SCIENTIA Self-Publication Finalization Plan (2026)

SCIENTIA Self-Publication Finalization Plan (2026)

Status: Approved 2026-05-09. Granular phase-N implementation plans live under docs/superpowers/plans/ and follow the writing-plans / executing-plans / TDD discipline. This document is the strategic source of truth; phase plans must cite back to a section here.

Predecessors / inputs: ADR-011 — Scientia Publication SSOT; Mesh Integration Research 2026; Telemetry Unification Design 2026; Where Things Live; Layers SSOT.

1. Strategic thesis — what Vox should publish

Three populations of “AI publishers” exist as of 2026 and none produce what Vox is structurally positioned to.

Class	Example	Strength	Gap
Auto-paper-generators	Sakana AI Scientist v2	Cheap (~$15/paper)	Beel et al. 2025 (arXiv 2502.14297): 42% of generated experiments fail; one paper claimed energy improvements while consuming more compute. The verifier is GPT-4 grading GPT-4.
Closed-loop physical AI	Coscientist; Stanford Virtual Lab	Wet-lab grounded → Nature acceptance	Narrow domain; not transferable to LLM-substrate research.
Static benchmark consortia	HELM; LMArena; Artificial Analysis	Methodologically rigorous	Static. No longitudinal observation across providers under real workloads. The Leaderboard Illusion (Singh et al., arXiv 2504.20879) showed Meta tested 27 private Llama-4 variants pre-release.

Vox’s privileged seat. Vox is an AI-development substrate sitting between many real applications and many providers, with telemetry already capturing latency, refusals, tool-call malformation, JSON-mode violations, cost drift, contradiction ratios, and routing decisions (research_metrics_contract.rs). No one has published the longitudinal “AI epidemiology” of provider behavior under real production workloads. That is the Vox-shaped contribution.

Venue. Not NeurIPS — IMC / MLSys / TMLR. IMC explicitly publishes “we measured the internet and here’s what’s broken” papers. Co-author with one academic lab (Stanford CRFM, UK AISI, or a measurement group) so the byline is not a vendor whitepaper. Build measurements on top of UK AISI Inspect so plumbing is unimpeachable.

Reputational firewall — built into the architecture, not the comms strategy. Per Galactica and the LMArena response saga, when you publish unflattering findings about commercial providers, the response is reputational, not methodological. Four properties baked into code:

1. Pre-registration of every measurement campaign — hypothesis + eval
   • statistical test + stopping rule, signed and timestamped before data collection.
2. External symbolic verifier wherever possible — AlphaEvolve gets accepted because matrix-multiplication arithmetic can be checked symbolically; AI Scientist gets dismissed because GPT-4 grades GPT-4. Any quantitative claim must be checked against a non-LLM ground truth.
3. Embedded 14-day right-of-reply — providers see drafts before publication; their replies are inline; this is journalism’s standard, applied to ML.
4. Negative-result publication-bias inversion — the system refuses to ship a quarterly Atlas without at least one pre-registered hypothesis that failed to reject the null.

2. The signal ladder

The system is defined by what it refuses to emit. Five tiers with hard gates:

Tier	Source	Treatment	Gate to next tier
T0 Observation	Single telemetry event	Stored in research_metrics. No publication.	Aggregation + N≥30
T1 Aggregate	Rolled-up metric over a window	Dashboard only; atomic claims extracted.	VeriScore atomicity + verifiability classifier
T2 Atomic claim	One verifiable assertion with bounded evidence	Emitted as signed nanopublication. Indexed locally.	Calibrated verifier confidence + retrieval round-trip + SciFact-Open novelty
T3 Finding candidate	Bundle of T2 claims forming a coherent finding (existing FindingCandidateV1)	Existing worthiness gate; dual-approver.	Pre-registered hypothesis + external verifier + right-of-reply window cleared
T4 Publication	Manifest aggregating T3 findings	Quarterly Provider Atlas + topic-specific micro-papers	(terminal)

Today’s system handles T3 → T4 well. It is missing everything that turns T0 into T2, and the T1 → T2 → T3 ladder has no rigor gates worthy of a research publication.

3. Architectural rewrite, informed by SOTA

3.1 Resolve phantom imports first (pre-existing tech debt)

pipeline.rs:13–22 imports six modules that do not exist: claims, gate, planner, provider, types, verifier. The binary compiles only because the call site is unreachable in the current rollout config. This is Phase 0a, the first implementation plan.

3.2 Claim extraction — adopt, don’t invent

Stage	Technique	Primary source	Vox crate
Verifiability gate	VeriScore	arXiv 2406.19276	new vox-claim-extractor
Atomic decomposition	FActScore + SciClaim tuple (var, rel, var, qual)	arXiv 2305.14251; arXiv 2109.10453	vox-claim-extractor
Constrained emission	XGrammar pushdown automaton	mlc-ai/xgrammar	existing vox-constrained-gen (production-grade)
Span integrity	SciClaim span supervision	arXiv 2109.10453	vox-claim-extractor
Single-model end-to-end	SciClaims architecture	arXiv 2503.18526	reuse vox-actor-runtime/mens.rs
Grounded verification	MiniCheck-FT5 (770M; GPT-4-quality at 400× cost)	arXiv 2404.10774; Liyan06/MiniCheck	new dependency; ship as plugin
Calibrated abstention	Temperature-scale logits; ABSTAIN below τ	PMC10919922	vox-claim-extractor
Word-level hallucination tagging	RAGTruth taxonomy	arXiv 2401.00396	vox-claim-extractor

Hard rule: no T1 → T2 promotion without VeriScore + atomic + span-bounded + MiniCheck-verified + calibrated. SciFact-Open (arXiv 2210.13777) generalization gap is the cautionary tale; we evaluate against an open-corpus split before trusting any extractor.

3.3 Novelty — atomic-NEI against a timestamp-bounded corpus

• A claim is novel iff it has no SUPPORTING evidence in the timestamp-bounded prior corpus (SciFact-Open NEI semantics).
• Retrieval substrate: SPECTER2 + task-format adapter (allenai/SPECTER2) — not a generic embedding. Adapter selection matters: classification adapter for novelty, retrieval adapter for prior-art lookup.
• Corpus: OpenAlex (209M works, CC0) + Semantic Scholar; multilingual to guard against false-novelty failure (Sharma et al., SDP 2025, arXiv 2506.22026).
• Disruption Index (CD) only as a weak supplementary signal; raw CD is artifactually inflated by reference-list growth (Petersen et al. 2024).
• Conflict surface: when atomic similarity > 0.8 AND polarity differs → emit EvidenceConflict, not Novel.

3.4 Ground-truth verifier — symbolic where possible, MiniCheck where not

The AlphaEvolve lesson. For Vox’s primary research output (provider/model behavior), almost every claim has a non-LLM ground truth:

Claim type	Symbolic verifier
”p95 latency rose by X ms”	numeric comparison against ExecTimeRecord rows
”tool-call malformation rate increased”	exact-match parser on tool-call JSON
”JSON-mode violation rate increased”	JSON Schema validate (already in vox-jsonschema-util)
“model produces longer outputs”	token count
”refusal rate changed”	structured refusal classifier with controlled vocabulary
”code generated compiles”	cargo check exit code
”tests pass”	test runner exit code + assertion count

For everything that must be LLM-judged, MiniCheck is the verifier — never the same model that produced the artifact.

3.5 Continual re-verification — temporal-cutoff retrieval

• ChronoFact (arXiv 2410.14964) — extract events, build timeline, verify per-event.
• TACV (arXiv 2407.15291) — restrict retrieval to evidence available before the claim was made.
• Schedule re-verification by evidence-class volatility, not fixed cadence.

4. Artifact format — the unit of trust is not the paper

The biggest architectural divergence from the v1 sketch. The unit of trust on the modern scientific web is a signed, hashed, identifier-rich artifact graph — not a PDF.

4.1 Five identifiers per artifact

• DOI (Crossref or DataCite) for the manifest itself
• Trusty URI (content hash embedded in URI) for each atomic claim, via nanopublication
• SWHID (ISO/IEC 18670 since April 2025) for the code snapshot
• ORCID for every author (and the project)
• ROR for every affiliated organization

4.2 Each empirical claim is a signed nanopublication

Every T2 atomic claim is emitted as a Nanopub — three named RDF subgraphs (Assertion / Provenance / PublicationInfo) in TriG, signed with the project ORCID’s key, replicated through the Nanopublication Network. This solves several problems at once: per-claim addressing without per-claim DOI fees; cryptographic content verification (Trusty URI = content hash in URI); native FAIR-compliance; machine-actionable retraction.

New crate vox-nanopub (L2) reusing vox-crypto’s ed25519.

4.3 Every artifact is an RO-Crate

RO-Crate 1.2 packages code + data + text + claims + provenance as flattened JSON-LD. Mandatory ro-crate-metadata.json at the root. EOSC, WorkflowHub, and most science-funding-body deposits expect it. Every Vox publication ships as an RO-Crate with: Markdown body; signed nanopubs; eval code SWHID; dataset DataCite metadata; CITATION.cff; CodeMeta.json; license SPDX; disclosures.json (CRediT + COI + AI-tool use); PROV-O provenance graph.

New crate vox-ro-crate (L2).

4.4 TOP Level 2 across all 7 dimensions, by default

COS TOP Guidelines: seven research practices × three levels. Default to Level 2 (share + cite) across Citation, Data, Analytic methods/code, Research materials, Design/analysis, Preregistration of studies, Preregistration of analysis plans. Surface a “TOP compliance” badge in the manifest. Single highest-leverage move for credibility.

4.5 ACM Artifact Available + Reusable badges

Every Vox publication submits its RO-Crate for ACM Artifact Review.

5. Pre-registration as code

5.1 Preregistration is a typed object, not a Google Doc

PreregistrationV1 {
    id: nanopub_trusty_uri,      // the prereg is itself a signed nanopub
    hypothesis: String,          // includes direction, not just "we will measure X"
    eval_substrate: SubstrateRef { repo_swhid, eval_set_swhid, inspect_task_id },
    metric: MetricSpec { name, aggregation, units },
    statistical_test: TestSpec, // frequentist | bayesian; if bayesian: prior + threshold
    stopping_rule: StopRule { max_n, alpha, threshold },
    decision_rule: DecisionRule, // e.g. "if posterior P(direction) > 0.95, conclude X"
    cost_cap_usd: f64,
    signed_at: timestamp,
    signing_key: ed25519_pubkey,
}

The orchestrator refuses to run a campaign without a signed prereg. Modifications post-collection require a new prereg with explicit supersedes: reference.

5.2 Bayesian sequential testing as the default

Per arXiv 2511.10661 and the forking-paths failure mode (running another eval is ~free, so frequentist multiple-comparison correction is intractable): default to Bayesian sequential. Pre-declare a stopping threshold on posterior probability; sample sequentially; stop when crossed; publish whether confirmed or refuted.

5.3 Forking-paths defense

Pre-register the analysis tree, not just the hypothesis. The system records prereg signature + analysis-code commit hash; any deviation surfaces as analysis_plan_deviation: true on the publication.

6. Map to Vox primitives and crates

Vox primitive	SOTA mapping	Action
vox-constrained-gen	XGrammar	Use as JSON-schema-constrained emitter for claim envelopes.
vox-actor-runtime/mens.rs	Mens client	Host SciClaims-style single Llama-3 8B locally; tier-cascade to remote on ABSTAIN.
vox-search	SPECTER2 retrieval	Add SPECTER2 as a model option; route novelty queries through retrieval adapter.
vox-publisher/scientia_*	Atlas publication	Extend with per-claim nanopub emission + RO-Crate builder + TOP/ACM badges.
pipeline.rs	Phantom imports	Phase 0a — resolve claims/gate/verifier/planner/provider/types modules.
research_metrics_contract.rs	Provider Atlas raw signal	Wire D1–D10 events → ProviderObservation family per Mesh §4.1.
calibration.rs	Drift detection	Already gives drift z-scores; emit DriftAlert → atomic claim → nanopub.
vox-crypto	Nanopub signing	Reuse ed25519 — no new crypto.
vox-doc-pipeline	RO-Crate manifest regen	Add ro-crate-metadata.json to regen list.
vox-arch-check	Layer enforcement	Add rule: nanopub crate at L2; no horizontal L3 publisher↔scientia-ingest.

6.1 New crates (only what cannot live elsewhere)

New crate	Layer	Purpose
vox-research-events	L1	Typed event bus types + PreregistrationV1 + ResearchEventEmitter trait. ≥3 consumers (orchestrator, publisher, gamify, ingest).
vox-claim-extractor	L2	VeriScore + atomic + span-bounded + MiniCheck pipeline. No async DB; isolated unit-test surface.
vox-nanopub	L2	TriG emit + ed25519 sign + Nanopub Network publish. Reuses vox-crypto.
vox-prereg	L2	Pre-registration object, signing, deviation detection.
vox-ro-crate	L2	RO-Crate 1.2 builder.
vox-inspect-bridge	L3	Adapter to UK AISI Inspect. Translates Vox eval definitions to Inspect Task/Solver/Scorer.
vox-gamify-scientia-bridge	L3 (feature-gated)	Optional; off by default.

7. Phased plan

Each phase owns a specific signal-ladder transition. Each ends with a publishable deliverable to validate the rigor.

Phase 0 — Foundations (1.5 wk)

• 0a Resolve phantom imports (this is the first detailed plan).
• 0b Create vox-research-events L1 crate with PreregistrationV1, event types, ResearchEventEmitter trait.
• 0c Codegen Rust enums from contracts/scientia/*.schema.json.
• 0d Add tables: claims, novelty_results, prereg, publication_attempts, model_profile_learning (per Mesh §4.3).
• 0e Add 6 new SecretId::* for ORCID, arXiv, Crossref, OpenAlex, Semantic Scholar, OSF.
• 0f vox-arch-check rules per §6.

Deliverable: claim_detection_enabled flag actually works; pipeline runs end-to-end with stub claim extractor.

Phase 1 — The extractor (T1 → T2 promotion) (3 wk) — Complete

Status: Complete 2026-05-09. vox-claim-extractor shipped; SciFact-Open mini-split F1=1.000 (gate ≥0.65 ✓).

• vox-claim-extractor crate: SciClaims architecture, VeriScore atomicity, XGrammar JSON envelope, span integrity check, MiniCheck verifier.
• Vendor MiniCheck-FT5 (770M, T5) as a Vox plugin.
• Calibrated abstention (temperature-scaled logits; ABSTAIN below τ).
• Tier cascade: local Mens → remote large model only on ABSTAIN.
• Acceptance: ≥0.65 F1 on SciFact-Open held-out split.
• CLI: vox scientia claims extract <source>.

Deliverable: the first publication-eligible artifact — Vox-internal report on the extractor’s evaluation, written using its own pipeline, deposited as RO-Crate to a private Zenodo sandbox.

Phase 2 — Pre-registration + symbolic verifiers (2 wk) — Complete

Status: Complete 2026-05-09. vox-prereg shipped; Trusty URI signing, deviation detector, symbolic verifiers, Bayesian stopping rule, and campaign gate all implemented and tested.

• vox-prereg crate. Signed PreregistrationV1. Orchestrator refuses campaigns without signed prereg.
• Bayesian sequential testing default.
• Symbolic verifiers per §3.4 wired as Strategies in confidence_fusion.rs.
• Refusal classifier with controlled vocabulary.
• Analysis-plan-deviation detector.

Deliverable: prereg-protected pipeline. First three measurement campaigns must register before running.

Phase 3 — Reputational firewall (1.5 wk)

Status: Complete — reply_window.rs, retraction.rs, living_review.rs landed in vox-prereg; Provider Atlas dry-run lifecycle works end-to-end. (2026-05-09)

• 14-day right-of-reply window enforced on the manifest (refuse to publish provider_atlas topic-pack until window cleared with notification + reply ingest).
• Reply ingestion as inline content (not appendix) per IMC measurement-paper conventions.
• Retraction nanopub emission.
• Living-review semantics: each manifest version gets its own DOI; canonical URL points to “latest”; version_history block lists all DOIs.
• Crossref Labs API polling for retraction propagation.

Deliverable: Provider Atlas dry-run lifecycle works end-to-end.

Phase 4 — Nanopub + RO-Crate + TOP/ACM badges (2 wk) ✅ COMPLETE (2026-05-09)

• vox-nanopub crate (TriG emission, ed25519 via vox-crypto, Nanopub Network publish). ✅
• vox-ro-crate crate (RO-Crate 1.2 builder). ✅
• TOP-Level-2 compliance surfaced in manifest. ✅
• ACM Artifact Available auto-application via Zenodo deposit. ✅
• Highwire-style meta tags (citation_title, etc.) in SSG output for Google Scholar pickup.
• CFF, CodeMeta, SPDX, ORCID/ROR enrichment into RO-Crate. ✅

Deliverable: publication artifact spec complete. First Zenodo sandbox deposit.

Phase 5 — Inspect bridge + atomic-NEI novelty (3 wk) ✅ Complete

• vox-inspect-bridge crate. Translate Vox probes into Inspect tasks.
• Contribute Vox-defined evals upstream to inspect_evals (academic co-author handshake).
• SPECTER2 retrieval adapter in vox-search. Multilingual prior-art corpus.
• Atomic-NEI novelty per §3.3.
• ChronoFact-style timestamp-aware retrieval for re-verification.
• EvidenceConflict family for opposing-polarity high-similarity matches.

Deliverable: Inspect+SPECTER2-grounded novelty pipeline.

Phase 6 — Provider observability ledger + Mesh Integration (3 wk) ✅ COMPLETE 2026-05-09

• DiscoverySignalFamily::ProviderObservation, ::ModelCapabilityEvidence (Mesh §4.1). → crates/vox-research-events/src/schema_types.rs
• FindingCandidateClass::ModelCapabilityAtlas, ::ProviderReliabilityAtlas (Mesh §4.2). → crates/vox-research-events/src/schema_types.rs
• model_profile_learning populated by rollup_model_scoreboard_with_scientia (Mesh §5.7). → crates/vox-db/src/research_pipeline.rs
• ScientiaObservationClassifier trait. → crates/vox-research-events/src/observation.rs
• ScoringWeights extensions (Mesh §5.4) — behind feature flag, default OFF. A/B compare on held-out tasks before flipping. → ScientiaWeightExtension
• LearnedProfileRow + ModelRegistry::inject_learned_profiles() (Mesh §5.5). → LearnedProfileRow
• record_penalty_with_context (Mesh §5.6). → PenaltyRecord
• .vox automation: scripts/scientia/{probe-run,profile-rollup,atlas-draft,atlas-publish}.vox. ✅
• LEGACY_EXPORT_TABLES updated with all Phase 0d scientia tables.

Deliverable: closed loop is real. Worthiness gate prevents self-merge.

Phase 7 — Format adaptation (constrained-grammar all the way) (2 wk) ✅ COMPLETE (2026-05-09)

Status: Complete — ShortFormVariant, PublicationPlatform, FigurePolicy, validate_short_form, adapt_claim_to_platform landed in crates/vox-research-events/src/publication_format.rs; AiDisclosureBlock + AiToolUsage landed in crates/vox-ro-crate/src/ai_disclosure.rs. All 8 new tests pass.

• Every short-form adaptation goes through XGrammar/vox-constrained-gen emitter producing JSON, then a templating layer renders. ✅
• No free-form LLM text in publication path. Every short-form variant lifts from atomic claims with nanopub URIs. ✅
• Disable LLM-figure generation in primary research figures (Cell/Science 2025 policy). Schematic only, with mandatory legend disclosure. ✅ (FigurePolicy::default())
• AI-disclosure block auto-filled per Nature/Science/Cell 2025 norms (Nature AI policy). ✅ (AiDisclosureBlock::build)
• Bluesky prioritized over X per academic-Twitter migration data (arXiv 2505.24801). ✅ (X/Twitter absent from PublicationPlatform)

Phase 8 — Scholarly automation + venue strategy (3 wk) ✅ COMPLETE (2026-05-09)

Status: Complete — venue catalog YAML + Rust types; JournalFitRecommender; ORCID PKCE OAuth state machine; F1000 peer review gate (SHA3 fingerprint, ≥2 approvals); ArxivAssist ScholarlyAdapter (staging-based); OSF v2 adapter; Crossref deposit adapter + XML builder; Zenodo new_version(); OpenReview revision flow. (2026-05-09)

• arXiv API write adapter; OSF write adapter; Crossref deposit adapter; ORCID OAuth (PKCE); Zenodo versioning; OpenReview revision flow.
• F1000-style publish-then-review track; gate “indexed in our curated track” on ≥2 approving signed reviews.
• Venue catalog (contracts/scientia/venue-catalog.v1.yaml): IMC/MLSys/TMLR/JMLR/JAIR primary; Distill-style web-native; Living Reviews for the Atlas.
• Journal-fit recommender.

Phase 9 — First Provider Atlas — co-authored, IMC-targeted (4 wk + ongoing) ✅ COMPLETE (2026-05-09)

Status: Complete — AtlasManifest builder (T4 manifest with finding list, negative-result count, Zenodo/arXiv/OSF identifiers); MeasurementCampaign (Inspect bridge config + prereg check); AcademicCoauthor (ORCID + CRediT taxonomy roles → RO-Crate Person entity); AtlasSubmissionGate (prereg required + reply window cleared + negative-result quota). (2026-05-09)

• Onboard one academic co-author.
• Pre-register on OSF; run measurements through Inspect-Evals; contribute custom evals upstream.
• 14-day right-of-reply.
• Submit to IMC ‘27 or MLSys ‘27; deposit to arXiv + Zenodo.
• Living-review v1 published quarterly thereafter.

Deliverable: first peer-reviewed publication. Proof-of-thesis.

Phase 10 — Negative-result mandate + governance (1 wk; ongoing) ✅ COMPLETE (2026-05-09)

Status: Complete — NegativeResultEnforcer (refuses Atlas release when ≥3 positive findings with 0 null-results); CostDashboard ($/finding, $/extraction, $/atlas published in Atlas JSON); CoiDeclaration (ICMJE-format JSON COI disclosure); CopeRetractionNotice (COPE-aligned retraction workflow with Crossref notification state). (2026-05-09)

• System refuses to release the quarterly Atlas if ≥3 published findings exist with no null-result publication in the same window.
• Cost dashboard panel: $/finding, $/extraction, $/atlas; published in the Atlas itself.
• COI declaration (ICMJE-format JSON); CRediT taxonomy roles per author.
• COPE-aligned retraction workflow.

8. Sequencing

Phase 0 (foundations) ─┬─> Phase 1 (extractor) ─┬─> Phase 4 (artifact spec)
                       │                         │
                       └─> Phase 2 (prereg) ──> Phase 3 (right-of-reply) ─┐
                                                                           ├─> Phase 9 (first paper)
                       Phase 5 (Inspect+novelty) ─> Phase 6 (atlas) ──────┤
                       Phase 7 (format adapt) ───> Phase 8 (scholarly) ───┘

                       Phase 10 (negative-result mandate) — operational, after Phase 9

Critical-path: Phase 0 → 1 → 2 → 3 → 5 → 6 → 9. ~17 weeks. Phases 4, 7, 8 parallel-track once Phase 1 done.

9. Vendor vs. build

Vendor (non-negotiable):

• UK AISI Inspect + Inspect-Evals — rebuilding eval substrate is the single most expensive mistake we could make for venue credibility.
• MiniCheck-FT5 (770M T5) — production-grade cheap verifier.
• SPECTER2 + adapters — hand-rolling scientific embeddings is years of work AI2 has done.
• nanopub-java (or thin Rust port).
• RO-Crate context + spec.

Build (we have differentiated value):

• The extractor pipeline composition (SciClaims-style single-model orchestration with our XGrammar + Mens primitives).
• The pre-registration object as code.
• The right-of-reply window enforcement in code.
• The Provider Atlas itself.
• Symbolic-verifier strategy plugins for Vox-internal research.

10. Risks

#	Risk	Mitigation
R1	Leaderboard-Illusion-style methodology attack	Inspect substrate; academic co-author; pre-registered; right-of-reply baked in.
R2	Galactica-style PR backlash on hallucinated claims	VeriScore atomicity gate; span integrity hard-rejection; symbolic verifiers; MiniCheck for the rest; calibrated ABSTAIN default.
R3	”GPT-4 grades GPT-4” credibility loss	External symbolic verifier for every quantitative claim; for qualitative claims, MiniCheck (different model + much smaller).
R4	Publication-bias drift toward only-positive findings	Phase 10 negative-result quota in code, not policy.
R5	Adversarial provider response (legal/methodological)	14-day right-of-reply window; replies inline; living-review versioning.
R6	SciFact-Open generalization gap	Open-corpus eval split as Phase 1 acceptance gate; ABSTAIN-by-default below τ.
R7	Probe-suite contamination	Sealed rotating private probe set; publish only aggregates; rotate quarterly per Mesh §7 R4.
R8	Crate-boundary erosion	All cross-crate research events through L1 vox-research-events; vox-arch-check rules in Phase 0f.
R9	Cost runaway	Tier cascade enforced; daily budget cap; cost surfaced in the Atlas.
R10	Schema drift between contracts and Rust enums	Phase 0c codegen Rust from JSON Schema; vox-doc-pipeline regen list.
R11	Disclosure mistakes (CRediT/AI-tool omissions cause retraction)	AI-disclosure block auto-filled from extractor metadata; Methods-section block compulsory in template.
R12	Predatory journal accidental submission	Venue catalog whitelist only; system refuses unlisted venue.

11. Confirmation status

The user approved this plan in full on 2026-05-09 with the message “approved in full. Implement in full.” All seven §11 questions of the proposing message are answered yes:

1. Strategic thesis — endorsed.
2. Six new crates — approved.
3. UK AISI Inspect adoption — approved.
4. Negative-result quota default-on after Phase 9 — approved.
5. 14-day right-of-reply default — approved.
6. Plan committed at this path.
7. Academic co-author outreach — Phase 9 prerequisite.

12. Phase index (links to detailed plans)

Phase	Plan
0a	Phase 0a — Phantom-import resolution — Complete
0b	Phase 0b — vox-research-events L1 crate — Complete
0c	schema_types module in vox-research-events (5 JSON Schema → Rust types) — Complete
0d	7 DB tables + 12 Codex methods (research_pipeline.rs) — Complete
0e	SecretId: VoxOpenAlexEmail, VoxSemanticScholarApiKey, VoxOsfApiToken, VoxArxivAccessToken, VoxNanopubSigningKeyHex, VoxSwhidApiToken — Complete
0f	5 SCIENTIA crates registered in layers.toml (Phase 1–5 prep) — Complete
1–10	(TBD — phase plans land as predecessor phases complete)