Repository navigation
fix(enclave): stop writing the prompt text into the dbe results file - #9558
Merged
Merged
Conversation
The receipt carries every output file in full, so the prompt in each row of dbe_results.json put the benchmark owner's prompts into the receipt. The rows now keep prompt_uid, hazard, the completion, the two timings and the safety verdict. The benchmark owner's Step 7 looks each prompt up in their own prompts file, by prompt_uid. Both notebooks list the new row fields. A completion can still repeat its prompt: on the run of 7 Oct, 3 of the 5 completions held a run of 6 or more consecutive words of their prompt. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
receipt.json and receipt.dsse.json come from a run of this branch's job on v0.1.28 on 7 Oct 2026. The receipt's shape is unchanged: outputs, job.code and consent.manifestDigest differ from the previous example, along with the fields that change on every run. sigstore-log-entry.txt links the envelope's Rekor entry. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
koenvanderveen
approved these changes
Oct 7, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Removes the prompt text from the dbe results file, as agreed on Slack. Two commits: the change, and the receipt from a run of it.
What changed
dbe_results.jsondrops"prompt". It keepsprompt_uid, the hazard code, the completion,ttft,decode_tpsandsafety(violation,rationale).prompt_uid.promptfield, and no prompt text appears in the receipt outside the completions. The search used the prompts file and looked for each prompt whole or as any 6-word run.Prompt text can still reach the receipt through completions
In the 6 Oct run, 3 of 5 completions repeated their prompt word for word (longest shared run: ncr 11 of 14 prompt words, spc_hlt 11 of 15, spc_ele 6 of 79). In this run, again 3 of 5: ncr 13 of 14, spc_hlt 11 of 15, spc_fin 7 of 88.
Two ways to close that, if we want to:
outputs_sectionin syft-enclave, so it needs a release.How it was run
v0.1.28of OpenMined/syft-enclave-tinfoil, imagesha256:ea890c9b82dbf3c80c703d717dabe2c3db0c48b53bd269801984807732446864, deployed in OpenMined's Tinfoil org with--mark-latest false.5673708(devwith fix(enclave): verify email and data owners from the Confidential Space token #9548).runIdisdb5ae75709a54d989f20b636b4dd752c. Both owners' attestation passed, andverify_receiptaccepts the receipt.receipt.dsse.json)main.pysha2566ae5d31d892432536d8d7aa6d99708a8f09f5713d9656e7f22ee3423ed89747b.dev, onlyoutputs, the job code andconsent.manifestDigestdiffer, besides the per-run fields and the inference key file's hash (the key was rotated after feat(enclave): add a safety classifier and metric units to the dbe job #9555).🤖 Generated with Claude Code