v0.20.7

Fixed

• Preserve per-query arXiv discovery failures in previews, shortlists and run
  summaries. Distinguish successful empty searches from API errors and malformed
  feeds. Surface paper revisions with captured and discovered dates; recheck
  saved previews against the current library and keep revision updates preview-only.
• Treat missing, invalid and mixed-timezone freshness evidence as unknown.
  Compare explicit offsets in UTC without the filesystem timestamp tolerance.
• Replace complete active claim generations, including successful zero-claim
  retirement, while retaining append-only history. Refuse incomplete synthesis
  and changed-during-generation evidence. Keep exact citation paths and derived
  origin; exclude answers, audits, reports and hidden history from retrieval.
• State numeric-only versus prose verification coverage. Propagate strict
  refusals to nonzero CLI exits and retain refusal receipts without replacing
  previously accepted artifacts. Add ingest --verify consistently.
• Treat rejected site seeds as failed captures with a reason, rather than empty
  success. Refused paper publication fails its CLI/MCP item, and an all-failed
  paper batch stops before synthesis.
• Reuse unchanged local captures only when source, analysis settings and bound
  verification evidence match. Preserve history for explicit reruns and recover
  interrupted publication. Record zero-cost no-ops in the run ledger.
  Prose model revision and threshold changes invalidate reuse. Checker failures
  remain visible and strict mode fails closed when an installed checker fails.
• Use the direct Gemini 3.8 Flash default, curated current candidates, UTC price
  steps, and a scheduled refusal for the retiring Deep Research preview.
  Preserve historical model and pricing records.
• Make local setup and failure hints route-aware; keep diagnostic paths copyable,
  support -h, shorten root help and home, align diagnostic rows, and remove
  routine animation and screen clearing. Keep color-independent state words.
• Load trusted packaged theme resources installed as hardlinks while retaining
  single-link confinement for mutable user palettes.
• Pin HHEM's foundation configuration and tokenizer as well as its outer code;
  invalidate local reuse when either revision changes. Raise patched optional
  torch, fsspec and setuptools floors and retain explicit advisory limitations.

Added

• Twelve validated web and export palettes, Dark/Light/System appearance,
  persisted choices, safe TOML and bounded GitHub archive installation, CLI theme
  lifecycle commands, accessible preview cards and copy announcements. Server
  and export share tokens and component styles.
• A CPU-only optional entailment recipe and measured tail-evidence probe beyond
  512 tokens. Scope two known pinned-model metadata diagnostics to debug while
  preserving other warnings and failures.
  Validate the separately published torch CPU wheel on Windows Python 3.15.
• Numbered first-run instructions and local installed-wheel, CPU and browser
  validation records. These remain distinct from the later 1.0 semantic gates.

Changed

• Default to Python 3.15 with blocking Linux 3.12 through 3.15, Windows and macOS
  coverage. Retain the 3.12 floor. Refresh compatibility wheels, dependency audit
  tooling and immutable Actions pins; build the container from the frozen lock.
• Retain the hash-matched native plugin evaluation for unchanged canonical
  skills and eval cases; regenerate distribution version metadata.

Documentation

• Record the authorized dependency-ordered execution plan for issues #40
  through #57, including another bug hunt, local zero-dollar journeys, and
  exact-commit CI before publication.
• Record 0.20.6 as the published maintenance release and move unfinished
  evidence-state correction, evaluation, operator, and freeze-time security
  work to 0.20.7 through 0.20.10 without changing their acceptance gates.

Verification evidence

Local validation passed 7,680 tests with 95.40% combined coverage and branch measurement, zero Pyright diagnostics and $0 local operator/CPU journeys. Final main CI passed.

Publication receipt records matching PyPI/GitHub wheel and sdist digests, all 432 runtime files matching the tagged source, eight canonical skill archives, the CycloneDX SBOM, two cryptographically verified publish attestations and the isolated installed public-wheel smoke.

