      "outputTokens": 144,
+      "raw": "ANSWER: 5"
+    },
+    {
+      "id": "mrr-5",
+      "kind": "mrr",
+      "months": 60,
+      "model": "claude-sonnet-5",
+      "want": 14887.453040647115,
+      "answeredBy": "claude-sonnet-5",
+      "got": 14887.45,
+      "relError": 2.0424226399734698e-7,
+      "thinkingTokens": 1403,
+      "outputTokens": 1416,
+      "raw": "ANSWER: 14887.45"
+    },
+    {
+      "id": "cash-6",
+      "kind": "cash",
+      "months": 36,
+      "model": "claude-sonnet-5",
+      "want": 5822959.304807226,
+      "answeredBy": "claude-sonnet-5",
+      "got": 5822959.31,
+      "relError": 8.917756667733204e-10,
+      "thinkingTokens": 5427,
+      "outputTokens": 5441,
+      "raw": "ANSWER: 5822959.31"
+    },
+    {
+      "id": "ltv-7",
+      "kind": "ltv",
+      "months": null,
+      "model": "claude-sonnet-5",
+      "want": 2346.842105263158,
+      "answeredBy": "claude-sonnet-5",
+      "got": 2347.37,
+      "relError": 0.0002249383269790866,
+      "thinkingTokens": 33,
+      "outputTokens": 46,
+      "raw": "ANSWER: 2347.37"
+    },
+    {
+      "id": "two-phase-8",
+      "kind": "two-phase",
+      "months": 36,
+      "model": "claude-sonnet-5",
+      "want": 21262.901931922686,
+      "answeredBy": "claude-sonnet-5",
+      "got": 21262.91,
+      "relError": 3.794438472858277e-7,
+      "thinkingTokens": 7349,
+      "outputTokens": 7362,
+      "raw": "ANSWER: 21262.91"
+    },
+    {
+      "id": "runway-9",
+      "kind": "runway",
+      "months": null,
+      "model": "claude-sonnet-5",
+      "want": 11,
+      "answeredBy": "claude-sonnet-5",
+      "got": 11,
+      "relError": 0,
+      "thinkingTokens": 205,
+      "outputTokens": 216,
+      "raw": "ANSWER: 11"
+    },
+    {
+      "id": "mrr-0",
+      "kind": "mrr",
+      "months": 36,
+      "model": "claude-haiku-4-5",
+      "want": 38115.30678167854,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 38115,
+      "relError": 0.000008048778940644374,
+      "thinkingTokens": 7710,
+      "outputTokens": 7723,
+      "raw": "ANSWER: $38,115"
+    },
+    {
+      "id": "cash-1",
+      "kind": "cash",
+      "months": 60,
+      "model": "claude-haiku-4-5",
+      "want": 4864506.921884278,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 4864544.82,
+      "relError": 0.000007790741452534025,
+      "thinkingTokens": 3873,
+      "outputTokens": 3887,
+      "raw": "ANSWER: 4864544.82"
+    },
+    {
+      "id": "ltv-2",
+      "kind": "ltv",
+      "months": null,
+      "model": "claude-haiku-4-5",
+      "want": 7800.000000000001,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 7800,
+      "relError": 1.1660188484268308e-16,
+      "thinkingTokens": 161,
+      "outputTokens": 173,
+      "raw": "ANSWER: 7800"
+    },
+    {
+      "id": "two-phase-3",
+      "kind": "two-phase",
+      "months": 60,
+      "model": "claude-haiku-4-5",
+      "want": 18578.74834305274,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 18553,
+      "relError": 0.0013859029993464173,
+      "thinkingTokens": 13085,
+      "outputTokens": 13097,
+      "raw": "ANSWER: 18553"
+    },
+    {
+      "id": "runway-4",
+      "kind": "runway",
+      "months": null,
+      "model": "claude-haiku-4-5",
+      "want": 5,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 5,
