D Research v3.2.1

Highlights

D Research v3.2.1 makes research workflows easier to deploy, more reliable across
languages, and better suited to publication-quality output.

• More relevant local search — Use a production-capable semantic backend when
  available, with a fast deterministic fallback that works without additional services.
• Cleaner bibliographies — Export richer BibTeX records for articles, books, and
  conference papers while preserving safe fallback behavior for incomplete metadata.
• More reliable multilingual research — Improve language detection while keeping
  processing fully local.
• Seamless upgrade — Keep existing research plans, evidence ledgers, routes, and
  installation paths. No migration or new mandatory dependency is required.

What’s improved

Semantic retrieval

• The default auto mode uses an installed sentence-transformers model when available.
• Environments without the optional model backend automatically use deterministic
  local-hashing; the test stub is selected only through explicit configuration.
• Model loading and encoding failures now return controlled, actionable diagnostics.
• Stricter index validation rejects malformed vectors, duplicate keys, non-finite values,
  invalid entries, and blank queries before retrieval.

Citation export

• Validated metadata sidecars support conservative @article, @book, and
  @inproceedings BibTeX entries.
• Structured personal names and literal corporate authors or editors retain their
  intended identity in BibTeX and CSL output.
• Conflicting DOI identities and ambiguous metadata fail closed instead of producing a
  misleading citation.
• Incomplete rich metadata continues to fall back safely to @misc.

Language detection

• Add deterministic local langdetect through the optional language-detection extra.
• Keep the built-in trigram detector as an offline, dependency-free fallback.
• Preserve explicit backend selection for repeatable workflows and test fixtures.

Upgrade

No workspace migration is required.

Update an existing Git checkout:

git fetch --tags
git checkout v3.2.1

Install D Research as a new agent skill:

git clone --branch v3.2.1 --depth 1 \
  https://github.com/d-init-d/d-research-skill.git \
  .agents/skills/d-research

Install the production-quality optional backends only when needed:

python -m pip install -e ".[embeddings,language-detection]"

Compatibility

Component	Support
Python	3.10 or newer
Node.js	18 or newer
Browser runtime	Playwright 1.61.1 (locked)
Existing workspaces	No migration required
Core runtime dependencies	No new mandatory dependency

Quality and security

• Candidate exact-SHA CI passed across Python 3.10–3.12 and Node.js 18/20/22.
• All 27 adversarial acceptance scenarios passed.
• All 14 real-browser Chromium smoke groups passed on local fixtures.
• Real optional-backend tests passed on Python 3.10 and 3.12 without model downloads.
• Package validation passed with the frozen 199-path boundary.
• Dependency audit reported 0 vulnerabilities.
• The stable release contains no executable, dependency, workflow, route, or package-path
  drift from the tested v3.2.1-rc.2 candidate.
• The annotated v3.2.1 tag is SSH-signed and verified by GitHub.

Downloads

GitHub provides the standard Source code (zip) and Source code (tar.gz) archives
generated from the signed v3.2.1 tag.

Technical release assurance

The stable tag is bound to commit
dc07d4902361ddf15ff0dd093faa0784b2fd47ab and tag object
4f797f0cb0f75539edfc9bc9332ca4dd041e881c. The executable candidate is
520915764a97d717aaf4682e02b8aae5dc511d2f (v3.2.1-rc.2).

Live dogfood retained 128 canonical baseline/candidate bundles under one Grok Build
0.2.101 / grok-4.5 configuration. Every task completed or produced its expected
policy refusal; no task failed or remained not-run. Tier 1 and Tier 2 both produced a
SAME promotion verdict.

At the maintainer's direction, publication proceeded without waiting for an independent
GitHub review or reviewer-signoff.json. No independent reviewer sign-off, custom release
archive/checksum, or green stable-promotion provenance attestation is claimed. This
assurance limitation does not alter the shipped executable tree, which remains identical
to the tested release candidate.

Audit trail: release PR #14 ·
tag workflow ·
raw evaluation evidence

Full changelog: v3.2.0...v3.2.1 ·
Detailed candidate notes

D Research v3.2.1-rc.2

D Research v3.2.1-rc.2

v3.2.1-rc.2 Release Notes

D Research v3.2.1-rc.2 supersedes rc.1 with a release-assurance correction
discovered by the first live candidate-tag run. The skill behavior, helper
implementations, dependency graph, evidence-ledger schema, and read-only
research contract remain unchanged from rc.1.

