DeepSeek Responses API audit — public evidence download
Release: 2026-09-04, version 1.0
Original execution: 2026-08-19 UTC
Article: https://chat-deep.ai/research/deepseek-responses-api-compatibility-audit/
Publisher: Chat-Deep.ai, an independent publication, not affiliated with DeepSeek.

WHAT IS INCLUDED
- public-results.json: the byte-identical sanitized derivative prepared after the two original runs, with 48 model-case records, outcome categories, reported usage, source run IDs and evidence SHA-256 values.
- verify-public-results.mjs: a new, offline-only reader check for the JSON hash, unique result count, model totals, failure categories, token sums and dated price estimate. Requires Node.js; run: node verify-public-results.mjs . It does not send requests, need an API key, or write files.
- RETAINED-RECORDS-VERIFICATION.txt: the publisher's September 4 offline verification of retained source records, with limitations. This is not external certification.
- SHA256SUMS.txt: checksums of the included files other than the checksum list itself.

RESULTS
42 selected core records plus six separately executed optional Web Search records = 48.
34 PASS and 14 FAIL; each of Flash and Pro has 24 results, 17 PASS and seven FAIL.
FAIL categories: six missing final messages after Web Search execution, six incorrect response media types, one JSON Schema content-conformance failure, one exact-text trailing-period difference.
Usage is provider-reported: 281681 input tokens (235776 cached), 4009 output tokens (44 reasoning), 285690 total. Forty records report usage; eight do not. Missing usage is not zero.
The frozen peak-rate estimate is USD 0.246114, not an invoice, balance debit or measured total spend.

WHAT READERS CAN AND CANNOT VERIFY
This download allows inspection of all 48 sanitized records, full evidence identifiers, failure labels, reported usage and arithmetic. The included verifier checks the downloadable bytes and internal consistency.
It is NOT a complete raw request/response archive or a self-contained rerun kit. Raw provider responses, request bodies, account-balance observations, private manifests and credentials are not included. Their recorded hashes are identifiers, not a substitute for access to those private bytes. The publisher rechecked them locally; readers cannot independently repeat that private-source comparison from this download alone.
The superseded core Protocol 1.10.0 fixture, runner and protocol source bytes were not recovered during this September 4 review. Their hashes remain anchored in retained run records. The later Protocol 1.11.1 files must not be presented as the old core sources.
No API test was rerun for this evidence release, and no new model result was created. The September 4 date describes evidence access and offline verification, not a new benchmark date.

HISTORICAL METADATA
The original JSON contains claims_state=COMBINED_LOCAL_QA_PASS_PUBLICATION_NOT_AUTHORIZED. This records its historical preparation state on August 19. The publisher has now authorized this public evidence release. The old JSON bytes and hash are preserved rather than silently rewriting historical metadata.

INTERPRETATION
One sample per model-case does not estimate reliability, variance or a general success rate. PASS/FAIL refers to exact assertions, not overall compatibility, model quality, a score or a winner. Web Search cases assess response shape, not truth or source quality. Vision Exp was not tested; current provider behavior may differ from the August 19 snapshot.

SOURCE IDENTITIES
Core: RSP1100-CORE-20260819T190014Z; Protocol 1.10.0; 42 selected records.
Manifest: b297b7e1005c3905b03ee01840145d832568fbc51ee36fa76bdc1139a6ccc4dc
Canonical tree: 026b7257085220df7cf48cdf34c4aefaf8e70ad09e8b1b94d0b0f2ebc109a44a
Optional: RSP1111-OPTIONAL-20260819T212207Z; Protocol 1.11.1; six selected records.
Manifest: f27b622e7025454476b7f928d55d9a76305667c53d08fc9fe5b0ff7d114f003c
Canonical tree: 7df2bcd2a7fee2178299c921d76523a4c94e67e8e7a193fc36e13b2865367cb4
public-results.json: b84a59b5c73c8a62ba7a029e543b5fdf01bb53c24e54e3fe33ab0ec279bbd9d9

INTEGRITY LIMIT
Checksums establish identity relative to named files. They do not independently prove that a historical live execution occurred or that every original assertion was well designed. This release does not claim independent certification or full raw-run reproducibility.