+      "relError": 0,
+      "thinkingTokens": 355,
+      "outputTokens": 366,
+      "raw": "ANSWER: 5"
+    },
+    {
+      "id": "mrr-5",
+      "kind": "mrr",
+      "months": 60,
+      "model": "claude-haiku-4-5",
+      "want": 14887.453040647115,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 14887.54,
+      "relError": 0.000005841116855149435,
+      "thinkingTokens": 10160,
+      "outputTokens": 10173,
+      "raw": "ANSWER: 14887.54"
+    },
+    {
+      "id": "cash-6",
+      "kind": "cash",
+      "months": 36,
+      "model": "claude-haiku-4-5",
+      "want": 5822959.304807226,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 5831410,
+      "relError": 0.0014512715529021876,
+      "thinkingTokens": 7051,
+      "outputTokens": 7064,
+      "raw": "ANSWER: 5831410"
+    },
+    {
+      "id": "ltv-7",
+      "kind": "ltv",
+      "months": null,
+      "model": "claude-haiku-4-5",
+      "want": 2346.842105263158,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 2346.32,
+      "relError": 0.00022247140614481596,
+      "thinkingTokens": 421,
+      "outputTokens": 434,
+      "raw": "ANSWER: 2346.32"
+    },
+    {
+      "id": "two-phase-8",
+      "kind": "two-phase",
+      "months": 36,
+      "model": "claude-haiku-4-5",
+      "want": 21262.901931922686,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 21271,
+      "relError": 0.00038085432097844377,
+      "thinkingTokens": 14239,
+      "outputTokens": 14251,
+      "raw": "ANSWER: 21271"
+    },
+    {
+      "id": "runway-9",
+      "kind": "runway",
+      "months": null,
+      "model": "claude-haiku-4-5",
+      "want": 11,
+      "answeredBy": "claude-haiku-4-5-20251001",
+      "got": 11,
+      "relError": 0,
+      "thinkingTokens": 972,
+      "outputTokens": 983,
+      "raw": "ANSWER: 11"
+    },
+    {
+      "id": "mrr-0",
+      "kind": "mrr",
+      "months": 36,
+      "model": "claude-fable-5",
+      "want": 38115.30678167854,
+      "answeredBy": "claude-fable-5",
+      "got": 38115.34,
+      "relError": 8.715218178934132e-7,
+      "thinkingTokens": 1430,
+      "outputTokens": 1443,
+      "raw": "ANSWER: 38115.34"
+    },
+    {
+      "id": "cash-1",
+      "kind": "cash",
+      "months": 60,
+      "model": "claude-fable-5",
+      "want": 4864506.921884278,
+      "answeredBy": "claude-fable-5",
+      "got": 4864506.92,
+      "relError": 3.873522928947228e-10,
+      "thinkingTokens": 1182,
+      "outputTokens": 1196,
+      "raw": "ANSWER: 4864506.92"
+    },
+    {
+      "id": "ltv-2",
+      "kind": "ltv",
+      "months": null,
+      "model": "claude-fable-5",
+      "want": 7800.000000000001,
+      "answeredBy": "claude-fable-5",
+      "got": 7800,
+      "relError": 1.1660188484268308e-16,
+      "thinkingTokens": 33,
+      "outputTokens": 44,
+      "raw": "ANSWER: 7800"
+    },
+    {
+      "id": "two-phase-3",
+      "kind": "two-phase",
+      "months": 60,
+      "model": "claude-fable-5",
+      "want": 18578.74834305274,
+      "answeredBy": "claude-fable-5",
+      "got": 18578.75,
+      "relError": 8.918508557902054e-8,
+      "thinkingTokens": 1148,
+      "outputTokens": 1161,
+      "raw": "ANSWER: 18578.75"
+    },
+    {
+      "id": "runway-4",
+      "kind": "runway",
+      "months": null,
