Skip to content

Protect HR tasks and personnel data at model boundaries - #21

Merged
gadzbi123 merged 8 commits into
mainfrom
hr-purpose-guard
Oct 4, 2026
Merged

gadzbi123 merged 8 commits into
mainfrom
hr-purpose-guard

Conversation

@gadzbi123

@gadzbi123 gadzbi123 commented Oct 4, 2026 •

Copy link
Copy Markdown
Contributor

HR requests could ask the model to judge employees under a harmless X-Purpose header, and confidential purpose text could be recorded even when a request was blocked. This adds HR purpose enforcement before both the task model and Jev, plus protected synthetic personnel-data tools.

  • Block configured review, ranking and termination purposes and common named-person judging requests, including conversation history, on both API adapters.
  • Classify employee files as DP30, apply employee scope, and refuse retrieval when results exceed the destination ceilings. Keep ordinary handbook questions available.
  • Withhold sensitive or unclassified demo previews. Hide HR purpose headers from new audit/session reports and all early identity denials.
  • Add the HR demo persona/scenario, policy validation and regression coverage. Assessment and demo instructions: docs/hr-review.md.

Validation: final integrated pytest suite 289 passed, 1 live-network test skipped; git diff --check passed. Implementation and independent security review used Claude Opus. Final focused independent Opus verification passed: 116 HR/reporting tests, no new blockers. Accepted limitations are tracked below.

Known limits: deterministic signatures do not cover every paraphrase. Confirmed conservative false positives for technical performance wording and negated statements are tracked in #20. Existing audit records are preserved; historical free-text purposes are not rewritten.

Board task: Enforce X-Purpose + HR role (no judging employees), https://github.com/orgs/TheRealStartup/projects/1.

gadzbi123 and others added 8 commits October 4, 2026 01:32
… run

query_datalake takes only a query name from the `datalake` catalog in policy.yaml
(demo/datalake.py holds the synthetic datasets). Unknown queries, queries the role may
not run, and queries whose result could not be sent on (model limit, or Jev's while the
AI check is on) are refused before they run, with one neutral reason. A result's class
comes from the query the gateway let through; a missing or different label withholds it.
Lower-class views exist only as declared transformations (sector counts for the deal
team, small groups dropped). tests/test_datalake.py; decisions.md D7.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@gadzbi123
gadzbi123 marked this pull request as ready for review October 4, 2026 00:14
@gadzbi123
gadzbi123 merged commit 7edad6b into main Oct 4, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant