Skip to content

Releases

TraceWarrant 0.8.1 - preserve collector failures

Choose a tag to compare

@G26karthik G26karthik released this 15 Oct 03:35
4a25775

TraceWarrant 0.8.1 - preserve collector failures

Fix two reproduced collector error paths. An exception that evaluates to False
could be propagated while recorded as OK/completed; exception boolean conversion
could also invalidate capture or replace the first clock fault. Presence checks
now use is not None, without invoking caller boolean conversion. When recording
and cleanup both fail after acquisition or during release, the original body
error takes priority, otherwise the first recording error; cleanup is its cause.
Both paths share synchronous release validation and close invalid returned
coroutines. Actual cleanup is called once; uncertain captures refuse export.

Public APIs, formats, scheduling and prior scientific conclusions are unchanged.
Python 3.11+, Apache-2.0; core runtime dependencies remain empty. This is bounded
maintenance of the completed experimental scope, not a new scientific study.

Qualified clean source: 4a2577524924bded8daa9d51eea240812e65c2b2. All five source CI jobs passed:
328 tests per Windows/Ubuntu Python 3.11/3.13 matrix, fresh installed core/
locked-SDK qualification, measured collection cost/memory and preservation plus
analytical validation. Eleven new tests include eight failures reproduced on
original source and two further cleanup/clock combinations before their fix;
the success-cleanup control passed before/after. The installed-wheel probe
explicitly records a false-valued exception as failure and verifies token release.
SHA256SUMS binds all twelve other release assets; hosted evidence is permanent.

Install the attached wheel and extract collection-example.zip:

python -m pip install ./tracewarrant-0.8.1-py3-none-any.whl
python collection/collect.py --output new-trace.json
tracewarrant inspect new-trace.json --output new-inspection.json
tracewarrant inspect collection/framework-trace.json --output framework-inspection.json

Closed SQLite resource coverage is 2/2 with one declared dependency. Pinned
Pydantic AI 2.51.0/OTel SDK 1.45.0 TestModel keeps separate companion traces and
partial coverage 2/4; graph/prediction suitability remain unsupported and
GPU demand UNKNOWN. TestModel does not demonstrate real task quality/adoption.
Callbacks/completeness are declarations; local release does not prove remote
cancellation. No native scheduler monkeypatch or global network/provider setup.

Exact-source five paired fresh-process SQLite runs (500 operations/10,000 rows)
measured 101.958 us added per operation,
24.38% loop overhead, and
34.48 MiB peak process memory. Installed SDK
peak: 115.48 MiB. Final serialization
is measured separately. Costs are material for short operations, not a general
overhead, agent capacity or speedup result. Local/hosted measurements meet the
provisional 2 GiB ceiling; this machine has about 31.2 GiB, not tested 16 GiB
hardware. Raw samples and hosted platform variation are included.

All 64 protected files, old tags and negative studies are preserved. Portable
performance/quality replay and rehashed false-report refusals remain qualified.
HotpotQA remains failed (4/8 exact, mean F1 0.541667); FRAMES established no added
decision value. No model rerun, regrading, paid provisioning, private answer
publication or PyPI upload occurred. Production readiness, adoption and model
superiority remain unestablished. No required maintenance gate remains after
public download/install verification is recorded.

Fix and acceptance,
interface,
public proof.

TraceWarrant 0.8.0 - usable observation collection

Choose a tag to compare

@G26karthik G26karthik released this 14 Oct 16:01
c8ae1af

TraceWarrant 0.8.0 - usable observation collection

TraceWarrant now supplies a bounded stdlib recorder for explicitly owned Python
operations and resources, plus explicit optional OpenTelemetry SDK export and a
pinned Pydantic AI companion example. Capture actual enqueue/acquire/release,
opaque pool identities, known outcomes and declared finish-to-start edges without
copying an application collector. Scheduling and cleanup remain caller-owned.
No application payloads or arbitrary labels are accepted by the recorder API.
Core runtime dependencies remain empty; Python 3.11+, Apache-2.0.

Qualified source: c8ae1af15e8d3e60972e72c5bcca27ad30775714. All five exact-source CI jobs passed:
Windows/Ubuntu Python 3.11/3.13 with 317 tests per matrix job, isolated installed
core and locked-SDK integration qualification, paired collection cost and memory
measurements, preserved workflows and research, plus analytical validation.
SHA256SUMS binds all twelve other release assets; hosted evidence is permanent.