+      "model": "claude-fable-5",
+      "want": 5,
+      "answeredBy": "claude-fable-5",
+      "got": 5,
+      "relError": 0,
+      "thinkingTokens": 104,
+      "outputTokens": 114,
+      "raw": "ANSWER: 5"
+    },
+    {
+      "id": "mrr-5",
+      "kind": "mrr",
+      "months": 60,
+      "model": "claude-fable-5",
+      "want": 14887.453040647115,
+      "answeredBy": "claude-fable-5",
+      "got": 14887.46,
+      "relError": 4.6746430468813303e-7,
+      "thinkingTokens": 488,
+      "outputTokens": 501,
+      "raw": "ANSWER: 14887.46"
+    },
+    {
+      "id": "cash-6",
+      "kind": "cash",
+      "months": 36,
+      "model": "claude-fable-5",
+      "want": 5822959.304807226,
+      "answeredBy": "claude-fable-5",
+      "got": 5822959,
+      "relError": 5.2345759298359134e-8,
+      "thinkingTokens": 1585,
+      "outputTokens": 1597,
+      "raw": "ANSWER: 5822959"
+    },
+    {
+      "id": "ltv-7",
+      "kind": "ltv",
+      "months": null,
+      "model": "claude-fable-5",
+      "want": 2346.842105263158,
+      "answeredBy": "claude-fable-5",
+      "got": 2346.84,
+      "relError": 8.970621215001095e-7,
+      "thinkingTokens": 42,
+      "outputTokens": 55,
+      "raw": "ANSWER: 2346.84"
+    },
+    {
+      "id": "two-phase-8",
+      "kind": "two-phase",
+      "months": 36,
+      "model": "claude-fable-5",
+      "want": 21262.901931922686,
+      "answeredBy": "claude-fable-5",
+      "got": 21263,
+      "relError": 0.000004612168067540336,
+      "thinkingTokens": 2161,
+      "outputTokens": 2172,
+      "raw": "ANSWER: 21263"
+    },
+    {
+      "id": "runway-9",
+      "kind": "runway",
+      "months": null,
+      "model": "claude-fable-5",
+      "want": 11,
+      "answeredBy": "claude-fable-5",
+      "got": 11,
+      "relError": 0,
+      "thinkingTokens": 101,
+      "outputTokens": 111,
+      "raw": "ANSWER: 11"
+    }
+  ]
+}

transcript.md

@@ -0,0 +1,431 @@
+# Transcript — 2026-08-16
+
+Models: claude-opus-5, claude-sonnet-5, claude-haiku-4-5, claude-fable-5. Seed 20260812. Tools and MCP disabled.
+
+Command (prompt over stdin):
+
+```
+claude -p --output-format json --model <model> --tools "" --strict-mcp-config --mcp-config no-mcp.json
+```
+
+Expected values are computed in `run.ts`, never by a model.
+
+## mrr-0
+
+**Question**
+
+```
+A SaaS starts at $8,000 MRR. Each month, in this exact order: MRR first grows by 7.0%, then 2.4% of the resulting MRR churns away. What is MRR after 36 months?
+```
+
+**Expected** (computed in code): 38115.30678167854
+
+### claude-opus-5
+
+```
+ANSWER: 38115.31
+```
+
+Parsed: 38115.31 · error: 0.0000% · reasoning tokens: 936
+
+### claude-sonnet-5
+
+```
+ANSWER: 38115.31
+```
+
+Parsed: 38115.31 · error: 0.0000% · reasoning tokens: 5896
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: $38,115
+```
+
+Parsed: 38115 · error: 0.0008% · reasoning tokens: 7710
+
+### claude-fable-5
+
+```
+ANSWER: 38115.34
+```
+
+Parsed: 38115.34 · error: 0.0001% · reasoning tokens: 1430
+
+## cash-1
+
+**Question**
+
+```
+A company starts with $900,000 in cash. It spends a fixed $40,000 every month. Revenue is $18,000 in month 1 and grows 5.0% every month after that (month 2 revenue = month 1 revenue x 1.05, and so on). Cash after 60 months = starting cash + all revenue collected - all spend. What is it?
