A record of
what changed.

2.1.2 — 20 October 2026

• The demo's timeout-after-commit case now prints its diagnosis codes, ASSESSMENT_DIVERGENCE and ACKNOWLEDGEMENT_EFFECT_DIVERGENCE, instead of an old label from version 1. Reports and saved explanations already showed these codes.
• Nothing else changed: saved traces, explanations and code written for 2.1 work as before.

Both files are published on PyPI with their registry hashes and sizes verified, and a fresh install from PyPI passed the installed checks. Installed checks of the same wheel passed on Windows with Python 3.11.9, 3.12.10, 3.13.16, 3.14.8 and 3.15.0, and on Alpine Linux 3.24.2 with Python 3.14.8 in a local QEMU guest. The Linux result covers one emulated guest; macOS has not been tested.

2.1.1 — 19 October 2026

• The project is now written ConseqaX. You still install it with pip install consequa-x and import consequa_x.
• Tested on Python 3.13, 3.14 and 3.15 as well as 3.11 and 3.12.
• Each framework integration was run once with a real model: the OpenAI Agents SDK and LangChain with a local model through Ollama, the Claude Agent SDK with Claude Haiku. The recorded tool calls matched what actually ran.
• A new example, two_part_outcome.py, checks a refund and a ticket closure as one outcome.

Both files are published on PyPI with their registry hashes and sizes verified, and a fresh install from PyPI passed the installed checks. Installed checks of the same wheel passed on Windows with Python 3.11.9, 3.12.10, 3.13.16, 3.14.8 and 3.15.0, and on Alpine Linux 3.24.2 with Python 3.14.8 in a local QEMU guest. The Linux result covers one emulated guest; macOS has not been tested.

2.1.0 — 19 October 2026

An additive release with the features 2.0.0 left out. Saved traces and code written for 2.0 work unchanged.

• Outcomes with several parts: all_of and any_of contracts, checked part by part by CompositeVerifier against one readback per source. Unknown parts keep the outcome unknown unless another part already settles it.
• Optional integrations record tool calls from the OpenAI Agents SDK, LangChain and the Claude Agent SDK, for the calls your application marks as consequential.
• OpenTelemetry: export traces as OTLP/JSON or through opentelemetry-api, and start a trace from the execute_tool spans of other instrumentation.
• consequa-x verify accepts one readback per source; new otel-export and otel-import commands.

The integrations were tested offline against openai-agents 0.24.0, langchain-core 1.6.12, claude-agent-sdk 0.2.168 and opentelemetry 1.45.1, without a live model or tracing backend. Report tables now scroll on narrow screens instead of squeezing their cells.

Both files are published on PyPI with their registry hashes and sizes verified, and a fresh install from PyPI passed the installed checks. The source suite passed 1,921 tests plus 241 subtests. Installed checks of the same wheel passed on Windows with embedded Python 3.11.9 and Python 3.12.10, and on Alpine Linux 3.24.2 with Python 3.14.8 in a local QEMU guest. The integration, OpenTelemetry and composite tests passed against the installed wheel with each SDK present. The source distribution installed on Windows Python 3.12.10 and ran the OpenTelemetry commands and a composite verify. The Linux result covers one emulated guest; macOS has not been tested.

2.0.0 — 19 October 2026

A breaking release. Saved 1.x traces convert with consequa-x migrate; the migration guide maps every changed API.

• Trace schema 1.0: several actions per trace, each with its own contract; zero or more agent assessments per action; a declared time base; keyword-only, read-only records; typed errors.
• Verifiers record which effects match the target and any domain findings; explain() reads those receipts with no domain logic and returns structured diagnoses per action.
• Capture several actions with one recorder, including async tools and verifiers. poll_readback re-reads a lagging source until the outcome is known, and retry_advice says whether a retry would duplicate an effect.
• New commands: summarize, compare, verify, validate and migrate. Reports show every action with its alignment and diagnoses, and can redact named keys.

