FactCircuit v0.3.2

FactCircuit v0.3.2

This release makes the first installation and complete offline example
reproducible. FactCircuit remains a research preview; the model-backed semantic
adapter is experimental.

• Added factcircuit quickstart --output DIRECTORY, which writes a readable
  summary, frozen inputs, resource config and full provenance trace.
• Added explicit budget configuration and existing-output protection.
• Added factcircuit --version and retained the newsverify compatibility CLI.
• Documented macOS/Linux and Windows installation, expected results, configuration,
  troubleshooting, and a minimal custom-claim local model input.
• Included the previously developed archived-response replay and local GGUF
  inference workflows. The provided GGUF setup selects Qwen, not Astra.
• Added isolated wheel installation checks outside the checkout, with first-run
  CI coverage on Linux, macOS and Windows.

Start with the installation guide.
Source archives include experiments and setup scripts. The wheel contains the
offline core CLI and adapter API. Downloaded weights and local configurations
are not distributed.

The offline example uses fictional evidence and hand-authored annotations.
The release checks do not generate fresh model responses or establish accuracy.
The earlier two-case Astra comparison and its known full-evidence diagnostic
failure remain documented in the repository. This release does not claim a fix
for that experimental semantic error or a new model-performance improvement.

FactCircuit v0.3.1 — Trace every claim. Close the evidence loop.

FactCircuit is live

Accuracy Tracing / NewsVerify Harness now has a shorter, clearer name: FactCircuit.

What changed

• The GitHub project and Python distribution are now FactCircuit.
• New integrations use the factcircuit import package and command.
• Existing newsverify imports and commands remain supported for compatibility.
• Current documentation, CI, security links, package metadata, and release tooling use the new name.
• Historical v0.2/v0.3 reports, raw traces, and NVScore identifiers remain unchanged as audit evidence.

Validation

• 254 regression tests passed.
• Python 3.11, 3.12, and 3.13 are covered by CI.
• The factcircuit-0.3.1 wheel built, installed, imported, and ran the offline demo successfully.
• The frozen Historical 2023 contract audit passed without model or live-news calls.

This is a branding and compatibility release. It does not change the previously published two-case model result or claim a new real-world accuracy measurement.

FactCircuit v0.3.0 — Auditable verification loops (pre-rename release)

Changelog

All notable changes to Accuracy Tracing are documented here. The project uses
semantic versioning for its public releases.

0.3.0 — 2026-09-06

Highlights

• Added the staged validation loop: seven single-responsibility semantic stages
  with deterministic Python validation and judgement assembly.
• Added bounded retrieval feedback with exact task attribution, immutable
  material versions, reanalysis of affected history, and explicit stop reasons.
• Added fail-closed target grounding for actors, actions, objects, quantities,
  time scopes, negation, comparisons, and cross-event conflicts.
• Added the frozen Historical 2023 comparison harness with cutoff enforcement,
  gold isolation, raw call traces, resource accounting, and offline scoring.
• Expanded evaluation, repair, release, provenance, and staged semantic
  regression coverage to 250 passing tests.

Observed pilot result

On the frozen two-case post-hoc pilot, the original monolithic adapter scored
1/2 and the staged adapter scored 2/2. The staged arm used 5.5× as many calls
and 3.17× as many total tokens. A separate full-evidence diagnostic failed in
the staged arm, so the run-level status is has_errors even though all scored
main cases completed. See the
full report.

This small pilot is an engineering signal, not a general accuracy estimate or
an equal-compute causal result.

Compatibility

• Python 3.11 or newer.
• The core package remains standard-library-only.
• Model-backed experiments use the optional model dependency group and run
  from a source checkout.
• Legacy demo, verify, and benchmark commands remain available.

0.2.0 — 2026-09-05

• Introduced the bounded provenance trace engine and fixed-target evaluation
  toolkit.
• Added immutable evidence versions, lineage-aware source grouping,
  reanalysis, explicit budgets, and deterministic offline fixtures.
• Added the MIT license, packaging metadata, contribution guide, and CI.