Install the wheel and extract collection-example.zip:

python -m pip install ./tracewarrant-0.8.0-py3-none-any.whl
python collection/collect.py --output new-trace.json
tracewarrant inspect new-trace.json --output new-inspection.json
tracewarrant inspect collection/framework-trace.json --output framework-inspection.json

The closed SQLite example has one explicit dependency and owned-resource coverage
2/2. The Pydantic AI 2.51.0/OTel SDK 1.45.0 TestModel example intentionally retains
two companion traces and partial coverage 2/4: model operations are unowned,
dependency graph and prediction suitability unsupported, GPU demand UNKNOWN.
TestModel never establishes real task quality. No auto-complete native graphs,
scheduler monkeypatch, global provider or network exporter is installed.

Recording has a material cost for short operations. This machine's five paired
fresh-process runs (500 real SQLite operations on 10,000 generated rows each)
measured median 114.919 microseconds added
per operation
, 29.21% loop elapsed
overhead
, and 35.29 MiB peak process memory.
Final serialization is excluded from the loop and reported separately in the raw
asset. Installed SDK peak: 116.07 MiB.
All measured processes met the provisional 2 GiB ceiling; actual hardware was
about 31.2 GiB, so this is not a test on a 16 GiB machine, an agent capacity result,
a negligible-overhead claim or a speedup. Hosted measurements vary by platform.

Cancellation before acquisition fabricates no release; after acquisition it uses
the caller's real cleanup. Faults/abandoned captures refuse export, bounded buffers
fail explicitly, and cleanup does not swallow the original body exception.
Callbacks and completeness are caller assertions, not authenticated evidence.
Single event-loop/thread collection does not repair cross-host clocks; local
release does not establish remote cancellation or physical GPU capacity.

Portable quality/performance evidence retains its original replay/refusal behavior.
Performance remains performance-only; HotpotQA still fails (4/8 exact, mean F1
0.541667), with retrospective chronology. All 64 protected files, historical tags
and failed studies are preserved. No new model/service execution, regrading, paid
resource, private answer publication or PyPI upload occurred. This closes all four
bounded V10 gates as an experimental toolkit increment; production readiness,
external adoption and model superiority remain unestablished.

Interface and limits,
V10 scope,
verified closeout.

TraceWarrant 0.7.1 - final acceptance and integrity patch

Choose a tag to compare

@G26karthik G26karthik released this 14 Oct 14:22
ecc0ebe

TraceWarrant 0.7.1 - final acceptance and HTML integrity patch

Fix a demonstrated integrity gap in performance verify-bundle: a supplied HTML report with an updated checksum was accepted even when its presentation disagreed with replayed evidence. The verifier now compares HTML bytes against recomputation, as quality bundles already do. Four regression cases reproduced the gap before the fix. Existing formats, report generation, policies, JSON decisions and quality interfaces remain compatible. Use the same version for export/replay. Python 3.11+; no runtime dependencies.

Qualified source: ecc0ebe1c224a9345b355d38ef44a0a0c6f1ac61. All five exact-source CI jobs passed: Windows/Ubuntu Python 3.11/3.13, 292 tests per matrix job, fresh installed-byte/workflow qualification and preservation, plus analytical validation. Verify all ten release assets with SHA256SUMS.

After installing the attached wheel and extracting portable-evidence-examples.zip:

python -m pip install ./tracewarrant-0.7.1-py3-none-any.whl
tracewarrant verify-bundle frames-performance --output verified-performance
tracewarrant verify-quality-bundle hotpot-quality --output verified-quality

Expected exits: performance 0, quality 1. The performance result is explicitly performance-only: it does not assess answer quality or establish model superiority. The retained HotpotQA result still fails (4/8 exact; mean F1 0.541667); its roster/receipt is retrospective. Both directories contain native SHA256SUMS. Adding arbitrary files or editing reports makes verification invalid; even updating the checksums cannot make false HTML agree with recomputed evidence. Invalid bundles return 2 without output. Outputs must be new.

This patch completes the final acceptance stage of the agreed experimental open-source toolkit, including the original ten V3 criteria and the bounded V4-V8 workflows. No additional engineering stage is scheduled. Completion does not establish production readiness, external adoption, authentic measurements/chronology or a model win. It does not revive the rejected capacity-planner thesis.

All prior tags, frozen evidence and failed studies are preserved. No model run, regrading, private answer publication, paid/cloud service or PyPI upload occurred. Apache-2.0. Acceptance map, finite completion boundary, quality bundles.

TraceWarrant 0.7.0 - portable quality evidence

Choose a tag to compare

@G26karthik G26karthik released this 14 Oct 13:13
da3da5d

TraceWarrant 0.7.0 - portable application quality evidence

Export exact quality inputs with bundle-quality, then replay their checksums and JSON/HTML offline with verify-quality-bundle. Changed evidence, missing/unlisted entries, symlink files, incomplete binding and rehashed false reports are refused. Existing quality formats/gates and performance workflows remain compatible. Python 3.11+; no runtime dependency, model, simulator or network requirement.

