CROSSWIRE SOURCE / EVIDENCE

Behavioral interoperability / finite models

Evidence before
connection.

How can independently developed systems connect through an ambiguous interface? CROSSWIRE keeps the competing interpretations, probes an opaque target, and checks a contract across the surviving models.

Get the source Inspect the evidence

RELEASE / crosswire-formal-synthesis-2026-11-03

This is the source and evidence portal. Browse the research, download the release, and inspect scientific receipts here. Full simulations run in the local Python workbench.

Frozen release / local validation

Recorded 2 November 2026
Synthetic finite-state experiments

666Automated tests passedFrozen-source full suite
44Study rows checked & replayedIndependent public checks + exact replay
22Matched backend configurationsFlat and factored behavior agreed

These are local validation results. Backend repetitions represent the same underlying target conditions; they do not add independent targets. Read the verification record ↗

02 / INSIDE THE RELEASE

Plan. Preserve. Check.

01 / BUDGETED EXPERIMENTS

Look ahead before querying.

Depth-two experiment trees weigh hypothetical responses against a reset-and-action budget. The worker executes the checked root query, observes the response, then replans.

02 / EXPLICIT AMBIGUITY

Keep the candidate set visible.

Flat and factored backends represent the same finite adapter class. Structural factor counts describe representation; they do not measure memory compression or establish end-to-end scaling.

03 / INDEPENDENT EVIDENCE

Reconstruct the proposed tree.

A separate checker enumerates models, response branches, safety guards and costs. Source-bound replay checks recorded scientific decisions against an actual rerun.

03 / OBSERVED RESULTS

A useful result includes its limits.

Research & methods ↗

GUARDED GENERATED-TARGET PILOT

No coverage gain in the two-parent pilot.

Lookahead changed acquisition costs while useful coverage stayed equal across the three methods. This small exploratory study does not establish a general advantage.

Mean calibration calls, per backend, over two structural parents. Call caps include resets and actions.
Probe methodCap 4Cap 8
Entropy3.07.0
Branching3.07.0
Lookahead3.56.0
CAP 4 / USEFUL COVERAGE0 / 2Parents succeeded · all methods
CAP 8 / USEFUL COVERAGE1 / 2Parents succeeded · all methods

Two parents leave wide uncertainty. Label seeds and backend repetitions do not enlarge that denominator. No calibration or deployment hazards were recorded in this guarded pilot.

Inspect the saved pilot report ↗

SEMANTIC-CORRUPTION NEGATIVE CONTROL

Raw agreement has a blind spot.

Hidden meanings can change while raw observations stay compatible with the public model. The monitor may admit a plan even when the private semantic audit finds a hazard.

All four semantic-corruption control runs were admitted and hazardous.

WHAT THE GUARANTEES REQUIRE

  • The target belongs to the declared finite model class.
  • Behavior is stationary and resets are deterministic.
  • Semantic anchors are truthful.

Observable compatibility leaves model inclusion unknown. Candidate counts are not calibrated harm probabilities. Neither matching outputs nor a content hash establishes semantic grounding or real-world safety.

Inspect the control report ↗

04 / SOURCE & PROVENANCE

Take the release with you.

Browse Space files ↗
RESEARCH PAPER / LATEX

Formal Synthesis Paper

Mathematical theory, decision-theoretic VoI probing, anytime-valid Ville martingales, and certified runtime shielding.

crosswire_formal_synthesis.tex ↗ DOWNLOAD / SOURCE ARCHIVE

Source, tests and evidence

Inspect the implementation, reproduce the bounded studies, and build the local workbench.

crosswire_lab-0.1.0.tar.gz ↓ DOWNLOAD / PYTHON WHEEL

Install the packaged lab

The release wheel includes the Python runtime and local browser workbench assets.

crosswire_lab-0.1.0-py3-none-any.whl ↓ INSPECT / RELEASE MANIFEST

Check release identities

Find the source revision and artifact hashes. A digest identifies content; it does not authenticate experimental claims.

release-manifest.json ↗
GITHUB COMMITView tagged source ↗
PYTHON SOURCE HASH5e8274ec5a4577237407f94379c87b7f9b774befac85cd62f1b14f11dd9c0fe1

05 / RUN THE FULL WORKBENCH

The experiment runs on your CPU.

Extract the source archive, open a terminal in its folder, and activate a Python 3.11+ virtual environment. Install the lab and launch its loopback browser workbench.

The workbench provides configuration, live candidate inspection, study comparisons, receipt export and exact replay.

Installation & quickstart ↗
LOCAL PYTHON / 127.0.0.1
$ python -m pip install -e ".[dev,baselines]"
$ python -m crosswire serve --port 8781 --open