Fruth 0.1.9

Fruth 0.1.9 separates runtime responsibilities behind existing contracts while preserving accepted work, evidence bindings and durable response state.

• Focused owners for preparation, evidence, Closure, materialization, continuation and durable commit make the core easier to maintain and extend.
• Process startup and cleanup verify ownership; Whisper admission and Python bootstrap/dependency checks are hardened.
• PDF extraction and rendering run in bounded, owned child processes. PDF continuation preserves selected-input authority and existing repair identities.
• MLX-VLM Flux 3 support, installation guidance and the documented Codex access boundary are updated.
• The source package carries exact validation dependency records. The README explains how consistent vocabulary connects code, tests and documentation for developers and AI coding agents.

Release validation reconciles 6,551 successful regression checks and one intentional compatibility-shim skip across recorded source epochs. The existing 380-case offline gate, 38 profiles and 198 owner checks retain their exact source evidence; the unchanged-runtime comparison and matching checker were verified. The archive rebuilt identically. Historical live campaigns and the five reference workflows retain their actual dates and scope.

See the changelog, validation status and attached release-validation summary.

Fruth 0.1.8

Fruth 0.1.8

This release turns the completed bounded Hypothesis and semantic-core investigation round into stronger runtime contracts. The studies followed caller intent through intake, execution, saved-file reads, review, cancellation and recovery. Counterexamples became fixes and maintained regressions.

• Keep accepted requirements and caller authority intact through control and repair.
• Require current, exactly bound content reviews for named saved-file workflows, and reject foreign or stale evidence references.
• Continue independently owed review work through bounded recovery while respecting cancelled branches and current-frame authority.
• Preserve the read/target contract during HTML syntax repair and deliver every accepted individual image report.
• Restore request-correlated PCC Pro-to-Cloud fallback and recognize the updated Apple Foundation Models CLI availability output.

The corrected runtime passed its matching 380-case offline gate and 198 owner checks. Local, AFM and PCC live campaigns, targeted recoveries and refreshed reference examples retain their actual source identities and limits. These bounded results strengthen specific guarantees; they do not establish general semantic correctness or unconditional progress.

See the changelog and validation status.

Fruth 0.1.7

Fruth 0.1.7

Fruth 0.1.7 strengthens how a response continues, preserves its accepted results,
and verifies later work. It also expands local image generation and embeddings
across supported backends.

• Durable continuation: preserve accepted sibling results through cancellation
  and retry, prevent stale worker publication from undoing acknowledged controls,
  and retain exact repair identities and evidence through completion.
• Local images: discover compatible MLX-VLM image checkpoints through the
  installed loader, expose supported diffusion-step controls, validate saved
  images, and release idle image memory after requests. Image inspections can
  use a compatible alternate vision instance after a transient failure.
• Embeddings: support backend-neutral text embeddings and qualified image or
  audio inputs, with explicit capability checks and malformed-vector rejection.
• Storage and workflows: configure independent archive and Self-Attack
  locations; preserve learning and external stores during full local archival.
• Hardening: retain deterministic regressions from bounded Hypothesis studies,
  including control/worker overlap and selected persistence/recovery faults.
  The portable test runner isolates ordinary checks from opt-in research campaigns.

Local models, AFM and PCC Cloud each have complete 170-case live conformance
coverage on their recorded source snapshots. The current report also documents
matching 380-case offline gates and five refreshed reference workflows, including
saved-file dependencies, audio, images and a portable website bundle. PCC coverage
combines two bounded windows and a successful audio retry.

Full changelog ·
Validation evidence and scope ·
Reference workflows

Fruth remains experimental. Conformance validates the exercised runtime
contracts; it does not guarantee model answer quality or prove general recovery
and progress properties. See the packaged release scope and known limitations.

Fruth 0.1.6

Fruth 0.1.6

This patch release brings together the authority, evidence and successor-lineage
hardening completed since 0.1.5, plus current examples and validation records.