+```
+
+**Expected** (computed in code): 4864506.921884278
+
+### claude-opus-5
+
+```
+ANSWER: 4864506.92
+```
+
+Parsed: 4864506.92 · error: 0.0000% · reasoning tokens: 1018
+
+### claude-sonnet-5
+
+```
+ANSWER: 4864506.96
+```
+
+Parsed: 4864506.96 · error: 0.0000% · reasoning tokens: 3281
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 4864544.82
+```
+
+Parsed: 4864544.82 · error: 0.0008% · reasoning tokens: 3873
+
+### claude-fable-5
+
+```
+ANSWER: 4864506.92
+```
+
+Parsed: 4864506.92 · error: 0.0000% · reasoning tokens: 1182
+
+## ltv-2
+
+**Question**
+
+```
+ARPU is $240/month, gross margin is 78%, monthly customer churn is 2.4%. Using LTV = ARPU x gross margin / monthly churn, what is LTV?
+```
+
+**Expected** (computed in code): 7800.000000000001
+
+### claude-opus-5
+
+```
+ANSWER: 7800
+```
+
+Parsed: 7800 · error: 0.0000% · reasoning tokens: 27
+
+### claude-sonnet-5
+
+```
+ANSWER: 7800
+```
+
+Parsed: 7800 · error: 0.0000% · reasoning tokens: 0
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 7800
+```
+
+Parsed: 7800 · error: 0.0000% · reasoning tokens: 161
+
+### claude-fable-5
+
+```
+ANSWER: 7800
+```
+
+Parsed: 7800 · error: 0.0000% · reasoning tokens: 33
+
+## two-phase-3
+
+**Question**
+
+```
+A SaaS starts at $26,000 MRR. Each month, in this exact order: MRR grows, then churn is applied to the grown figure. Growth is 9.0% for months 1 through 6, and 3.0% from month 7 onwards. Churn is 4.0% every month. What is MRR after 60 months?
+```
+
+**Expected** (computed in code): 18578.74834305274
+
+### claude-opus-5
+
+```
+ANSWER: 18578.75
+```
+
+Parsed: 18578.75 · error: 0.0000% · reasoning tokens: 1665
+
+### claude-sonnet-5
+
+```
+ANSWER: 18578.73
+```
+
+Parsed: 18578.73 · error: 0.0001% · reasoning tokens: 10793
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 18553
+```
+
+Parsed: 18553 · error: 0.1386% · reasoning tokens: 13085
+
+### claude-fable-5
+
+```
+ANSWER: 18578.75
+```
+
+Parsed: 18578.75 · error: 0.0000% · reasoning tokens: 1148
+
+## runway-4
+
+**Question**
+
+```
+A company has $300,000 in cash and no revenue. Monthly burn is $60,000 in months 1, 2 and 3, then increases by $10,000 every 3 months (so it steps up in month 4, again in month 7, and so on). In which month number does the cash balance first go below zero?
+```
+
+**Expected** (computed in code): 5
+
+### claude-opus-5
+
+```
+ANSWER: 5
+```
+
+Parsed: 5 · error: 0.0000% · reasoning tokens: 90
+
+### claude-sonnet-5
+
+```
+ANSWER: 5
+```
+
+Parsed: 5 · error: 0.0000% · reasoning tokens: 133
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 5
+```
+
+Parsed: 5 · error: 0.0000% · reasoning tokens: 355
+
+### claude-fable-5
+
+```
+ANSWER: 5
+```
+
+Parsed: 5 · error: 0.0000% · reasoning tokens: 104
+
+## mrr-5
+
+**Question**
+
+```
+A SaaS starts at $12,000 MRR. Each month, in this exact order: MRR first grows by 4.0%, then 3.5% of the resulting MRR churns away. What is MRR after 60 months?
+```
+
+**Expected** (computed in code): 14887.453040647115
+
+### claude-opus-5
+
+```
+ANSWER: 14887.45
+```
+
+Parsed: 14887.45 · error: 0.0000% · reasoning tokens: 338
+
+### claude-sonnet-5
+
+```
+ANSWER: 14887.45
+```
+
+Parsed: 14887.45 · error: 0.0000% · reasoning tokens: 1403
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 14887.54
+```
+
+Parsed: 14887.54 · error: 0.0006% · reasoning tokens: 10160
+
+### claude-fable-5
+
+```
+ANSWER: 14887.46
+```
+
+Parsed: 14887.46 · error: 0.0000% · reasoning tokens: 488
+
+## cash-6
+
+**Question**
+
+```
+A company starts with $420,000 in cash. It spends a fixed $40,000 every month. Revenue is $18,000 in month 1 and grows 11.0% every month after that (month 2 revenue = month 1 revenue x 1.11, and so on). Cash after 36 months = starting cash + all revenue collected - all spend. What is it?
