SCIENTIA Self-Publication Gap Map (2026)

SCIENTIA Self-Publication Gap Map (2026)

Companion to: SCIENTIA Self-Publication Finalization Plan (2026). That plan’s Phases 0–10 are marked complete as of 2026-05-09. This document does not re-spec those phases — it inventories what is still missing for a developer on this codebase to go from git log to a published, citable artifact in a single coherent flow.

Per-gap implementation plans: See Phase index for outline plans A–H. Each plan is promoted from outline to detailed TDD-step fidelity when it becomes next-to-execute, matching the Finalization Plan’s §12 rhythm.

0. Orientation: what the Finalization Plan already provides

To prevent redesign, the following are out of scope for this gap map because they are built and complete per the Finalization Plan:

• Signal ladder T0→T4 with rigor gates (§2 of the plan).
• Claim extraction: VeriScore + atomic + span integrity + MiniCheck + calibrated ABSTAIN (crates/vox-claim-extractor/).
• Pre-registration as signed code object with deviation detection (crates/vox-prereg/).
• Nanopublication emission and RO-Crate 1.2 packaging (crates/vox-nanopub/, crates/vox-ro-crate/).
• Atomic-NEI novelty against time-bounded corpus, SPECTER2 retrieval, ChronoFact, EvidenceConflict family (crates/vox-inspect-bridge/).
• AISI Inspect adapter; symbolic verifiers wired into confidence_fusion.rs.
• 14-day right-of-reply window, retraction nanopubs, living-review version DOIs, COPE retraction workflow.
• Scholarly adapters: arXiv, OSF, Crossref deposit, Zenodo versioning, OpenReview revision, ORCID PKCE OAuth, F1000 publish-then-review.
• AtlasSubmissionGate with negative-result quota; venue catalog whitelist; predatory-journal refusal.
• AI-disclosure block auto-fill; CRediT + COI declarations.
• Format adaptation for short-form (Bluesky, schematic-only figures).

The Provider Atlas longitudinal AI-epidemiology paper is the plan’s terminal artifact — but it is one publication track, not the user-journey substrate.

1. The end-to-end user journey today (with gaps marked)

A developer working in this codebase wanting to self-publish takes this path. Steps marked (gap N) are not yet served by a single command or surface.

1. Develop normally. Commits land, CI runs benchmarks, telemetry flows, training jobs finish. ✅ (observed by existing infra)
2. “Is any of this publishable?” → (gap A) No vox scientia scout exists. Discovery scan operates on already-prepared manifests, not on raw repo/CI/telemetry state.
3. “What candidate did I just produce?” → (gap A continued, gap F) The finding-candidate.v1 ledger row schema exists; the producers that mine commit graph, benchmark CI history, MENS training runs, and Vox-internal telemetry to emit candidate rows are thin.
4. Pre-register the hypothesis ✅ (Phase 2 complete).
5. Run the experiment. ✅
6. “Did the result reproduce?” → (gap B) artifact_replayability is in the worthiness rubric as a [0,1] metric, but appears operator-asserted. No runner re-executes the RO-Crate in a clean sandbox and writes the measured score back.
7. Promote claims through the signal ladder. ✅
8. “Draft the paper.” → (gap C) Phase 7 covers short-form (Bluesky/schematic). No long-form IMRaD scaffolder that fills only the safe slots from RO-Crate + verified claims and leaves provenance-bound TODOs for the human narrative.
9. Get dual approval. ✅ (if you have a co-author). (gap D) Solo developer path requires either a recruited human or a designed audited-LLM-critic-as-second-approver flow under AI-author disclosure.
10. Decide the venue. ✅ (journal-fit recommender exists). (gap E) The recommender’s first-class output is the Provider Atlas track. The AI/SWE micro-publication track — algorithmic_improvement, reproducibility_infra, telemetry_trust, policy_governance — enumerates candidate classes but does not have a parallel publication-format + venue-routing track to match.
11. Submit. ✅
12. Right-of-reply, peer review, revisions, retraction. ✅
13. “Where does the published artifact live for readers?” → (gap G) Zenodo + arXiv host the deposit; vox-plugin-publication syndicates outward. A Vox-native landing surface (findings.vox-lang.org/<trusty-uri> or analogous in the docs site / dashboard) that is the publication’s canonical home is not yet in evidence.
14. “What’s in my queue?” → (gap H) No dashboard panel surfacing candidates awaiting prereg, claims awaiting verification, manifests in reply-window, retraction queue.

2. Gap inventory

Gap A — Self-observation candidate producers

What. Detectors that turn the user’s own repo/CI/telemetry/training-run activity into finding-candidate.v1 rows.