• Local API effects require trusted Host/Origin admission and an explicit action
  declaration. Invalid preference encodings no longer reset saved preferences.
• Caller intent remains separate from planner-prepared payload. Explicit file
  targets survive extraction, and accepted bindings are checked before writes.
• Fulfillment and canonical reuse require matching saved-file identity and
  digest evidence. Conflicting and duplicate evidence stays visible through
  recovery; successive image branches must qualify using their own results.
• Bounded Hypothesis property and stateful campaigns exposed reproducible
  evidence/lineage defects. The resulting corrections have deterministic
  regressions and positive controls; the campaigns are not an exhaustive proof.
• Response text correction and monitor phase attribution now preserve the
  relevant response/branch identities. Uncorrelated timing stays unknown.
• Cleanup protects legacy outputs and retained learning/history evidence.
  Axios 1.20.0 is bundled locally with its license and provenance.
• PCC conformance tooling supports bounded windows, quota stopping and compatible
  evidence recovery. Live quota recovery remains a separate validation task.

The package includes the five refreshed October 10 reference workflows with
saved artifacts and checksums. The complete local and AFM live matrices remain
tied to their recorded source revisions; later corrections have their own
offline evidence. See the packaged
validation chronology
and changelog for details.

The experimental 0.x support scope is unchanged. Historical PCC evidence remains
separate. Native shortcut importability and new live provider runs were not
repeated during packaging.

Release preparation covered all 181 maintained test modules and two Research
modules: 5,338 distinct test cases passed, with one expected skip for a
non-distributed development shim. Five stale test fixtures were refreshed.
Independent archive builds are byte-identical; packaged reference artifacts,
manifest and public-source boundaries were verified.

See the full changelog.
The attached .sha256 file verifies the Fruth source archive.

Fruth 0.1.5

Fruth 0.1.5

Maintenance release improving branch progress and dependency repair validation.

• Independent ready work continues while another branch waits for availability;
  due waiters are rechecked in deadline order within existing resource gates.
• Additive graph repairs reject unresolved or ambiguous dependencies and cycles,
  with validation repeated at application time.
• New regressions cover these cases; scheduling and architecture documentation
  describe the resulting behavior.

Existing dependency, cancellation, evidence and retry-budget boundaries remain.
The experimental 0.x support scope is unchanged.

Validation: 4,581 distinct offline tests passed, with one expected
skip for a development-only compatibility shim excluded from the source package.
All 161 test modules and Research validation fixtures were covered.
Syntax, dependency, local documentation links, archive contents and reproducible
build checks passed. No new live-provider conformance claim is made.

Source archive: fruth-0.1.5.tar.gz

SHA-256: cb640fd2b20b6d4556796c31f155163a4927217706da4278664d0a005ddf276d

Fruth 0.1.4

Fruth 0.1.4 includes the complete source package with both signed Apple PCC shortcuts and assisted first-use setup.

• SQLite is the sole normal response Index, with transactional per-response updates, authenticated completeness and Ledger coverage. The append-only Ledger, frozen lineage, snapshots and evidence remain preserved.
• Explicit migration, reconstruction and current-history rollback/export tools support existing histories. Runtime does not maintain a JSON mirror or silently switch back to JSON.
• Archive, cleanup and reviewed-learning retention recognize SQLite Index generations. Relocated-history evidence paths and persistence guidance are corrected.

Download fruth-0.1.4.tar.gz and its checksum. Both Fruth PCC.shortcut and Fruth PCC Pro.shortcut are in fruth_integrations/shortcuts/ inside the tarball. Backend runtimes and model weights remain separate prerequisites.

For existing JSON histories, preserve all saved state, finish active work, stop all writers and verify a complete backup before following the explicit migration commands in the README. Already migrated histories need no repeat migration.

The experimental macOS/Apple Silicon support scope remains unchanged. Historical conformance reports retain their original verdicts; this release does not claim a new live conformance result.

Validation: 4,540 distinct offline tests passed, with one expected development-only shim skip. Synthetic native PDF tests, source syntax, local documentation links, exact manifests and deterministic builds passed. No live inference or new conformance campaign was run.

Fruth 0.1.3

Fruth 0.1.3

Fruth 0.1.3 adds local Apple Foundation Models and optional Apple Private Cloud
Compute, native macOS PDF processing, improved follow-ups and provider recovery,
and reusable Research evidence review. The experimental 0.x support scope
remains unchanged.

• Apple AI: on-device text and image analysis through the installed fm
  transport, with instance controls, streaming and request-bound context.
• Apple PCC: optional cloud text/image execution through user-installed
  Shortcuts, with recorded tier fallback and provider-limit evidence.
• PDF handling: native PDFKit/CoreGraphics page rendering, retained pypdf
  text extraction, fresh processing for incoming PDFs and ordered page dispatch
  through the selected OCR/vision path. Rendering and transcription accuracy
  remain separate guarantees.
• Follow-ups and fallback: reuse saved image/audio evidence, carry full
  selected replies and multiple artifact references, and recover eligible
  provider failures through bounded compatible alternatives while respecting
  explicit targets, refusals and evidence gates. Public artifact lists preserve
  registered dependency identities, consistently with downloadable bundles.
• Research: maintained offline candidate audit, exact response inspection,
  explicit Gold curation and retained-evidence validation. New installations
  initialize empty stores; private cases, reviews and retained evidence are not
  shipped.
• Documentation and packaging: current backend diagram, PCC setup guide,
  dependency/setup instructions and reproducible source archives with checksums.

Installation and upgrading

The supported environment is macOS on Apple Silicon with Python 3.11 or newer.
Extract the source archive, enter fruth-0.1.3, and run ./fruth start. A fresh
installation creates its Python environment and installs requirements; the
initial installation needs internet access. Backend runtimes and model weights
are separate. Apple AI requires available Apple Intelligence and fm; PCC
requires the Shortcuts described in the included setup guide.

For an existing environment, update requirements explicitly before the next
start/restart:

.venv/bin/python -m pip install -r requirements.txt

Preserve existing settings, state/, artifacts/ and Research stores. An empty
registry or store in a fresh release is not a migration over saved user data.

Conformance evidence

The dated September 25–27 campaigns remain unchanged. AFM recorded 168 passing
cases and two correct audio/transcript mismatch rejections, leaving successful
handoff coverage incomplete for those two cases. PCC recorded 126 passing cases;
remaining coverage includes provider-quota blocks, one sequence timeout and one
correct mismatch rejection. Neither campaign recorded invariant findings; both
retain their native INCOMPLETE verdict.

These campaigns predate the later PDF, provider-failover and Research changes.
They are historical integration evidence, not a full-conformance claim for every
0.1.3 path. Apple model output quality, guardrails, availability and usage limits
remain provider-dependent. The observed four-page AFM transcription had reading
errors despite complete page coverage.

Fruth's code and project documentation remain Apache-2.0. The separately licensed
bundled font retains its attribution and OFL license. Python dependencies,
Apple frameworks, backend runtimes and model weights are not bundled.

Release validation

The isolated offline checks passed 4,426 distinct tests, with one expected skip
for a development-only compatibility entrypoint excluded from source releases.
This includes native PDF rendering, Apple transports/fallback, response and
artifact contracts, and Research verification. Initial failures were resolved
through fixes and explicit reruns. No new live conformance campaign was run.

The attached source archive is the curated 467-file distribution; its SHA-256
checksum is attached separately. Backend runtimes, model weights and private
runtime/Research data are excluded.

PCC shortcut download added September 28

Download fruth-pcc-shortcuts.zip and its .sha256 checksum from the assets below.
The ZIP contains importable Fruth PCC (Cloud) and Fruth PCC Pro (Cloud Pro)
shortcuts, import instructions and file checksums. Import them explicitly in
Apple's Shortcuts app, keeping the exact names and avoiding duplicates. Apple's
model availability, access requirements and usage limits still apply.

Both exports were inspected and accepted by Shortcuts' import preview; the
focused packaging checks passed 46 tests. The original 0.1.3 tag, source archive
and source checksum are unchanged. The shortcuts and updated setup guide are
also available in the repository's main branch.

Fruth 0.1.2

Continues Ollmo 0.1.1 under the Fruth name, with legacy-input cleanup and
migration fixes. The existing engine and experimental 0.x support scope are
carried forward.

• Rename the project, Python packages, CLI and project configuration to Fruth;
  describe the former Ghost responsibility as interpretive inference.
• Use the canonical semantic_role_ids list for explicit role selection; remove
  the former public ghost_mode aliases. Preserve internal intent orientations,
  role sets, precedence and advisory authority.
• Set the web/API default to port 5011, preserving backend ports 11434–11600 and
  the existing startup, shutdown, restart, discovery and recovery workflow.
• Package maintained Research code and initialize missing empty stores during
  normal startup. Keep accumulated Research data and accepted learning separate.
• Normalize working-frame role summaries through the canonical request owner,
  including malformed inputs, deduplication and metadata precedence.
• Require contract-specific evidence before skipping already-completed-looking
  branches, including reads of saved files used by later outputs.
• Add one bounded retry for failed vision-analysis evidence when a compatible
  alternative is available, preserving the original branch and source binding.
• Reconcile accepted media and file outputs before final link repair and Closure;
  prefer exact verified producer paths to avoid unintended extra artifacts.
• Update project documentation, citation concepts, copyright attribution and the
  architecture diagram's persistence, observation and rebase authority boundaries.
• Replace migrated examples with five recorded Fruth reference runs, retaining
  their response, monitor, artifact and checksum evidence.
• Record the 18–19 September full conformance pass: 380 deterministic cases and
  170 live cases. See the dated report
  for tested scope and limitations.

The attached source archive includes a complete SHA-256 manifest. Its sibling .sha256 file verifies the archive itself.

Ollmo 0.1.1 — Stabilization and evidence hardening

Ollmo 0.1.1 — stabilization and evidence hardening

Release date: September 12, 2026. Tag: v0.1.1. These notes compare the
released source with the August 1, 2026 v0.1.0 public tag. They describe the
accumulated changes, not a new product architecture or a promise of production
stability. The release remains a solo-maintained experimental 0.x project.

What changed

Requested work stays attached to its evidence

Requested file identities and counts survive reduced detector results and graph
rebuilds. Category-scoped deferral keeps “do not create SVG” from suppressing
unrelated requested text files. Deferred work stays visible; it does not become
silently waived or fulfilled. Saved-file consumers receive the actual authorized
saved bytes with producer, response, branch, phase, path and digest bindings.
Changed or missing evidence prevents unsupported completion.
Explicit image reservations also survive comma/sentence boundaries and scoped
pronoun/ordinal references while remaining non-executable until promoted.

Artifact registration and public outputs follow the accepted canonical frame.
Equal bytes in different files do not erase distinct obligations. HTML/CSS/image
composition, counted image prompts, linked-artifact rebind, exact repair targets
and exported bundles have stronger checks against the actual saved outputs.

Audio, images and continuation are more reliable

TTS generation has stronger source, duration/completeness and physical integrity
checks, with bounded evidence-driven recovery for supported cases. Direct
TTS-to-STT work binds its transcript to its own generated source. Image preparation
and retry paths preserve exact per-branch prompts and current provider evidence.
Semantic review, deferred branch continuation, retained context and public output
projection have additional regression coverage. Explicit MLX reasoning controls
and audio transport handling are also improved.

These changes strengthen supported workflows; they do not guarantee model quality,
add automatic authority to model prose, or remove truthful BLOCKED outcomes.

Durable state and diagnosis have clearer boundaries

Frame storage, Index publication/verification, Epoch binding and Readiness
retention have been hardened. Private snapshot serialization/size preparation and
single-use observation reuse avoid repeated transformations while retaining fresh
source, identity, integrity and authority checks. No benchmark percentage is a
release guarantee.

Causal/transition telemetry and bounded state-flow diagnostics make convergence
and persistence costs easier to inspect. Late Fill telemetry has a distinct
observer owner. Readiness epoch-mismatch diagnostics report bounded physical
file changes without attributing an unproven writer or rereading payloads solely
for diagnostics. Observers remain separate from runtime execution authority.

Reproducibility and public guidance

The package now includes Self-Attack's deterministic/fake harness, explicit live
coverage scopes, resumable evidence capture, two compact corpora and five curated
reference packages. The documentation audit reconciled response truth, repeated
Closure, artifact identity, lookup/hydration and finalization ordering. Accepted
learning remains available as explicitly enabled advisory hints; it cannot replace
current evidence or independently authorize work.

Upgrading from 0.1.0

The supported platform remains macOS on Apple Silicon with Python 3.11 or newer.
Model weights and backend packages are installed separately. The canonical
Responses endpoints remain /api/responses and /v1/responses; their limited
compatibility surface is documented in Responses Contract.
This release does not establish a stable 1.0 API or a blanket downgrade guarantee.

1. Retain the old installation and back up its configuration, populated model
   registry, saved artifacts and durable state together before changing code.
   Preserve the Ledger and referenced sidecars as a set.
2. Extract the new archive into a separate directory and install its requirements
   using the installation instructions.
   Its generated empty model_ports.json is for a fresh installation; do not copy
   it over an existing populated registry.
3. Stop the old control plane explicitly before switching the running installation.
   Do not mix old/new processes or replace active state during execution. Preserve
   existing state rather than running clean/reset commands as an upgrade step.
4. Check existing response/artifact lookup after switching. Stricter evidence
   gates can expose previously hidden missing artifacts, invalid references or
   unresolved obligations as blocked/repair-needed states. Report those states;
   do not delete evidence or relax gates to make them appear complete.