Qualified source: da3da5d799a7c1ed351c40ad45ffc8cd1e381119. All five exact-source CI jobs passed: Windows/Ubuntu Python 3.11/3.13, 288 tests per matrix job, fresh installed-byte/workflow qualification and preservation, plus analytical validation. Verify all ten assets with the release SHA256SUMS.

After installing the attached wheel and extracting application-quality-example.zip:

python -m pip install ./tracewarrant-0.7.0-py3-none-any.whl
tracewarrant verify-quality-bundle hotpot-quality --output verified-quality

Expected exit 1: all eight declared tasks bind, but exact match 0.5 and mean answer F1 0.541667 still fail the original 0.75 floors. The ZIP contains a native bundle, with its own SHA256SUMS. No README belongs inside it. Export your own sample with bundle-quality INPUT --policy POLICY --output NEW_DIRECTORY; optionally supply both --plan PLAN --freeze RECEIPT. Output directories must be new and the export parent must exist. A failed policy is valid evidence; invalid/tampered bundles return 2 without a report.

The retained roster/receipt is retrospective, created AFTER failed V5 collection. Integrity does not establish authentic authorship, independent grading, preregistration chronology or initial roster completeness. A caller can replace all evidence consistently. Verify using the exporting version and retain independent published digests/collection evidence separately when needed. No model run or answer regrading is represented.

Experimental Apache-2.0 release. Production readiness, adoption and added model decision value remain unestablished. Prior tags and failed studies remain unchanged. No PyPI publication. Quality bundle guide, V5 results.

TraceWarrant 0.6.0 - frozen quality plans

Choose a tag to compare

@G26karthik G26karthik released this 14 Oct 08:53
086123b

TraceWarrant 0.6.0 — frozen quality plans

Bind an exact quality policy and declared task roster with freeze-quality, then optionally require --plan/--freeze when running qualify-application. Changed plan/policy bytes, omitted/substituted tasks or mismatched workload/sample identities are refused before quality reporting. quality-schema plan and quality-schema freeze export the new fixed formats. Python 3.11+; no runtime dependency, model, simulator, grading service or network requirement.

Original sample/policy formats and unbound JSON decision semantics remain compatible. Matching bound samples retain completion/multiple-quality gates and refusal of missing scores. Binding verifies exact bytes and declared membership, not authentic preregistration chronology, honest execution or completeness of the initial plan.

Qualified source: 086123bc5bf3b3f8b114716de2f337f0bfa6ef81. All five exact-source CI jobs passed: Windows/Ubuntu Python 3.11/3.13, 260 tests per matrix job, fresh installed-byte/workflow qualification and historical preservation, plus analytical validation. Verify all downloaded assets with SHA256SUMS.

After installing the attached wheel and extracting application-quality-example.zip:

python -m pip install ./tracewarrant-0.6.0-py3-none-any.whl
tracewarrant qualify-application hotpot-quality/sample.json --policy hotpot-quality/policy.json --plan hotpot-quality/plan.json --freeze hotpot-quality/freeze.json --output quality-report

Expected exit 1: binding succeeds for all eight declared tasks, but exact match 0.5 and mean answer F1 0.541667 still fail the original 0.75 floors. This roster/receipt was created AFTER the failed V5 study and is explicitly retrospective, not a new preregistration. No model runs, answer regrading or successful-agent claim is represented. Editing policy bytes or task membership yields exit 2 without a report. For a real study, publish the plan/policy/receipt before collection and retain independent chronology evidence. A local created-at value is not a trusted timestamp.

Experimental Apache-2.0 code release. Production readiness, external adoption and added model decision value remain unestablished. No prior release tag or failed scientific result changed. No PyPI publication, paid/cloud call or outreach. Quality-plan workflow, V5 results.

TraceWarrant 0.5.0 — application-quality qualification

Choose a tag to compare

@G26karthik G26karthik released this 14 Oct 08:30
0bbc550

TraceWarrant 0.5.0 — application quality before performance modeling

Check completion and several exact/graded quality criteria before performance predictions exist. qualify-application consumes content-free offered-task scores and an explicit policy, writes JSON/standalone HTML, and returns a CI exit code. quality-schema sample and quality-schema policy export the fixed formats. Python 3.11+; no runtime dependency, model, simulator, grading service or network requirement.

Missing/null required scores remain unsupported. Failed/censored cases stay in the offered denominator. Binary fraction and mean-score semantics are distinct; duplicate identities and malformed values are refused. Existing V3 artifact/V4 job formats, Python imports and CLI compatibility are preserved.

Qualified source: 0bbc550519114e82d5a70858ea093786f77fd89d. All five exact-source CI jobs passed: Windows/Ubuntu Python 3.11/3.13, 244 tests per matrix job, fresh installed-byte/workflow qualification and historical preservation, plus analytical validation. Verify all downloaded assets with SHA256SUMS.