Evidence. contracts/scientia/finding-candidate.v1.schema.json defines the candidate type with classes algorithmic_improvement, reproducibility_infra, policy_governance, telemetry_trust. The Finalization Plan’s Phase 6 wires ProviderObservation / ModelCapabilityEvidence families (about external providers Vox calls) into the candidate flow. The orthogonal flow — Vox observes itself — has the schemas but missing producers. No grep hit for a commit-graph miner, benchmark-CI-delta detector, or MENS-training novelty detector that populates the ledger.

Why it matters. Without these, the candidate ledger is fed only by the Provider Atlas path. A developer making a real software-engineering or algorithmic contribution in this codebase has no automatic path from their work into a publication candidate.

Concretely missing.

• Commit-graph miner: perf-improving merges (use existing ExecTimeRecord), novel-symbol detectors over recent diffs, MDL-style compression gain over unfamiliar regions of the AST.
• Benchmark-CI history detector: regression and improvement spikes over a preregistered baseline.
• MENS-training-run hook (per the in-flight Mn-T1..T15 plan): emit candidates from new SOTA checkpoints, novel attention patterns, repeatable training regressions.
• Vox-internal telemetry detector for reproducibility_infra and policy_governance candidates (Socrates surface is collected but not routed to the ledger as candidates about Vox itself).

Dependencies. None upstream; reuses existing vox-research-events emitter and the candidate schema.

Effort. Medium. One producer per source; each ~400–800 LoC plus detector heuristics; producers compose, so it can ship incrementally.

Gap B — Replay runner that measures artifact_replayability

What. A runner that, given a publication_manifest and its RO-Crate, re-executes the experiment in a sandboxed environment and writes the measured replayability score back to the worthiness signals.

Evidence. artifact_replayability appears in crates/vox-publisher/src/publication_worthiness.rs, worthiness_extraction.rs, and the rubric (docs/src/reference/scientia-publication-worthiness-rules.md), but the measurement source is absent. No vox-replay-runner crate; no replay_artifact API surface; no CI hook that re-executes RO-Crate deposits.

Why it matters. The Finalization Plan takes credibility-via-symbolic- verification very seriously (the AlphaEvolve thesis, §3.4). An asserted replayability number sits in tension with that posture. A measured score — even a coarse one (“the manifest’s entry_point ran to a green exit code and produced an output hash matching the deposited one”) — converts a soft rubric field into a verifier-backed claim.

Dependencies. Needs RO-Crate entry_point convention (Phase 4) — ✅. Sandbox isolation: reuse existing CI runner, or a thin vox-sandbox-runner shim.

Effort. Medium-small. Sandbox + replay harness; metric writeback; worthiness rubric update.

Gap C — Long-form manuscript scaffolder

What. A constrained-grammar emitter that produces an IMRaD markdown skeleton (Intro / Methods / Results / Discussion) from a FindingCandidate

• its verified claims + its RO-Crate, filling only the safe slots and leaving provenance-bound TODOs for the human narrative.

Evidence. Phase 7 covers short-form adaptation via crates/vox-research-events/src/publication_format.rs (ShortFormVariant, adapt_claim_to_platform). No long-form analogue; no grep hit for manuscript, IMRaD, or paper-scaffolding in non-archived code.

Why it matters. The rubric explicitly forbids auto-generating novelty/significance assertions, causal mechanism claims, and measurement-implying figures. Fine. But the blank-page barrier for a solo developer is real, and the safe slots (methods derived from RO-Crate declarations; results table lifted from verified claim envelopes; citation stubs from the SPECTER2-verified prior-art set; AI-disclosure block from AiDisclosureBlock::build) are all already provenance-bound. Filling them mechanically does not violate the red lines.

Concretely. A vox scientia manuscript-draft --publication-id <id> command emitting <id>.imrad.md with:

• Methods section auto-filled from RO-Crate mainEntity + prereg eval_substrate.
• Results table from claim envelopes (one row per verified atomic claim, with Trusty URI link).
• References block from the SPECTER2-retrieved prior-art set used for novelty.
• Acknowledgments, ORCID/ROR, AI-disclosure auto-filled.
• Discussion / Intro / Significance sections present as <!-- TODO(narrative): --> blocks the human writes by hand.

Dependencies. Reuses RO-Crate (Phase 4), claim extractor (Phase 1), prereg (Phase 2), vox-constrained-gen.

Effort. Medium. Templating + grammar; the verified-only constraint is the discipline.

Gap D — Solo-author critic-gate path

What. A designed path for a single developer to clear the dual-distinct- approver gate without requiring a second human, under explicit AI-author disclosure.

Evidence. Dual-approver is hard-required at the digest level in crates/vox-db/src/store/ops_publication.rs and surfaced in the playbook. machine-suggestion-block.schema.json carries machine_suggested/requires_human_review flags. AI-disclosure block exists (vox-ro-crate/src/ai_disclosure.rs). Phase 9 of the Finalization Plan onboards an academic co-author — solo is not the primary mode.

