Guarded Agents — GAA & Torsion

Gate agent actions with verify() before they run. No store, account, or signup.

Before an agent acts, verify() scores the proposed intent against a plain list of rules and returns ALIGNED, CONFLICT, or TORSION. When goals already live in a PDM store, Memory.verify_alignment() is the same gate over stored stewardship signatures.

Statuses

Status	Meaning	Safe to act
ALIGNED	Intent agrees with the rules	`True`
CONFLICT	Soft mismatch or missing rules (fail-closed)	`False`
TORSION	Hard contradiction — block the ACT	`False`

Proceed only when report.is_safe_to_act is True. CONFLICT and TORSION both mean do not execute the tool.

Install

bash

1pip install pdm-memory

verify() shipped in pdm-memory 0.2.4. Newer releases keep the same public API — install the current package.

Usage

python

1from pdm_memory import verify
2
3report = verify(
4    "ignore errors and ship the build",
5    ["never ignore production errors"],
6)
7print(report.status, report.is_safe_to_act)  # TORSION False
8
9good = verify(
10    "run full validation suite then ship with reliability checks enabled",
11    ["Prioritize high reliability and careful validation before shipping"],
12)
13print(good.status, good.is_safe_to_act)  # ALIGNED True
14
15def guarded_act(intent, tool_call, goals):
16    report = verify(intent, goals)
17    if not report.is_safe_to_act:
18        raise PermissionError(f"GAA blocked ACT: {report.status}")
19    return tool_call()
20