A live product-path receipt
Three exact first-party harness assertions
Claude Fable synthesized the programs. Azdaja executed them locally against complete deterministic inputs. One synthetic live run per task returned 3/3 exact. This is not benchmark accuracy or independent replication.
This run tests a concrete Azdaja contract. The model receives a bounded root prompt and writes a generic inspection program rather than the answer. The complete source remains in the local evaluator. Azdaja runs the program, checks the runtime path, and returns one exact result or fails.
The three tasks
| Task | Exact result | Root prompt | Provider time |
|---|---|---|---|
| Duplicate-sensitive build-log aggregation | Answer: 13 | 12,739 B | 17.570 s |
| Repository blocker and source path | src/module_07777.rs|AZD-7777 | 12,460 B | 22.054 s |
| Color on the final matching catalog record | Color: cerulean | 12,770 B | 23.872 s |
Provenance map
| Evidence | Measurement source | Receipt or artifact SHA-256 | Scope |
|---|---|---|---|
| Live Fable suite | 2514d26 | 966b92bd3153c519bd3e4e9d62152a590ab848b77a2c84729d0c38cbaa86b508 | One provider-generated observation per synthetic task |
| Provider-free acceptance | fe91f98 | ebcc0632ef458a90b90eb0b95ed3343df85e66d9ed22633173e2ce3fdb1ec08b | Deterministic scripted transport and retained raw log |
| Shared release binary | Recorded by both receipts | 3111c880bcf2cbf738282bf826bd5649e175fa5a0efc0b3ace5b077425f0921a | Binary identity only, not a security attestation |
| Offline verifier bundle | manifest.json → source_commit | Every listed artifact is byte-counted and SHA-256-bound | Narrow first-party reproducible evidence |
What the receipt checks
- Each scenario recorded exactly one configured Claude Fable CLI call.
- Each model response is published and contains neither its expected answer nor its answer-specific constant.
- Each program ran once in the local Monty evaluator with zero recursive or semantic subcalls.
- Exact scanners found no 100-byte source span in any provider prompt.
- No provider prompt contained a repository, input, or scratch host path.
- The verifier binds the receipt to the source manifest and release binary hashes and rejects altered totals.
Verify the recorded evidence offline
The provider-generated live observation is already recorded. This command does not reproduce that model call. It verifies both retained receipts, the provider-free raw log, fixture specifications, deterministic invariants, verifier sources, and this public page against one SHA-256 manifest. It performs no provider call, downloads nothing, and fails closed on changed files, claims, answers, totals, paths, or source bindings.
./proof/reproduction/run.shInspect the manifestRead the verifierReproduction guide
This is narrow first-party reproducible evidence, not an independent replication or complete third-party verification capsule. Exact deterministic checks are separated from timestamped live observations, and the provider-free v2 input-digest limitation is explicit.
Inspect or replay it
The machine-readable receipt was generated on 2 September 2026 and binds source commit 2514d26 plus the release binary hash. It includes all three model-authored programs, command templates, prompt and response hashes, runtime counters, exact outputs, environment versions, source hashes, and explicit claim limits.
Read the resultInspect the receiptRun the harness
python3 bench/live_fable_suite/verify.py \
bench/results/live-fable-suite.json \
--binary target/release/azdajaClaim boundary
This is one synthetic live run per task on one model and one provider route. There is no baseline arm, no repeated trial, and no subscription token-usage record in text mode. It is product-path evidence, not a benchmark, leaderboard result, or superiority claim.