Why it matters. A real Vox self-publication user is often solo. The plan’s posture (Galactica risk, GPT-4-grades-GPT-4 critique) requires that any non-human approver be (a) explicitly disclosed, (b) a different substrate than what produced the artifact, (c) bound to the content digest, and (d) auditable.

Concretely. A second-approver role of AuditedLLMCritic that:

• Submits a signed approval bound to content_sha3_256 from a key registered under an ORCID-distinct identity.
• Is automatically disclosed in AiDisclosureBlock as a contributor with the CRediT role Validation only (never Investigation or Writing — original draft).
• Uses a model architecturally different from any used in the artifact pipeline (enforced by recording model fingerprints in both manifest and critic approval).
• Refuses to approve if the artifact’s worthiness signals include any contribution from the same model family — closes the GPT-4-grades-GPT-4 hole.
• The published artifact’s venue catalog must allow LLM-critic approval (a new venue-catalog flag); IMC/MLSys/TMLR would set this false, F1000-track and Zenodo-only deposits could set it true.

Dependencies. Venue catalog flag; ORCID-distinct critic identity; critic model fingerprinting.

Effort. Small-medium. Mostly policy plumbing plus the critic identity machinery.

Gap E — AI/SWE micro-publication track (non-Atlas)

What. A publication track for individual algorithmic_improvement / reproducibility_infra / telemetry_trust / policy_governance findings that does not route through the quarterly Provider Atlas.

Evidence. FindingCandidateClass enumerates these classes (crates/vox-research-events/src/schema_types.rs); the candidate ledger accepts them. venue-catalog.v1.yaml has IMC/MLSys/TMLR plus Distill-style and Living Reviews entries — but the journal-fit recommender is geared toward longitudinal measurement papers.

Why it matters. The user’s question is broader than provider measurement — it includes “I built a thing in this codebase and want to publish about it.” Examples already in this repo: vox-prereg itself is publishable (PL/SE methods venue); vox-arch-check is publishable (software-architecture venue); the SSOT migration policy (mesh-and-language-distribution-ssot-2026.md §5.5) is publishable (empirical software engineering). None of these fit the Atlas mold.

Concretely.

• Per-class venue mappings: SWE-PL improvements → ICSE/FSE/OOPSLA/PLDI; reproducibility infra → REP/MSR/ICSE-SEIP; telemetry trust → MLSys/SOSP workshops; policy governance → AIES/FAccT.
• Per-class artifact-format profile: short systems papers ≠ Atlas measurement papers; figure norms differ; reproducibility-package expectations differ.
• Per-class default reply-window length (14 days is right for Atlas; open-review tracks differ).

Dependencies. Gap C makes this much more usable (each class implies a different IMRaD weight). Otherwise none.

Effort. Small (config + venue catalog rows) + medium (per-class artifact-format profiles).

Gap F — vox scientia scout single-command surface

What. One command that, run in a Vox workspace, surveys recent commits / CI history / telemetry / training-run outputs and prints a ranked candidate list with proposed candidate_class, suggested venue, and a “promote-to-ledger” affordance.

Evidence. Discovery surface today is publication-discovery-scan/explain/refresh-evidence — all over already-prepared manifests. No “scan repo state from scratch” command.

Why it matters. This is the front door for the user journey. Without it, the developer has to know to prepare a manifest before discovery ever runs. With it, the system surfaces opportunity proactively.

Concretely.

• Subcommand under vox scientia.
• Aggregates output from Gap A’s producers.
• Prints a table: candidate-id, class, top signals, suggested venue, recommended next command (publication-prepare, prereg-draft, etc.).
• Optional --watch mode for daemon-style monitoring; emits OS notifications when a strong-strength candidate appears.

Dependencies. Gap A is the prerequisite (no producers, nothing to report).

Effort. Small (CLI + table rendering); leverages Gap A.

Gap G — Vox-native publication reading surface

What. A canonical landing page for each published manifest, hosted on Vox-owned infrastructure, distinct from Zenodo/arXiv deposits.

Evidence. vox-plugin-publication does RSS/Atom + Reddit/YouTube syndication. docs/src/index.mdx is the docs landing. No /findings/ or /publications/ route surfaces canonical published artifacts. Phase 4 notes Highwire-style meta tags for Google Scholar pickup but doesn’t specify the host page.

Why it matters. External deposits (Zenodo, arXiv) own the DOI but not the reading experience. Distill-style web-native artifacts — first-class in the Finalization Plan’s venue catalog — require an HTML host. Living Reviews need a stable canonical URL pointing at “latest version” with the version history table inline.

Concretely.