Legacy/derived Index recovery remains supported by its current validation paths;
this release preparation does not run an automatic production-state migration.
Explicit maintenance/attestation is separate operator work. For rollback, retain
the matching old code and pre-upgrade data backup; newer writes are not assumed
to be understood by older code. See Truth Sources.

Validation and reference evidence

Release validation on September 12, 2026

The final full public-source suite passed: 3,394 tests, with one expected
skip and 942 passing subtests, in 11 minutes 12 seconds. It ran in a frozen,
disposable copy with network access blocked and writes confined to temporary
test storage. The skipped check concerns a development-only monitor shim that
public packages intentionally exclude; the packaged monitor entrypoint passed.

An initial run exposed nine failures that also reproduced against the latest
published source. Preparation corrected reserved-image candidate tracking,
updated stale/incomplete test fixtures and moved scheduling-test artifacts into
isolated storage. The affected checks passed (208 tests and 18 subtests, plus the
same development-only skip) before the final full run. Runtime limits and safety
assertions were preserved. Release, landing-page and reference-export checks
also passed: 70 tests, included again in the final full suite.

All five reference packages passed their existing checksum, file-set and
applicable bundle checks. No new live campaign or benchmark was performed.
This is release validation evidence for v0.1.1, dated September 12, 2026.

How to read the evidence

Evidence must retain its scope:

Evidence	What it establishes	Limit
Current release metadata, package and regression checks	Tested behavior and public-file integrity for the prepared source	Not proof of every live model/workflow
Deterministic/fake Self-Attack and its owner tests	Runtime authority and evidence invariants with controlled providers	Not live Ghost interpretation or provider quality
September 6 Self-Attack summary	Historical fake PASS and representative live-gate PASS: six of seventeen profiles, ten selected cases	Full live conformance remains INCOMPLETE; no new live campaign is implied
Five curated reference packages	Historical completed examples with saved outputs, identity-bound evidence and checksum manifests	Publication-copy verification is not a fresh execution on 0.1.1
Performance investigations	Measurements for their recorded source, workload and environment	No universal speedup or relaxation of integrity gates

The references cover a linked website with images/audio, continuous narration,
a TTS/STT roundtrip, three independently inspected images, and saved JSON read
by an HTML consumer. Their original dates, response/frame IDs, media and manifests
are preserved. Use the package verifier to check a downloaded example:

python3 scripts/export_reference_run.py --verify examples/reference-runs/2026-09-07-saved-json-read-html-verified

For new local deterministic evidence, follow Self-Attack.
Raw captures, production Ledgers, local benchmark corpora, credentials and model
files are excluded. Do not relabel a historical run as current release validation.

Known limits and release artifacts

The known limitations and release scope
remain applicable. In particular, ordinary finalizer persistence errors can be
logged without becoming request errors. HTTP success or an in-memory completed
frame alone is not a durable-commit receipt. Verify matching durable state and
saved-artifact evidence; Readiness retention does not repair this limitation.

The source archive contains an empty registry and a MANIFEST.sha256 covering
its staged files except the manifest itself. The archive's own SHA-256 is separate.
A mutable public-main checkout does not carry that archive manifest. Version,
release date and tag/commit should be recorded with any published result. Use
CITATION.cff for software attribution and the release date.

Download verification

Download ollmo-0.1.1.tar.gz and its .sha256 companion from the release assets.
In their download directory, verify with:

shasum -a 256 -c ollmo-0.1.1.tar.gz.sha256

Archive SHA-256: f440c9e6ac50b011f90d1127eaa52602ea1a5e1fc37afbf116a6db0c79a08a0e.

Source commit: d2dd41a.

Ollmo 0.1.0 – Experimental Release Candidate

The first public release of Ollmo – a local-first AI runtime substrate for durable, inspectable, and continuable work.

Before final freeze, an Ollmo response remains fluid but recorded. Intent opens a possibility space; promotion turns selected possibilities into explicit obligations; connected branches materialize outputs; runtime evidence and closure review determine what may be reported as fulfilled.

The reply summarizes the moment. The response preserves the work.

See the README for setup, architecture, capabilities, and current limitations.

The attached ollmo-0.1.0.tar.gz is the curated release archive and includes MANIFEST.sha256 for verification.