Correction evidence
Mines correction-shaped turns and their surrounding exchange from supported work logs.
Hermeneutic 0.1.12 · Python 3.10+ · Apache-2.0 published-release licenseOpen source · inspectable boundary
Correction-aware drift gate
Catch recurring AI drift with prior corrections.
Hermeneutic recovers correction evidence from supported work logs as advisory context and checks a draft for fixed English drift patterns.
$ printf '%s\n' 'Draft ready for review.' | uvx --from hermeneutic==0.1.12 hermeneutic gateRelease
v0.1.12
Published on PyPI
Runtime
Python 3.10+
Apache-2.0 published-release license
First run
One pinned command
No project files required
Scope
Bounded tool
Not a system-level guarantee
§ 01
The problem
Why this exists
Make the evidence boundary visible.
In repeated AI work, useful corrections can remain in logs while a later draft repeats a known drift pattern. Hermeneutic makes supported correction-shaped records available as advisory context and applies fixed English checks to a saved draft.
§ 02
Run one bounded check
Send two supplied drafts through the gate and see which wording the released rules flag, and which they let through.
Prerequisite. Python 3.10+ and uvx. The gate runs locally: no API key, no model call, no network call, and no mined log corpus.
INPUTRisky draft
Done. All 12 checks passed.
INPUTNeutral draft
Draft ready for review.
STEP 01Risky draft
$ printf '%s\n' 'Done. All 12 checks passed.' | uvx --from hermeneutic==0.1.12 hermeneutic gate
RISK — highest severity: high, listing completion_with_number, completion_with_all_quantifier and number_then_completion with the matched text. Exit code 1.
STEP 02Neutral draft
$ printf '%s\n' 'Draft ready for review.' | uvx --from hermeneutic==0.1.12 hermeneutic gate
PASS — no risk patterns matched. Exit code 0.
Verdict lineRISK or PASS, plus the highest severity when a rule fires.
Rule id and matched textWhich released rule fired and the exact span of the draft it read.
why: lineThe rule's own description of the obligation it is raising.
Exit code1 for a high or medium hit, 0 for a pass, so the check can sit in a hook or a script.
What this result does not establish. A hit is a lexical match on the draft's own wording. It marks a claim that now carries an evidence obligation; it is not evidence that the claim is false, and a pass is not evidence that the claim is true. This subcommand is the regex-only gate: it reads no correction corpus and no prior corrections, so the run says nothing about the retrieval path.
Hermeneutic 0.1.12 is published and maintained by Hermes Labs. Canonical repository: https://github.com/hermes-labs-ai/hermeneutic. Package: https://pypi.org/project/hermeneutic/.
§ 03
What it does
Advisory retrieval
Surfaces similar prior correction records as context for a new task.
Deterministic gate
Checks a draft with fixed English rules and reports an explicit result.
§ 06
Related concepts
GLOSSARYHermeneutic DriftA shift in the task, document, or referent a system takes to be in view.GLOSSARYEpistemic EngineeringEngineering how AI systems handle evidence, uncertainty, and justification.GLOSSARYRecursive interpretive conditioningHow retained interpretations reshape the conditions for later interpretation and action.
The boundary
Useful evidence is not a blanket guarantee.
Hermeneutic is an advisory drift check. It does not establish that a response, claim, or system is correct, complete, or ready for use.
Compare it with the other open-source tools →Read about hermeneutic drift →