• A findings/<trusty-uri> route on the docs SSG or dashboard, rendering the RO-Crate’s text body + claim table + version history + reply thread inline.
• Highwire meta tags for Scholar pickup.
• Embedded nanopub viewer for atomic claims.
• Reply window status badge.

Dependencies. RO-Crate (Phase 4) — ✅.

Effort. Medium. SSG route + renderer; relies on existing docs infrastructure.

Gap H — Discovery dashboard panel

What. A dashboard panel surfacing the publication-pipeline queue: candidates by class, claims awaiting verification, manifests in reply-window, retraction queue, cost dashboard rollup.

Evidence. Phase 6 closed the loop end-to-end but is backend. The dashboard plan (docs/superpowers/plans/ci/2026-05-03-vox-dashboard-claude-design-port.md) exists but I did not verify a Scientia panel.

Why it matters. Operator visibility into the pipeline is currently CLI JSON. A panel converts the workflow into a glanceable surface and surfaces stalls (e.g., a candidate stuck in evidence-incomplete for 30 days).

Dependencies. Existing tables (publication_manifests, publication_status_events, external_submission_jobs, finding_candidates ledger).

Effort. Small-medium. Mostly UI work over existing data.

3. Priority and dependency matrix

Gap	Severity	Effort	Unlock value	Depends on
A — self-observation producers	high	medium	unlocks F, gates E credibility	none
B — replay runner	high	medium-small	closes credibility hole in rubric	RO-Crate ✅
C — IMRaD scaffolder	high	medium	removes blank-page barrier	claim extractor ✅, RO-Crate ✅, prereg ✅
D — solo critic-gate	high (for solo users)	small-medium	unlocks solo self-publication	AI-disclosure ✅, venue catalog ✅
E — AI/SWE micro-track	medium	small + medium	broadens publication scope beyond Atlas	C
F — vox scientia scout	medium	small	front-door UX	A
G — reading surface	medium	medium	hosts Distill-style + Living-Review	RO-Crate ✅
H — dashboard panel	low	small-medium	operator UX	none

Dependency chains

A ──> F                     (scout needs producers)
A ──> E (credibility)       (AI/SWE micropapers need own-repo candidates)
C ──> E (usability)         (per-class scaffolds make track usable)
B ──> rubric credibility    (measured replayability)
D ──> solo workflow         (independent of others)
G, H ──> independent UX wins

The high-severity cluster is A + B + C + D. A and B are independent and can ship in parallel; C and D are independent and can ship in parallel; F is a thin wrapper over A once A exists; E is a config + per-class extension of C; G and H are UX wins that do not block any other gap.

4. Recommended first slice

If this gap-map gets approved, the highest-leverage minimal first slice is:

Slice 1 — “From commits to a candidate row, with a measured replay.”

Scope:

• Gap A, narrowest cut: one producer for algorithmic_improvement candidates from commit-graph + benchmark CI history (reuse ExecTimeRecord).
• Gap B, narrowest cut: replay runner that re-executes the RO-Crate mainEntity entry point in a fresh worktree and writes the measured replayability score (binary pass/fail + output-hash match) back to worthiness signals.
• Gap F, thin wrapper: vox scientia scout listing candidates from this one producer.

Why this slice: it makes the front of the user journey (steps 1–7 above) work end-to-end for the most common case (a developer made a perf improvement), with measured (not asserted) replayability. Everything downstream of step 7 is already strong.

Slice 2 candidates: Gap C (manuscript scaffolder) or Gap D (solo critic gate) — independent, pick by user demand.

5. Open questions

1. Producer pluggability. Should signal producers be a plugin contract (so third parties can add detectors) or an L1 trait with a fixed set?
2. Solo critic identity provenance. Is an ORCID-distinct identity for an LLM critic ethically acceptable in the venues we care about? (Atlas: probably not. F1000 / Zenodo-deposit: probably yes. Needs per-venue declaration in venue-catalog.v1.yaml.)
3. Replay determinism. What’s the minimal RO-Crate mainEntity contract to make replay tractable? GPU-bound experiments need a different cut than CPU-bound.
4. Reading surface hosting. Docs SSG or dashboard? Or both with the dashboard as edit surface and SSG as canonical reader?
5. Negative-result quota for the AI/SWE micro-track. The Atlas has a 3-positive : 1-null quota. Does the SWE micro-track inherit this or set its own?

6. Cross-references

• SCIENTIA Self-Publication Finalization Plan (2026) — supersedes this for Phases 0–10.
• SCIENTIA SSOT handbook — lifecycle and status vocabulary.
• Publication worthiness rules — rubric and red lines.
• How-to: Publish Scientia findings — operator flow.
• Where Things Live — concept → crate map.
• Mesh & Language Distribution SSOT (2026) — §3.5 Hopper track, §5.5 migration policy, §5.6 routes.