Skip to main content

Reversible cancellation and normalization with audit trails; fail-loud on contradiction

Project description

muqabalah (المقابلة) — Reversible Cancellation + Normalization

Part of the Mizan stack — the Arabic-first reliability scale for AI agents.

License: MIT Python 3.10+ Tests: 19 passing

The Balance operation as a standalone primitive. Companion to jabr. Reversible. Audited. Fail-loud on contradiction.


What it does

Takes a prompt with duplications, redundancies, or contradictions. Produces a canonical form with a full audit trail. The original is recoverable.

Critically: contradictions are never silently resolved. If two phrases conflict and one would have to be chosen over the other, balance() raises CancellationConflict and surfaces both options to the caller.

from muqabalah import balance, unbalance, CancellationContext

ctx = CancellationContext(
    user_normalizations={"u.s.": "United States", "usa": "United States"},
    contradiction_predicates=[("alice", "bob")],  # detect conflict
)

# Duplication + normalization
result = balance("I live in u.s. I live in u.s.", ctx)
print(result.output)
# I live in United States.
assert unbalance(result.output, result.trace) == "I live in u.s. I live in u.s."

# Contradiction — raises
try:
    balance("send to alice and send to bob", ctx)
except CancellationConflict as e:
    print(e.conflicts)  # → [{"predicate_a": "alice", "predicate_b": "bob", ...}]

Design properties

  1. Reversibility. unbalance(balance(p, ctx).output, trace) == p whenever balance succeeds. Verified by 19 tests.

  2. Fail-loud on conflict. When contradictions are detected, balance raises CancellationConflict with all conflicting spans. The library never picks a side silently.

  3. Trace integrity. Every removal records the exact removed text, span, and rationale. unbalance verifies both per-entry kept-text and a final input_hash round-trip. Wrong traces raise CancellationError.

  4. Determinism. Given the same (prompt, context, detectors), byte-identical output and trace.

Why fail-loud matters

Modern agent pipelines silently resolve contradictions all the time. A user says "send to A; also send to B" and the agent just picks one. The user never knows. The audit log shows only the chosen action.

muqabalah makes contradiction explicit. The agent must decide — and the decision is recorded in a separate step, not buried inside an opaque chain-of-thought.

Built-in detectors

Detector Behavior
ContradictionDetector Reports conflicts (no silent resolution); runs first
NormalizationDetector Replaces user-defined non-canonical forms with canonical ones
DuplicateDetector Removes literal duplicate sentences (keeps first occurrence)

Add a custom detector:

from muqabalah import Detector, DetectorResult, CancellationContext

class MyDetector:
    name = "my_detector"
    def find(self, input: str, context: CancellationContext) -> DetectorResult:
        # Return DetectorResult(actions=[...], conflicts=[...])
        ...

CLI

# Balance
muqabalah balance --prompt "..." --context ctx.json

# Round-trip verify
muqabalah roundtrip --prompt "Hello. Hello."

# Reverse balance to original
muqabalah unbalance --output "..." --trace-file trace.json

Where ctx.json is:

{
  "user_normalizations": {"u.s.": "United States"},
  "contradiction_predicates": [["alice", "bob"]]
}

Install

pip install -e .

Tests

pytest tests/ -v

19/19 pass on Python 3.10+. No runtime dependencies.

What it is not

  • Not a paraphraser. muqabalah only removes, replaces, or surfaces conflicts. It never rewrites for "clarity" or "tone."
  • Not a fact-checker. It detects self-contradictions declared by the caller in contradiction_predicates. It does not call out to a knowledge base.
  • Not a deduplicator for non-sentences. The default DuplicateDetector works on sentences. For other granularities, write a custom detector.

Composition with jabr

jabr and muqabalah are designed to compose:

from jabr import restore
from muqabalah import balance

# Pipeline: restore → balance
restored = restore(prompt, restoration_ctx)
balanced = balance(restored.output, cancellation_ctx)

# Both reversible:
from jabr import unrestore
from muqabalah import unbalance
recovered = unrestore(unbalance(balanced.output, balanced.trace), restored.trace)
assert recovered == prompt

Failure modes

See FAILURES.md.

License

MIT.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

muqabalah-0.1.0.tar.gz (16.7 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

muqabalah-0.1.0-py3-none-any.whl (13.4 kB view details)

Uploaded Python 3

File details

Details for the file muqabalah-0.1.0.tar.gz.

File metadata

  • Download URL: muqabalah-0.1.0.tar.gz
  • Upload date:
  • Size: 16.7 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.13

File hashes

Hashes for muqabalah-0.1.0.tar.gz
Algorithm Hash digest
SHA256 0f7682ec353cc0ef3a875e084cfcc1b7a01049102df8d9fbe6e7d64eca364500
MD5 79023a07d6bd2fbf8f067a405664c06c
BLAKE2b-256 ff539e70c7bbe7a9e2e26bc2a9d63c6bbaab478f120bbba58f7b14c544daca18

See more details on using hashes here.

Provenance

The following attestation bundles were made for muqabalah-0.1.0.tar.gz:

Publisher: release.yml on Moshe-ship/muqabalah

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file muqabalah-0.1.0-py3-none-any.whl.

File metadata

  • Download URL: muqabalah-0.1.0-py3-none-any.whl
  • Upload date:
  • Size: 13.4 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.13

File hashes

Hashes for muqabalah-0.1.0-py3-none-any.whl
Algorithm Hash digest
SHA256 6e7535b260a340ac0a10258735a7b1eaf7fdeb1b85d70e86a861c6333dc763a3
MD5 7e0b6576ae8271b93845d347c365f050
BLAKE2b-256 52dcdb896ffbf3d2ffdb27e32b1d1dc09658031f52576d3b655df4acaf90f8ef

See more details on using hashes here.

Provenance

The following attestation bundles were made for muqabalah-0.1.0-py3-none-any.whl:

Publisher: release.yml on Moshe-ship/muqabalah

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page