+```
+
+**Expected** (computed in code): 5822959.304807226
+
+### claude-opus-5
+
+```
+ANSWER: 5822959
+```
+
+Parsed: 5822959 · error: 0.0000% · reasoning tokens: 1440
+
+### claude-sonnet-5
+
+```
+ANSWER: 5822959.31
+```
+
+Parsed: 5822959.31 · error: 0.0000% · reasoning tokens: 5427
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 5831410
+```
+
+Parsed: 5831410 · error: 0.1451% · reasoning tokens: 7051
+
+### claude-fable-5
+
+```
+ANSWER: 5822959
+```
+
+Parsed: 5822959 · error: 0.0000% · reasoning tokens: 1585
+
+## ltv-7
+
+**Question**
+
+```
+ARPU is $49/month, gross margin is 91%, monthly customer churn is 1.9%. Using LTV = ARPU x gross margin / monthly churn, what is LTV?
+```
+
+**Expected** (computed in code): 2346.842105263158
+
+### claude-opus-5
+
+```
+ANSWER: 2346.84
+```
+
+Parsed: 2346.84 · error: 0.0001% · reasoning tokens: 45
+
+### claude-sonnet-5
+
+```
+ANSWER: 2347.37
+```
+
+Parsed: 2347.37 · error: 0.0225% · reasoning tokens: 33
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 2346.32
+```
+
+Parsed: 2346.32 · error: 0.0222% · reasoning tokens: 421
+
+### claude-fable-5
+
+```
+ANSWER: 2346.84
+```
+
+Parsed: 2346.84 · error: 0.0001% · reasoning tokens: 42
+
+## two-phase-8
+
+**Question**
+
+```
+A SaaS starts at $9,000 MRR. Each month, in this exact order: MRR grows, then churn is applied to the grown figure. Growth is 9.0% for months 1 through 6, and 3.0% from month 7 onwards. Churn is 1.5% every month. What is MRR after 36 months?
+```
+
+**Expected** (computed in code): 21262.901931922686
+
+### claude-opus-5
+
+```
+ANSWER: 21262.9
+```
+
+Parsed: 21262.9 · error: 0.0000% · reasoning tokens: 2055
+
+### claude-sonnet-5
+
+```
+ANSWER: 21262.91
+```
+
+Parsed: 21262.91 · error: 0.0000% · reasoning tokens: 7349
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 21271
+```
+
+Parsed: 21271 · error: 0.0381% · reasoning tokens: 14239
+
+### claude-fable-5
+
+```
+ANSWER: 21263
+```
+
+Parsed: 21263 · error: 0.0005% · reasoning tokens: 2161
+
+## runway-9
+
+**Question**
+
+```
+A company has $750,000 in cash and no revenue. Monthly burn is $60,000 in months 1, 2 and 3, then increases by $10,000 every 3 months (so it steps up in month 4, again in month 7, and so on). In which month number does the cash balance first go below zero?
+```
+
+**Expected** (computed in code): 11
+
+### claude-opus-5
+
+```
+ANSWER: 11
+```
+
+Parsed: 11 · error: 0.0000% · reasoning tokens: 157
+
+### claude-sonnet-5
+
+```
+ANSWER: 11
+```
+
+Parsed: 11 · error: 0.0000% · reasoning tokens: 205
+
+### claude-haiku-4-5 (answered by claude-haiku-4-5-20251001)
+
+```
+ANSWER: 11
+```
+
+Parsed: 11 · error: 0.0000% · reasoning tokens: 972
+
+### claude-fable-5
+
+```
+ANSWER: 11
+```
+
+Parsed: 11 · error: 0.0000% · reasoning tokens: 101