After installing the attached wheel and extracting application-quality-example.zip:

python -m pip install ./tracewarrant-0.5.0-py3-none-any.whl
tracewarrant qualify-application hotpot-quality/sample.json --policy hotpot-quality/policy.json --output quality-report

Expected exit 1: all eight HotpotQA requests completed, but exact match 0.5 and mean answer F1 0.541667 failed the original 0.75 floors. The example is a retrospective scalar export of the retained failed V5 study; no model runs, answer regrading, new experiment or successful-agent claim is represented. A sample-policy pass is not semantic correctness, performance-model acceptance, authenticated collection or production certification. Hashes establish byte integrity, not honest collection or policy chronology.

Experimental Apache-2.0 code release. Production readiness, external adoption and added model decision value remain unestablished. No prior release tag or failed scientific result changed. No PyPI publication, paid/cloud call or outreach. Quality workflow, V5 results.

TraceWarrant 0.4.0 — offline validation workflow

Choose a tag to compare

@G26karthik G26karthik released this 14 Oct 04:53

TraceWarrant 0.4.0 — offline validation workflow

Run an explicit quality/performance policy from one manifest, get standalone HTML/JSON and a CI exit code, and export verified portable evidence. Python 3.11+; no third-party runtime dependencies, service, simulator or model requirement.

  • New validate-job, bundle-job, verify-bundle and job-schema commands.
  • Explicit gates for worst primary error, replication count, environment/envelope, material ranking, baselines and measured quality. Required missing evidence fails closed.
  • Contained bounded input snapshots, exclusive outputs, escaped HTML without scripts/remote assets, and full bundle checksum/report recomputation.
  • Existing artifact schema 1.0, freeze semantics, workload_lab imports and workload-lab compatibility preserved.

Qualified source: a248708e203869a6efd529595a20988c57f35115. All five hosted jobs passed on that SHA: Windows/Ubuntu Python 3.11/3.13, 190 tests per matrix job, installed-package workflow checks and offline preservation, plus Ubuntu analytical validation. Exact local source/wheel/archive and fresh installed checks also passed. Download SHA256SUMS and verify every asset.

The portable example ZIP contains frames-strict and frames-performance directories. After installing the attached wheel and extracting the ZIP:

python -m pip install ./tracewarrant-0.4.0-py3-none-any.whl
tracewarrant verify-bundle frames-performance --output performance-report
tracewarrant verify-bundle frames-strict --output strict-report

The numerical-only policy returns 0; the strict policy deliberately returns 1 and explains failed exact-answer quality (original 0/48), calibration extrapolation and unestablished added decision value. Open the generated report.html. Both are retrospective replays; no new experiment or successful-agent claim is made. Checksums establish integrity, not collector honesty or author authentication.

All historical evidence/tags remain unchanged. Experimental Apache-2.0 release; production suitability and external adoption remain unestablished. No PyPI upload, paid calls, cloud provisioning or outreach. Workflow guide, repository.

TraceWarrant 0.3.1 — experimental validation toolkit

Choose a tag to compare

@G26karthik G26karthik released this 30 Sep 18:24

TraceWarrant 0.3.1 — experimental validation toolkit

TraceWarrant checks whether telemetry supports a performance claim and whether frozen predictions survive real interventions. The Python 3.11+ core runs locally with no third-party runtime dependencies and no simulator requirement.

This prerelease includes observation conformance, simulator-neutral prediction/measurement/protocol/freeze artifacts, independent error and intervention evaluation, first-class baselines, and comparison/output-integrity fixes. Both tracewarrant and compatible workload-lab commands are included; workload_lab imports and frozen wire formats remain unchanged.

Release source: b0d5c5dc2a23589245ad1c6fabdf4d07dd2b51b6. All five hosted CI jobs passed: Windows/Ubuntu Python 3.11/3.13 (150 tests each, lint/format, build, fresh installed-package qualification and preservation audit) plus Ubuntu analytical validation. The attached wheel/source pair also passed fresh Windows Python 3.13 installed qualification from that clean source, reproducing the frozen evaluation without simulator dependencies. The checksum file binds the attached artifacts; source-provenance.json and hosted-ci-evidence.zip retain the exact source and checks.

The FRAMES study retains its negative findings: the model selected the same intervention as the utilization heuristic, and strict exact-answer quality passed 0/48 held-out sessions. This experimental toolkit does not establish production capacity recommendations, quality-qualified agents or external adoption. All frozen historical evidence remains unchanged.

Download the wheel and verify it against SHA256SUMS, then install with:

python -m pip install ./tracewarrant-0.3.1-py3-none-any.whl
tracewarrant --help

No PyPI publication is part of this release. Apache-2.0. Repository, private security reporting.