Five optional Transformers version alerts remain visible with a path-specific assessment. This is not a clean optional dependency audit.

v0.20.6

Fixed

• YouTube capture no longer fails when a reused connection polls a closed
  deadline socket. The deadline and the metadata byte limits stay in effect.
• distill --cost-mode no-metered init chooses a local provider, does not
  validate a cloud key, and records the smallest installed model that bench
  rates as feasible. Preview under no-metered uses deterministic ranking when
  the configured route is metered.
• distill papers --preview saves that shortlist. --from-shortlist and
  --run-last-preview ingest the same papers and skip search and rerank.
  A fresh search that was not previewed still ranks as before.
• When Playwright's browser cannot be installed, launch uses a validated
  system Chromium. An exact page with no browser falls back to one bounded
  HTTPS fetch and records capture: http-static.
• distill bench stops before probing when Ollama cloud is enabled, and
  distill doctor reports that state. Optional-package hints follow uv or
  pipx. Setup outside a source checkout writes the per-user env file, which
  later commands load from any working directory.

v0.20.5

Fixed

• Add missing context limits for Gemini 3.8 Flash and current OpenRouter
  candidates, including provider-prefixed model IDs, so known routes avoid the
  4096-token unknown-model fallback. Price exact OpenRouter GLM and DeepSeek
  routes before supplier aliases; register distinct Prime and FlashX rates
  rather than inheriting cheaper family prices. Preserve previous bare supplier
  estimates. Preserve routed prices for versioned compatibility IDs and bare
  supplier model echoes when billed-cost metadata is absent.
• Preserve DeepSeek V4 Pro 0813's scheduled pricing qualifications. Use the
  full $1.32/$3.96 rate for budget authorization in every UTC window, rather
  than assuming the observed $0.66/$1.98 promotion always applies. Distinguish
  public endpoint minimums, conservative routing ceilings, and billed cost.
• Offer hyphenated native Claude minor-version IDs in the Anthropic catalog;
  retain dotted OpenRouter spellings in their routed catalog and price registry.
• Forward Grok 4.7 reasoning effort and omit deprecated sampling parameters for
  Gemini 3.8. Adapt LM Studio to the OpenAI SDK's supported HTTP client while
  preserving disabled redirects, disabled proxy inheritance for loopback
  inference, and disabled SDK retries. Preserve HTTP status errors even when
  the provider's message contains "refused". Refuse Deep Research polling when
  the provider omits its interaction ID, retaining the submission receipt and
  existing cleanup.
• Repair the dependency security gate found during documentation review. Raise
  runtime floors to pyjwt>=2.15.0, pypdf>=6.19.0, and urllib3>=2.8.0, and
  the development floor to virtualenv>=21.7.13. Refresh only those locked
  packages and virtualenv's required python-discovery dependency. The updated
  environment clears the 33 advisory rows reported by pip-audit on the prior
  lock, without advisory exclusions. Upstream fixes:
  PyJWT,
  pypdf,
  urllib3, and
  virtualenv.

Changed

• Refresh the locked API and capture stack to OpenAI 3.26.0, Google GenAI
  2.28.0, MCP 2.3.0, and yt-dlp 2026.8.19, with required Google Auth, jiter,
  and MCP types updates. Raise the OpenAI, Google GenAI, and yt-dlp package
  floors; bound both model SDKs to their supported major versions. The
  dated audit records primary evidence and
  compatibility checks. Defaults stay pinned pending research-quality evals.

Added

• Catalog Grok 4.7, native Claude Opus/Sonnet 5.5, and current OpenRouter
  GLM, Qwen, Grok, Gemini, and Claude candidates. Add GPT-6 Astra, GPT-6.1 Sol,
  and GPT-6 Luna metadata and long-context estimates to the reserved OpenAI
  registry. OpenAI remains an unimplemented route. Catalog review date:
  October 6, 2026; this is public metadata verification, not live API validation.

Documentation

• Emphasize dependable understanding in the README and require source-faithful
  summaries, explicit uncertainty, fair disagreement, visible collection
  boundaries, and responsible correction handling in repository instructions.
• Define the two-source correction acceptance episode in the research-desk
  doctrine and roadmap. Separate planned selective revision from existing
  capture, verification, claim history, audit, and refresh mechanisms; insert a bounded
  maintenance release under the blocker rule and shift the remaining scopes
  intact. Align both supporting release plans with the public sequence.

Validation

• The final release candidate passed 7,503 tests at 95.46% branch-enabled
  coverage on Windows/Python 3.12.13, with 4 skips, 9 deselections, and one
  upstream Starlette deprecation warning. The final billing corrections also
  passed 132 focused cost, execution, and real-SDK tests. The release requires
  full CI on its exact main commit before tagging;
  final run receipts are retained in the release pull request and workflows.
• Frozen resolution, lint, formatting, Pyright with warnings denied, import
  contracts, Bandit, dependency auditing, public contracts, both generated
  skill distributions, archives, and isolated installed-wheel checks passed.
• A fresh public catalog read covered all 19 curated routes, including
  scheduled overrides, and checked the four reviewed SDK and extractor
  versions against the latest published stable package metadata. No metered
  model call was used; default routes remain pinned pending semantic evaluation.
• Reuse the authorized native plugin behavior evidence from 0.20.1: all six
  current case and report hashes match the retained successful receipts, and
  skill, reference, eval, and bundled runtime content is byte-identical. Only
  package version metadata and generated distribution instructions change.
  No new native run or waiver is claimed.

Release evidence: PR #31, main CI, publication workflow, and PyPI 0.20.5. GitHub and PyPI wheel/source hashes match; the published wheel passed the isolated installation checks.

v0.20.4

Fixed

• Account for every URL a site crawl visits. _extract_page previously returned
  nothing for both a navigation failure and a page that rendered with no usable
  text, and the crawl loop discarded those URLs without a record, so an operator
  could not tell an empty page from a blocked one. A whole seed that exhausted
  its wall-clock or memory budget, or whose worker returned a result the parent
  refused to trust, also came back as an empty page list that read like a clean
  crawl of nothing. Each URL now leaves either a page or a typed capture receipt
  carrying its outcome, HTTP status, crawl depth, and a bounded detail.
• Bound the attempted pages in a site crawl, not only the captured ones. The
  crawl loop stopped at max_pages captured pages, but a page that failed to
  capture never advanced that count, so one successful page handing back a
  frontier of up to 512 links could be followed in full. A max_pages=8 request
  could therefore perform hundreds of navigations against the host. Attempted
  visits now carry their own ceiling proportional to the requested budget, and
  a crawl stopped by that ceiling records a receipt saying so.
• Raise the security floor on anyio to >=4.14.2. The previously resolved
  4.13.0 carries CVE-2026-63374, where a TLS connection to an internationalized
  domain name could be satisfied by a certificate issued for the IDNA 2003
  encoding of that name, and CVE-2026-64847, an undrained process-pool stderr
  pipe that wedges the awaiting call. anyio reaches Distill through httpx,
  openai, google-genai, mcp, and fastapi; the floor is declared directly
  so an installed distillr cannot resolve a vulnerable version.

Changed

• Wait on observed network quiescence, under a bounded ceiling, before running
  the page extractor. domcontentloaded fires before a client-rendered page has
  its content, so the previous flat 1200 ms wait captured a hydrating
  single-page app half-built while taxing a static page for nothing. A page that
  never goes idle is still captured rather than failed.
• Redact capture receipts at the persistence boundary. A crawl can follow a link
  whose query carries a session token, and browser exception text quotes the URL
  it was navigating to, so receipt URLs are reduced to the same scheme, host,
  explicit port, and path view that every other stored URL already receives,
  including URLs found inside exception detail.
• Report capture outcomes to the operator. Site ingest prints a per-crawl
  rollup, records one run-summary issue per failed URL, writes
  attempted_pages, captured_pages, failed_pages, capture_failures, and
  capture_failure_counts into the site manifest, and includes a failed count
  in the site status line. An empty crawl now says why it was empty.
• Record a page that could not be read as a warning rather than an error, so a
  single unreachable link no longer marks an otherwise clean site run partial.
  An outcome that stopped the whole crawl, such as an exhausted resource budget
  or a refused worker result, stays an error because work was genuinely left
  undone.
• Report unreadable pages over MCP. site_batch returns failed_pages on a
  seed row whenever a crawl could not read some of its URLs, because an agent
  reading the JSON payload never sees the console rollup and would otherwise
  treat a partial capture as complete. The field is omitted when nothing failed.
• Correct the roadmap workstream table, whose version targets had drifted two
  releases behind the versioned execution sequence and assigned future outcomes
  to already-shipped releases.

Added

• Add crawl_site_with_receipts and the SiteCrawlResult and CaptureFailure
  types to distill.ingestors.sites. crawl_site keeps its existing
  page-list contract and is now a thin wrapper, so no caller has to change.

• Document capture receipts where an operator or agent will look for them:
  the site.json field list in outputs.md, the website workflow in
  usage.md, and the site_batch response shape in mcp.md.

• Add a web capture fidelity section to the roadmap recording what site capture
  already does, what remains, and which published scraping techniques are
  declined and why. Density-threshold content pruning and anti-detection
  tooling are both declined, with reasons.

• Add pricing and catalog support for Zhipu AI GLM (z-ai/glm-5.3-flash,
  z-ai/glm-5.3) and Alibaba Qwen (qwen/qwen3.8-flash, qwen/qwen3.8-max-0902,
  qwen/qwen3.8-max) models via OpenRouter.

• Map z-ai and qwen author slugs in OpenRouter routing policy to direct registered
  model identifiers, enabling strict pre-call price ceilings, budget preauthorization,
  and exact immutable ledger accounting.

• Validate candidate models across ask (adversarial risk/hallucination fixtures),
  paper (extraction/limits), and video (two-pass synthesis) workloads under a
  strict spend budget with immutable ledger recording.

Internal

• Bump the browser worker result schema to version 2, adding a bounded,
  strictly parsed failures array beside pages.
• Split distill/ingestors/sites/scraper.py back under the 1000-line module
  cap by extracting records.py (the SitePage, SiteSeed, and SiteBatch
  types shared by the crawler, the manifest parser, and the worker boundary)
  and batch.py (site batch manifest parsing and its validators). scraper.py
  re-exports both, so no import changes.

v0.20.3

Changed

• Promote seven core subpackages (commands, pipeline, mcp, prompts, concepts,
  claims, agent_skills) to strict Pyright checking in pyproject.toml, bringing
  strict type enforcement across 64.7% of repository files with zero errors and
  zero warnings.
• Decompose commands and worker task modules into focused domain units,
  extracting distill.commands._command_cost and distill.worker._claim_envelope
  to ensure all modules stay cleanly under 915 lines.

Added

• Add comprehensive unit test suites covering CLI monitor option branches,
  backward-compatibility proxies, report materials, brief chunking edge cases,
  and concept contradiction malformed rows, raising total branch coverage to
  95.43%.

v0.20.2

Added

• Add pricing and catalog support for latest frontier models: Gemini 3.8 Flash
  (gemini-3.8-flash), Anthropic Claude Fable 5.1 (claude-fable-5.1) and
  Mythos 5.1 (claude-mythos-5.1), and DeepSeek models (deepseek-v4.1-flash,
  deepseek-v4-flash, deepseek-v4-pro, deepseek-v3.2, deepseek-chat, and
  deepseek-r1).
• Map deepseek author slugs in OpenRouter routing policy to direct registered
  model identifiers, enabling hard budget preauthorization, registered price
  ceilings, and accurate ledger accounting for OpenRouter DeepSeek models.
• Set deepseek/deepseek-v4.1-flash as the recommended default model when
  OpenRouter is configured, and provide OpenRouter models and pricing in
  distill provider list openrouter.
• Support global fallback workflow budgets (default=<usd> and all=<usd>)
  in DISTILL_COST_WORKFLOW_BUDGETS so unlisted commands inherit a single
  safety ceiling against unexpected spend.

v0.20.1

Changed

• Require MCP SDK >=2.2.0,<3 and lock the SDK and types to 2.2.0, with
  modern discovery, direct modern requests, and legacy stdio regression tests.
• Update Agent Plugins 1.0.0 guidance and generated packages to its published
  status. The canonical manifest schema remains unchanged.

Fixed

• Require fresh Ollama daemon and exact-model proof before each inference
  attempt. Cloud-enabled, unsupported, ambiguous, or remotely backed routes
  fail closed in every cost mode; anomalous remote usage is not recorded as
  local zero-cost inference. Doctor, setup, and benchmark cleanup reuse that
  admission boundary.
• Serialize and pace canonical arXiv query API requests within a process,
  including retries and redirects, under the total fetch deadline. Preserve
  failure diagnostics and close redirect responses when pacing fails.
• Isolate eval CLI tests from developer Ollama services and hardware probes.
• Support older FFmpeg versions when measuring decoded audio duration, while
  retaining the process deadline and compressed-timestamp regression check.
• Make the migration-error display test independent of terminal width.
• Grant native skill evaluations read-only reference access and enough turns
  to complete. Clarify that workflow rubrics grade the requested explanation,
  preserving every required step and the existing pass threshold. Use Sonnet
  judging after observed false negatives from the default judge.
• Verify that unexpected MCP tool failures preserve the internal exception
  while the SDK exposes its generic tool error, without depending on leaked
  internal exception text.