What changed since rc.1

• The default-branch provenance workflow now reads the workflow-run webhook
  from GitHub Actions' guaranteed GITHUB_EVENT_PATH environment variable.
• The previous expression-derived alias could resolve to an empty string in a
  live workflow_run, causing independent artifact selection to fail before
  checksum, tag-signature, archive-reproduction, and provenance checks ran.
• The dynamic contract and its mutation self-test now reject that alias and
  require both webhook validation stages to use GITHUB_EVENT_PATH directly.
• Candidate metadata, documentation, and the frozen stable-promotion contract
  now bind the release line to v3.2.1-rc.2.

The correction does not weaken any gate. The tag-triggered archive workflow
remains read-only, while OIDC and attestation write permissions remain confined
to the default-branch verifier after it independently selects the exact
upstream artifact.

Candidate scope

The production-capable optional upgrades introduced in rc.1 are retained:

• semantic retrieval defaults to an offline auto path that prefers local
  sentence-transformers and otherwise uses deterministic local-hashing;
• validated metadata sidecars enable conservative @article, @book, and
  @inproceedings export with safe @misc fallback;
• language detection supports deterministic local langdetect and a stdlib
  trigram backend without adding a mandatory dependency;
• CI exercises the real optional backends offline across the supported runtime
  matrix.

Compatibility

Python 3.10 or newer and Node.js 18 or newer remain supported. Existing
research-plan workspaces, evidence ledgers, routes, and install paths require no
migration. No mandatory runtime dependency was added.

Stable promotion gate

Stable v3.2.1 must be promoted from this exact candidate and requires:

1. Complete live Tier-1 and Tier-2 baseline/candidate dogfood under one
   identical runtime, model, tool configuration, and evaluator binding, with no
   failed or not-run tasks and all frozen comparison thresholds satisfied.
2. A GitHub-verified annotated v3.2.1-rc.2 tag, exact-SHA CI for the candidate
   and metadata-only stable commit, and pinned ancestry from v3.2.0.
3. Schema-1.2 promotion evidence recomputed from canonical schema-2.1 raw run
   bundles and bound to every score artifact by SHA-256.
4. An independent trusted GitHub APPROVED review on the exact stable commit,
   plus a reviewer sign-off bound to the promotion-manifest digest.
5. A metadata-only RC-to-stable transition, verified source archive and
   checksum, extracted-tree replay, independently reproduced archive, and
   GitHub build-provenance attestation.

Any executable, dependency, workflow, route, or package-path change after this
candidate is dogfooded requires another release candidate and a complete rerun.

Verification status

This document freezes the candidate contract; it does not pre-claim dogfood,
review, CI, archive, or provenance results. Those results are reported only
after the exact artifacts and GitHub checks exist.

D Research v3.2.0

D Research v3.2.0 is a major workflow, security, and reliability upgrade. Compared with v3.1.1—which focused on skill metadata compatibility—v3.2.0 upgrades the complete research lifecycle: planning, execution, evidence integrity, browser collection, synthesis, reporting, evaluation, and release verification.

Upgrade at a glance: schema-2.0 research plans, explicit execution gates, signed evidence ledgers, safer browser automation, deterministic resource limits, stronger citation/report validation, and a cross-platform release pipeline.

Highlights

• Research Workflow 2.0 — Separate research and synthesis phases with explicit plan_ready, execute_ready, synthesize_ready, and release_ready gates.
• Evidence you can audit — HMAC-signed ledgers, claim/process/blocker records, stable reproducibility assertions, and 100% authored-claim coverage before release.
• Safer public-data collection — Read-only browser enforcement, SSRF protection, robots and TLS policy, credential isolation, same-origin pagination, and bounded network/output usage.
• Production-grade reporting — Preserve authored narrative while generating evidence and references; reject unsupported claims, malformed references, unsafe paths, and invalid signatures.
• Deterministic quality controls — Structured dogfood and frontier benches, hostile-input fixtures, mutation checks, performance budgets, and fail-closed promotion validation.
• Reproducible delivery — Cross-platform CI, deterministic package boundaries, versioned source archives, SHA-256 checksums, archive replay, and build provenance.

What’s new since v3.1.1

Research planning and execution

• Introduces research-plan schema 2.0 with explicit research and synthesis task phases.
• Adds immutable approval fingerprints so changes to research intent automatically revoke stale approval.
• Adds portable output-tree locking to prevent duplicate, aliased, overlapping, or concurrently written task outputs across Windows and Unix-like systems.
• Replaces free-form completion checks with the canonical DRC-001–DRC-037 reproducibility contract.
• Adds a loss-preserving v1-to-v2 migration path with byte-for-byte backup, phase inference, custom-gate preservation, approval revocation, and re-rendering.

Evidence, citations, and reports

• Extends the evidence ledger with optional claim, process, and blocker record types.
• Requires real HMAC verification for signed-ledger gates through D_RESEARCH_LEDGER_KEY.
• Preserves the authored report narrative while regenerating bounded Evidence and References sections.
• Adds strict [ref:claim_id] linting, authored-claim coverage checks, and terminal synthesis-output requirements.
• Improves source scoring with separate base and adjusted totals, explicit review gates, and publication-based freshness.
• Hardens BibTeX/RIS export with deterministic unique keys, escaping, exactly one record terminator, and year-only normalization.

Browser, API, and access safety

• Adds connection-bound SSRF defenses for browser and direct-HTTP routes, including hostile redirects and IPv4-mapped IPv6 cases.
• Enforces read-only browser methods and blocks service-worker or WebSocket paths that could bypass the guarded request flow.
• Adds deterministic limits for request count, response size, aggregate network traffic, runtime, memory-sensitive extraction, and final output size.
• Strengthens robots handling, TLS verification, timeout behavior, credential redaction, cache isolation, and same-origin pagination.
• Prevents credential-bearing responses from entering shared caches unless private cache mode is explicitly selected.
• Records blocked access, truncation, and completeness limits instead of converting them into false success claims.

Evaluation and release engineering

• Adds schema-validated dogfood and frontier suites with 12 and 52 tasks respectively.
• Adds hostile-content, integrity, path traversal, credential leakage, false-completeness, degraded-runtime, and mutation regressions.
• Expands CI across Python 3.10–3.12 and Node.js 18/20/22, with real Chromium integration on Ubuntu and Windows.
• Adds deterministic package-manifest validation, clean archive replay without .git, SHA-256 verification, and GitHub provenance attestation.
• Adds exact-SHA and annotated-tag binding so release evidence cannot be reused for another revision.

Compatibility and upgrade

• Python: 3.10 or newer
• Node.js: 18 or newer
• Browser runtime: Playwright 1.61.1 (locked)
• Existing schema-v1 plans remain readable until v4, but migration is required before using the new synthesis and release gates.
• In-place migration creates research-plan.json.bak and revokes the previous approval. Review, render, and approve the migrated plan again before execution.
• Signed-ledger gates require D_RESEARCH_LEDGER_KEY; never store this key inside the research workspace.

See the v3.1.1 → v3.2.0 upgrade guide for Bash and PowerShell migration steps, verification, and rollback.

Verification

• 23/23 required local release checks passed.
• 27/27 adversarial acceptance scenarios passed.
• All 14 real-browser smoke scenario groups passed.
• Deterministic quality evaluation passed 3/3 runs.
• Promotion anti-spoof suite passed 46/46 checks.
• Python 3.10–3.12, Node.js 18/20/22, Ubuntu Chromium, and Windows Chromium CI passed on the exact release SHA.
• Dependency audit reported 0 moderate-or-higher vulnerabilities.
• Source archive replay, checksum verification, and GitHub provenance attestation passed.

Downloads

GitHub provides the standard Source code (zip) and Source code (tar.gz) archives generated from the v3.2.0 tag.

Release assurance and residual risk

The stable release is bound to commit 8156b460b358636e69d66c10c16800045d258446 and the audited candidate 2974893c77415686b6bcd1d05b6b1f6738a4f320.

The version-scoped maintainer decision records four external assurances that were explicitly waived rather than represented as completed work: GitHub-verified candidate and stable tag signatures, an independent GitHub reviewer, and uncontaminated live Tier-1/Tier-2 dogfood against v3.1.1.

Annotated tags, candidate ancestry, exact-SHA CI, semantic metadata freeze, archive replay, SHA-256 verification, and provenance attestation were non-waivable and passed.

CI evidence: candidate matrix, stable matrix, and source archive/provenance.