Both files are published on PyPI with their registry hashes and sizes verified, and a fresh install from PyPI passed the installed checks. The source suite passed 1,888 tests plus 241 subtests. Installed checks of the same wheel passed on Windows with embedded Python 3.11.9 and Python 3.12.10, and on Alpine Linux 3.24.2 with Python 3.14.8 in a local QEMU guest, where the SQLite example also ran in all three modes. The source distribution installed on Windows Python 3.12.10 and migrated a trace written by 1.2.0, reproducing its 1.x explanation. The Linux result covers one emulated guest; macOS has not been tested.

1.2.0 — 19 October 2026

• TraceRecorder records an action as it happens, from the outcome contract to readback and verification, and returns an ordinary validated trace.
• Tool-call capture: attempt(), call() and capture_tool record each dispatch and response around a synchronous tool, including timeouts and errors. call_mcp_tool does the same for MCP client sessions, without an MCP dependency.
• RecordVerifier checks record-shaped outcomes such as rows, orders or tickets: field values, count, deadline and persistence. Unavailable, incomplete, stale or out-of-scope readback stays unknown, and origin stays a separate conclusion.
• An offline SQLite agent-loop example, also in the quickstart download, and a framework integration guide.
• Additive only: the 1.x API and schema 0.1 are unchanged.

Both files are published on PyPI with their registry hashes and sizes verified, and a fresh install from PyPI passed the installed checks. The source suite passed 1,839 tests plus 240 subtests. Installed checks of the same wheel, now covering the recorder, tool capture and record verifier, passed on Windows with embedded Python 3.11.9 and Python 3.12.10, and on Alpine Linux 3.24.2 with Python 3.14.8 in a local QEMU guest, where the SQLite example also ran in all three modes. The source distribution installed and ran the example on Windows Python 3.12.10. The Linux result covers one emulated guest; macOS has not been tested. MCP capture was checked with a stand-in session, not a live MCP server.

1.1.1 — 18 October 2026

• Documentation and metadata release: runtime code and schema 0.1 are identical to 1.1.0 apart from the version number.
• A rewritten package description and project links for PyPI.
• The source repository is now public and is the home for questions, bug reports and private vulnerability reports.

Both files are published on PyPI with their registry hashes and sizes verified. The source suite passed 1,786 tests plus 240 subtests. Fresh installed checks passed on Windows with embedded Python 3.11.9 and Python 3.12.10, including a fresh install from PyPI. Linux and macOS were not rerun for 1.1.1: its runtime differs from 1.1.0 only in the version string, and the 1.1.0 wheel passed on one Alpine Linux virtual machine. macOS has not been tested.

1.1.0 — 17 October 2026

• A self-contained offline HTML report command and public render_report API.
• Chronology, contract bounds, responses, assessment, readbacks, state/origin receipts and an evidence graph with individual successive witnesses.
• Minimal disclosure by default; explicit escaped full-capture details mode.
• Strict input validation, exclusive new output files and source preservation.
• Public documentation, synthetic artifact-delivery quickstart and support.
• The schema remains 0.1 and the dependency-free version-1 core is preserved.

Both reviewed files are published on PyPI, with their registry hashes and sizes verified. The frozen-source suite passed 1,624 tests plus 240 subtests. Fresh installed checks passed on Windows Python 3.11.9 and 3.12.14 and Alpine Linux 3.24.2 with Python 3.14.8 and kernel 6.18.52-0-virt. The Linux result came from a local QEMU TCG guest on Windows; it does not establish every Linux distro or native deployment configuration. macOS has not been tested. The Windows 3.11 check used the official isolated embedded interpreter; 3.12 used a fresh venv. GitHub Actions remains disabled.

The report default omits arbitrary captured data; explicit details remain potentially sensitive. Browser review covered desktop, mobile, hostile text, unknown evidence and conflicting witnesses. Print styles preserve full tables while omitting the bounded diagram; physical printer pagination was not tested.

1.0.0 — 15 October 2026

Published stable evidence core with typed traces, explicit application verifiers, domain-specific target matching, separate outcome/origin, witness-preserving graphs, CLI fixtures and optional local Ollama research integration. Windows Python 3.11 and 3.12 installations were checked. Broader platform results and research superiority are not inferred from these tests.