Documentation

• Refine the canonical repository agent instructions with verified quality
  commands, implementation owners, evidence states, the existing scratch/index
  convention, and test-isolation requirements while preserving provider and
  corpus safety rules.
• Record the September 13 README and roadmap review, the evidence-generation
  acceptance matrix, and the proposed research evaluation and follow-up plan.
• Correct obsolete notebook source-discovery and research-persistence
  comparisons, and make the blocker-patch policy apply to the current release
  line.
• Reserve this release for maintenance and billing correctness. Evidence
  generation/origin work remains next in 0.20.2; research evaluations,
  operator evidence, and strict/security closure follow in 0.20.3 through
  0.20.5.

Upgrade notes

• Ollama now needs the supported local-only status contract published in
  version 0.34.0. Disable cloud on the daemon and restart it; upgrading an
  unsupported server is required. A Distill-only environment setting cannot
  change an already-running daemon. See local setup.
• arXiv pacing covers canonical query calls in one process. Operators still
  coordinate separate processes, machines, and other API or RSS clients.

Validation

• The frozen release candidate passed 7,339 tests with 95.29% branch coverage
  on Linux/Python 3.12, Ruff lint and formatting, production Pyright with
  warnings denied, all four import contracts, Bandit, and dependency auditing.
• Public contracts and both generated skill distributions passed validation.
  The wheel, source distribution, and skill archives built; an isolated wheel
  installation passed CLI, bundled-skill, runtime-asset, and editorial-default
  smoke checks.
• All six current native plugin behavior cases passed every criterion over
  three runs per arm, with a no-plugin comparison. Four unchanged cases passed
  in the full run; the worker and corpus cases passed focused reruns after
  rubric and fixture corrections. Input hashes and reported prompts/rubrics
  were checked against the final package. Failed attempts remain preserved.
• Native evaluation used Claude Code 2.1.270, Opus 5, and Sonnet judging under
  authorized account-backed runner usage, with API-key overrides removed for
  those processes. No waiver was used. These workflow cases do not establish
  end-to-end host installation or a research-quality baseline. Publication
  still requires successful CI on the exact tagged commit.

v0.20.0

Added

• Standalone perspective-editorial Agent Skill with private person or company
  context, host-provided research and document tools, complete editorial review,
  and optional Distillr/Retonr adapters. Reproducible skill, Agent Plugins 1.0.0,
  and Claude/Codex compatibility packages include behavior cases and CI checks.
• Opt-in distill editorial init, preview, capture, and run commands:
  private author perspectives, current publisher receipts, shared weekly and
  per-run budget admission, ideas and outlines, mandatory non-Anthropic paid
  rewriting, bounded evidence and editorial review, optional Retonr candidate
  checks, and unbranded Markdown and DOCX exports.
• A dated model comparison and implementation plan for the private editorial
  consumer. Explicit supplied-receipt input supports external research agents.

Changed

• The requested editorial release advances independently of the remaining
  research-desk hardening work. That work moves to 0.20.1 through 0.20.4;
  its completion and the 1.0 stability gates remain outstanding.
• CI and release builds include the standalone editorial skill archives and
  verify that the installed wheel contains the generic private-profile template.

Validation

• Local checks passed more than 7,200 tests with over 95% branch coverage, lint, formatting,
  production type checks, import contracts, security and dependency audits,
  package validators, and exact-wheel editorial smoke tests.
• The default AI news sample passed evidence and editorial review and was
  exported to Markdown and a visually checked three-page Word document.
• Six text-only standalone skill scenarios passed model review, including one
  bounded retry after an empty provider response. Native client installation
  and account-specific Cowork trials are not claimed by these checks.

Fixed

• CLI contract snapshots normalize path defaults across Windows, Linux and
  macOS, with regression coverage for both Windows and POSIX path objects.
• OpenRouter text requests now exclude separate per-request fees when enforcing
  registered token-price ceilings. OpenRouter, Ollama and LM Studio preserve
  completion reasons so editorial passes can reject truncated responses.
• Updated the HTTPX2 development dependency and lockfile to 2.12.0, resolving
  the five advisories reported by the dependency audit for HTTPX2 and HTTPCore2.
• The offline golden eval test now stubs both model judges, avoiding blocked
  network retries from its unmocked faithfulness pass.