Full changelog: v3.1.1...v3.2.0 · Detailed v3.2.0 notes

v3.1.1

Skill Metadata Compatibility Patch

v3.1.1 Release Notes

D Research v3.1.1 is a focused metadata-compatibility patch for the skill
entrypoint. It does not change the research workflow, routing logic, helper
scripts, evidence-ledger schema, or trigger wording. The release makes the
SKILL.md frontmatter more robust for YAML parsers by converting the long
description from a single plain scalar into folded block scalar syntax.

What's New

• SKILL.md now uses description: >- for the long skill description. This
  preserves the exact trigger language while avoiding ambiguity around
  colon-bearing text such as Triggers:.
• README.md and README.vi.md now include v3.1.1 in the v3.x release sequence
  so users can distinguish the metadata-hardening patch from the broader v3.1.0
  documentation-polish release.
• Package metadata now reports version 3.1.1 in pyproject.toml,
  package.json, and package-lock.json.

Why It Matters

The skill entrypoint is the first file a runtime reads when deciding whether and
how to load D Research. The previous one-line description was valid in many YAML
parsers, but long plain scalars containing colon-bearing phrases can be brittle
across stricter or less forgiving parser implementations. v3.1.1 removes that
parser ambiguity without narrowing the skill, changing trigger intent, or
touching the operational research instructions.

Compatibility

• No workflow behavior changes.
• No trigger text changes.
• No new dependencies.
• No evidence-ledger schema changes.
• No script CLI changes.
• Existing v3.1.0 workspaces, ledgers, reports, eval fixtures, and release tags
  remain valid.

Upgrade Notes

Pull v3.1.1 if you want the safest public skill metadata surface. No migration is
required from v3.1.0; this release only hardens the YAML frontmatter format while
preserving the same D Research behavior.

v3.1.0

Release Consistency and Documentation Polish

v3.1.0 Release Notes

D Research v3.1.0 is a product-polish release for the public documentation and
release surface around the register/jargon upgrade. It does not change the
runtime workflow shipped in v3.0.5 and v3.0.6; it makes the release artifacts,
eval documentation, and reference cross-links consistent enough to ship as a
commercial-quality skill package.

What's New

• Release notes now follow a consistent product format: a concise title followed
  by a vX.Y.Z Release Notes subtitle. The v3.0.5 and v3.0.6 release-note
  artifacts were updated to match the public release style used by earlier
  releases.
• docs/eval.md now correctly documents frontier bench 2.2 as 52 tasks across
  26 classes, including the register-jargon-recall class added in v3.0.6.
• references/frontier-search.md no longer repeats the same
  references/register-and-jargon-expansion.md See also entry twice.
• README.md and README.vi.md now summarize the v3.0.5, v3.0.6, and v3.1.0
  release sequence so users can see how the register/jargon work moved from
  method, to tool and bench, to release polish.

Why It Matters

The register/jargon upgrade is already functional and tested, but product-grade
skill packages also need clean release surfaces. Users should see the same naming
pattern across GitHub releases, release-note artifacts, README summaries, and eval
documentation. v3.1.0 removes the last small inconsistencies so the repository is
ready to ship, review, and install without caveats.

Compatibility

• No runtime behavior changes.
• No new dependencies.
• No evidence-ledger schema changes.
• No script CLI changes.
• Existing v3.0.x workspaces, ledgers, reports, eval fixtures, and release tags
  remain valid.

Upgrade Notes

Pull the new release for a cleaner public documentation surface. If you already
installed v3.0.6, no workflow migration is required; v3.1.0 preserves the same
register/jargon recall behavior and only polishes documentation, metadata, and
release consistency.

v3.0.6

Register Recall, Tooled and Benched

v3.0.6 Release Notes

D Research v3.0.6 finishes the register- and jargon-aware recall work introduced
in v3.0.5. v3.0.5 shipped the method — a bidirectional register ladder and a
guardrails-first companion. v3.0.6 makes that method runnable and
regression-protected: the cross-source recurrence rule is now a deterministic
helper script, and the capability is covered by the frontier eval bench.

This is a purely additive release. No existing behavior, evidence-ledger schema,
or script CLI changes.