v0.19.74

Added

• Python 3.15 prerelease watch. An advisory Linux CI lane now compiles the
  source with the current Python 3.15 release candidate, probes the locked
  dependency set, and runs the full unit suite and CLI smoke when that set is
  installable. Dependency gaps remain visible as workflow warnings. The
  supported production matrix remains 3.12 through 3.14 until the final release
  and cross-platform dependency wheels are available.
• Optional OpenRouter analysis route. A concrete lowercase author/model
  slug can now use OpenRouter as an explicit metered provider. The adapter
  requests Zero Data Retention and denies data-collection endpoints by default,
  disables hidden SDK retries, applies a registered upstream price ceiling when
  available, shapes token and temperature fields from the current no-cost
  endpoint capability catalog, uses a stable per-run upstream session, carries
  exact billed cost and selected-endpoint identity into telemetry and the usage
  ledger, and participates in doctor and distill eval.
• Budgeted OpenRouter smoke path. Provider setup now documents an explicit
  one-run route check with a hard aggregate doctor budget, followed by
  optional persistence only after validation succeeds.

Changed

• Versioned roadmap execution. The public roadmap now names the exact
  dependency order from 0.19.74 through 1.0.0, defines exit evidence for
  every release, forbids calendar and duration estimates, and requires roadmap,
  changelog, user-doc, example, schema, and evidence updates as implementation
  lands.

Safety

• Dynamic OpenRouter routing fails closed. Auto/free routers, moving aliases,
  and endpoint variants are rejected. no-metered blocks every OpenRouter route,
  and a hard budget rejects unregistered model prices before provider contact.
  Provider HTTP errors are reduced to status-class messages so account metadata
  from upstream error bodies cannot enter logs or terminal output. Payment
  failures are permanent, and key diagnosis reports provider-side spending
  limits without making an inference call.
• Browser doctor cleanup. Browser readiness now requires a real headless
  Chromium launch through the same unsafe-override refusal used by setup. This
  catches broken runtime dependencies and prevents Playwright driver tasks from
  leaking warnings during doctor shutdown.

Validation

• Complete release gate. The Windows Python 3.12 release suite passes 7,144
  tests with 4 expected skips and 9 deselected tests at 95.20 percent branch
  coverage. Ruff lint and format, generated contracts, generated distributions,
  and the frozen lock all pass.
• Native plugin evaluation unavailable. Claude Code 2.1.247 returned its
  early-access block before running the requested three-run, $5-capped
  with/without evaluation, so it incurred no model spend. Publication follows
  the maintainer's explicit release instruction. Substitute evidence is the
  deterministic cross-client distribution check and build, installed-wheel
  bundle smoke, complete static and security gates, and full branch-coverage
  suite.
• Live OpenRouter route. The installed public CLI completed its selected
  openrouter/x-ai/grok-4.6 doctor probe under a process-local $20 hard cap.
  OpenRouter reported $0.00069 billed cost for 207 input and 78 output tokens,
  and Distill persisted the provider, call, token, and exact-cost receipt without
  exposing the key.

v0.19.73

Added

• Research-desk product doctrine. The human analogue, evidence-portfolio
  model, research-state boundaries, source contribution contract, field-model
  synthesis, meaningful-change vocabulary, stopping states, evaluation
  fixtures, product boundaries, and feature-admission rubric now form one
  explicit product-quality standard.
• Evidence-anchor and atomic-claim handoff design. The accepted direction
  separates source assertions from downstream judgment, defines current claim
  generations and derived-origin preservation as the first trust correction,
  and stages digest-bound anchors, lineage, typed relations, and a versioned
  read-only claim packet behind semantic and authority gates.

Changed

• The roadmap now starts from the research role. Near-term work finishes
  evidence correctness and 1.0 stability first. Post-1.0 work proceeds through
  a research-desk evaluation baseline, inquiry maps, source-role handoffs,
  portfolio selection, one cross-source field model, meaningful refresh,
  selective updates, reading paths, and honest stopping.
• Agentic synthesis is framed as bounded field-model reconciliation. A
  novel thesis is optional, research sufficiency is a semantic judgment, and
  deterministic code owns identities, budgets, approvals, exact actions, and
  stop limits. Provider breadth and autonomous activity are enabling
  capabilities rather than product outcomes.

Validation

• Native plugin evaluation waiver. The required
  claude plugin eval command returned its early-access block before running.
  Publication proceeds under the explicit human waiver granted on 2026-08-30.
  Substitute evidence includes the current generated-distribution check,
  deterministic agent archive build, isolated Python 3.12 wheel smoke test,
  complete static and security gates, and the full branch-coverage suite.