What's New

• scripts/harvest_terms.py turns the "keep only terms recurring across two
  or more independent community sources" rule into a deterministic, offline,
  stdlib-only tool. Feed it tagged source<delimiter>term occurrences and it
  counts the distinct sources behind each candidate term, then labels each one
  confirmed (at or above the threshold) or candidate (below it). It never
  invents vocabulary — it only scores what the agent already harvested from
  fresh results. Run it via npm run terms:harvest or
  python3 scripts/harvest_terms.py harvest --threshold 2.
• A register-jargon-recall frontier eval class with two ground-truth tasks
  (FB-051, FB-052). One probes whether the agent can name and apply the
  bidirectional register ladder; the other probes the non-negotiable boundary
  that harvested vocabulary is a discovery layer, never evidence. The frontier
  bench moves from 2.1 (50 tasks / 25 classes) to 2.2 (52 tasks / 26
  classes) under its additive bench-version policy, with matching empty-score
  fixtures regenerated deterministically.

Why It Matters

A method that lives only in prose is easy to drift away from. v3.0.6 closes that
gap from both ends. The helper script makes the filtering rule executable and
auditable, so "recurs across at least two independent sources" is a command, not
a hopeful guideline. The new bench class makes the capability a tracked
regression target, so a future change that quietly weakens register-aware recall
shows up as a FAIL instead of slipping by unnoticed.

Determinism and Discovery-Only Discipline

harvest_terms.py is intentionally boring: same input, same output, no network,
no model calls. It groups terms case-insensitively, counts distinct source
labels, and applies a threshold. That determinism is the point — the tool
enforces the recurrence filter without becoming a new source of hallucinated
vocabulary. And the boundary from v3.0.5 still holds end to end: confirmed terms
are a discovery layer that opens recall, never evidence; every claim still
passes the source-quality rubric and the contradiction pass.

Also in This Release

• Removed three stale placeholder for task 2.x comments in
  scripts/pdf_extract.py. The subcommands they annotated (tables,
  to-ledger, self-test) were already fully implemented; only the misleading
  comments remained.

Compatibility

• No new runtime dependencies.
• No evidence-ledger schema changes.
• No CLI changes to existing scripts.
• Frontier bench 2.2 is additive; 2.1 score artifacts remain comparable on
  the shared task subset.
• Existing v3.0.x workspaces, ledgers, reports, and eval fixtures remain valid.

Upgrade Notes

Pull the new release and continue using the skill normally. To exercise the new
helper, pipe tagged occurrences into it:

printf 'forum-a\tbrain fog\nforum-b\tbrain fog\nblog-c\tbrain fog\n' \
  | python3 scripts/harvest_terms.py harvest --threshold 2

Run npm run self-test to confirm the full chain — including the new
harvest_terms self-test and the bench-harness consistency check — passes.

v3.0.5

Register- and Jargon-Aware Recall

v3.0.5 Release Notes

D Research v3.0.5 strengthens the recall layer of an agentic research run. It
adds a register- and jargon-aware discovery companion so the agent can match the
vocabulary of the people who actually hold the evidence — not the vocabulary the
model happened to be trained on. The result is higher recall on topics where the
authoritative discussion lives in a different register than the query: clinical
vs. lay, legal vs. street, standards vs. shop-floor, academic vs. community
jargon, or fast-moving emergent slang.

Crucially, the skill stores a process for harvesting and verifying
register-matched vocabulary at runtime, not a frozen word list. There is
nothing to maintain when slang changes — the update lives on the live web.

What's New

• A new opt-in companion, references/register-and-jargon-expansion.md, built
  guardrails-first around five core principles: the skill stores a process not a
  word list; vocabulary is harvested from fresh results at recall time, never
  recalled from model memory; harvested vocabulary is a discovery layer and
  never evidence; the register ladder is bidirectional; and native-language
  research principles stay intact.
• A bidirectional register ladder — formal → vernacular to open recall, and
  vernacular → formal to anchor every community term back to a primary source —
  so a run never gets trapped in the community basin without a primary anchor.
• A register_jargon_recall intake label that activates only when the evidence
  basin demonstrably uses vernacular, subculture, or domain jargon.
• An audit-grade templates/register-vocab-log.csv that records one row per
  harvested term — including its register level, source basin, independent
  source count, status (candidate / confirmed / rejected), and the claims
  it ultimately produced — so a reviewer can replay exactly which vocabulary was
  trusted and why.
• Wiring across SKILL.md, AGENTS.md, references/query-patterns.md,
  references/topic-decomposition.md, references/frontier-search.md, and
  references/multilingual-research.md so register expansion is reachable from
  the decision tree, the query fanout, topic decomposition, and gap-driven
  follow-up.

Why It Matters

A clinical, legal, standards, or academic query frequently under-recalls because
the people who hold the long-tail evidence use different words. Patients say
"brain fog", clinicians say "cognitive dysfunction". Practitioners abbreviate;
communities coin terms; slang moves faster than any maintained list. An agent
that searches only in the canonical register quietly misses entire source
basins — and an agent that descends into slang without anchoring back up ends up
citing community chatter as if it were fact.

v3.0.5 addresses both failure modes at the recall layer. The bidirectional
ladder opens recall and forces every community term back to a primary source
before it can support a claim. The guardrails make the hard part — not the
"add more slang" part — the center of the design.

Discovery, Never Evidence

The single most important boundary in this release: harvested vocabulary is a
discovery aid, never evidence. Terms are kept only when they recur across two or
more independent community sources; single-source terms stay flagged as
candidates and never drive a wide fanout on their own. Every resulting claim
still anchors a real source and still passes references/source-quality-rubric.md
and the contradiction pass. Person-related slang inherits the
references/person-aggregation.md privacy boundary, and slurs, harassment
vocabulary, and brigading terms are explicitly excluded from recall.

Compatibility

• No new runtime dependencies.
• No evidence-ledger schema changes.
• No script CLI changes.
• Existing v3.0, v3.0.1, v3.0.2, and v3.0.3 workspaces, ledgers, reports, and
  eval fixtures remain valid.

Upgrade Notes

Pull the new release and continue using the skill normally. The register layer
is opt-in: it activates only when recall is thin or the evidence basin uses lay,
community, or vernacular terms, and otherwise stays out of the way. For
audit-grade runs, copy templates/register-vocab-log.csv into your workspace to
keep a replayable record of which vocabulary was trusted and which claims it
produced.

v3.0.3

v3.0.3 Release Notes

D Research v3.0.3 hardens the earliest and most important decision in an
agentic research run: choosing the right research shape before opening sources.
This release expands Step 0 intake into a stronger classification controller for
teams that prefer auditability, recall, and correct routing over raw speed.

What's New

• First-class due_diligence_or_investigation routing for companies, projects,
  vendors, packages, public claims, provenance, trustworthiness, risk, and
  red-flag checks.
• First-class policy_or_standards_analysis routing for standards, RFCs,
  governance policies, compliance rules, and versioned normative texts.
• First-class creative_or_cultural_research routing for creative works, media,
  cultural history, trend analysis, archives, reception, and public discourse.
• Completeness-first depth selection for audit-grade, high-risk, due diligence,
  red-flag, and "speed is not important" requests.
• Intake cards now capture research depth, authority model/source basins, and
  red-flag or contradiction focus before discovery begins.
• research.config.example.json now exposes depth-control defaults under
  research.intake.

Why It Matters

Many research failures start before the first search query: the agent picks the
wrong authority model. A company risk check becomes a generic market overview, a
standards question becomes a blog-summary task, or a cultural trend question is
treated like a scientific consensus problem.

v3.0.3 fixes that failure mode at the routing layer. The new labels are
multi-label and composable, so they do not narrow the skill. A task can still be
due_diligence_or_investigation + technical_research + multilingual_local, or
policy_or_standards_analysis + public_url_analysis + monitoring_change, or
creative_or_cultural_research + dataset_collection.

Completeness-First Mode

When completeness-first mode is triggered, agents must use source maps, search
logs, evidence ledgers for key claims, independent recall expansion,
contradiction search, no single-basin completion claims, execution gates, and
explicit gap/blocker notes. This is the right default for risk-heavy research
where confidence matters more than speed.

Compatibility

• No new runtime dependencies.
• No evidence-ledger schema changes.
• No script CLI changes.
• Existing v3.0, v3.0.1, and v3.0.2 workspaces remain valid.

Upgrade Notes

Pull the new release and continue using the skill normally. Existing workflows
will behave the same unless the request benefits from the new routing labels or
completeness-first depth. For project-specific behavior, copy
research.config.example.json to research.config.json and tune the
research.intake depth settings.

v3.0.2

v3.0.2 Release Notes

Overview

D Research v3.0.2 introduces Step 0: Research Intake. Before an agent opens
sources or commits to a branch, it now classifies the request by research shape,
safety posture, expected artifact, freshness requirement, geography/language
scope, and required gates.

This release closes a practical failure mode in agent research: starting with
the wrong workflow. A person-related task should not begin as generic broad
research; a PRISMA request should not become a loose literature scan; a dataset
request should not end as prose; and a high-stakes topic should not lose its
primary-source posture.

What Ships

• references/research-intake.md: a detailed Step 0 controller with
  multi-label routing, hard-stop safety checks, route priority, output-artifact
  selection, ambiguity policy, and failure-mode guidance.
• SKILL.md and AGENTS.md now require intake before branch selection and
  source access.
• research.config.example.json now exposes research.intake defaults.
• README and Vietnamese README now position intake as lifecycle pillar 0.
• Version metadata bumped to 3.0.2.

Why It Matters

v3.0.1 hardened the end of the workflow with execution gates. v3.0.2 hardens
the beginning. Together they give agents a stronger operating envelope:

• classify correctly before searching;
• apply privacy and access boundaries before touching sources;
• route academic, systematic, dataset, URL, person, high-stakes, technical, and
  multilingual tasks into the right workflow;
• verify coverage and evidence before synthesis.

The intake remains deliberately multi-label. It does not narrow D Research to a
single taxonomy; it helps the agent compose the right branches when tasks
overlap.

Compatibility

• No new dependencies.
• No CLI changes.
• No evidence-ledger schema changes.
• Existing v3.0 and v3.0.1 artifacts remain valid.

Recommended Upgrade Path

1. Pull or install v3.0.2.
2. Review references/research-intake.md.
3. Keep research.intake.enabled and research.intake.multiLabel set to
   true for normal use.
4. Use research.intake.emitClassificationCard for audit-grade workflows where
   the route decision should be visible to reviewers.
5. Run npm run refs:check:decision-tree and npm run self-test before
   publishing a downstream package.

v3.0.1

#v3.0.1 Release Notes

Overview

D Research v3.0.1 is a focused workflow-hardening release for production agent
research. It takes the strongest orchestration discipline from the MiniMax
expert workflow and folds it back into the portable skill core without tying the
project to MiniMax, any specific subagent runtime, or a single source domain.

The result is a safer pre-synthesis layer: agents now have explicit gates for
source mapping, recall, source-basin coverage, identity/date discipline,
claim-level verification, blockers, and final readiness before presenting
non-trivial research as complete.

What Ships

• Portable execution gates in references/execution-gates.md.
• Optional worker-role contract for source mapping, recall auditing, public
  source hunting, data extraction, evidence verification, and report synthesis.
• Opt-in Vietnamese source discovery companion in
  references/vietnamese-source-discovery.md.
• Updated SKILL.md, AGENTS.md, README files, and configuration defaults.
• Version metadata bumped to 3.0.1.

Why It Matters

The v3.0 core already shipped a broad research stack: browser-first probing,
evidence ledgers, anti-bot fallback, frontier search, research-plan workspaces,
social archival, citation resolution, report generation, OCR, translation,
semantic retrieval, citation graphs, multi-format extraction, deduplication,
cache support, and provenance columns.

v3.0.1 does not add another heavy tool. It improves the operating discipline
around those tools. Agents are now less likely to stop after one promising
source, treat mirrors as independent evidence, overclaim completeness, infer
dates or identity from weak anchors, or synthesize before blockers and coverage
gaps are visible.

Compatibility

• No new runtime dependencies.
• No evidence-ledger schema change.
• No CLI breaking change.
• Existing v3.0 workspaces, ledgers, reports, and eval fixtures remain valid.
• Subagents remain optional. Hosts without subagent support can run every gate
  manually from the main agent context.

Recommended Upgrade Path

1. Pull or install v3.0.1.
2. Review references/execution-gates.md.
3. If your workflow involves Vietnamese or Vietnam-local sources, also review
   references/vietnamese-source-discovery.md.
4. Keep research.executionGates.enabled set to true unless you are running a
   deliberately minimal or benchmark-controlled workflow.
5. Run npm run refs:check:decision-tree and the usual self-tests before
   publishing a downstream package.

Release Positioning

This is a production patch release. It is intentionally conservative: the
surface area is mostly documentation, workflow routing, and configuration. The
upgrade improves research reliability without narrowing the skill's scope or
changing its safety model.