Skip to content
Vox: The AI-Native Programming Language
Internals — research note: investigation or findings, not a description of shipped behaviour.
Unverified: This page has not yet been anchored to automated code tests.

Deep Research Stage 2 Synthesis & Priorities (2026-08-01)

Deep Research Stage 2 Synthesis & Priorities

Section titled “Deep Research Stage 2 Synthesis & Priorities”

Date: 2026-08-01 Inputs: deep-research-cross-domain-methods-survey-2026-08-01.md, deep-research-domain-agnosticism-audit-2026-08-01.md, deep-research-gui-representation-design-2026-08-01.md, deep-research-implementation-divergence-audit-2026-08-01.md, plus the graphify structural map of vox-scientia/vox-research-shim/vox-search/vox-research-events/vox-publisher (6216 nodes, 12457 edges, 536 communities). Builds on deep-research-synthesis-and-priorities-2026-08-01.md (Stage 1).

Where Stage 1 landed vs. what Stage 2 found

Section titled “Where Stage 1 landed vs. what Stage 2 found”

Stage 1 shipped 13 tasks across two plans (trust/novelty core, multi-provider routing) plus 10 code-review fixes — genuinely substantial, verified working. Stage 2’s job was to test that work against reality on two axes: does it generalize beyond code/academic queries, and did it actually deliver everything the docs claimed. The answer to both is “mostly, with specific, fixable gaps” — not a redo, a continuation.

Prioritized gap list

Section titled “Prioritized gap list”

P0 — Fix now, near-zero risk

Section titled “P0 — Fix now, near-zero risk”
  1. web_dispatcher.rs’s source-authority ranking bias. github.com/docs.rs are boosted (1.15x) alongside arXiv/PubMed/doi.org, while Wikipedia/Reuters/BBC/AP get no boost at all (1.0x, same as a random blog). This is a live, measurable bias today — not latent. Fix: extend the authority tier to include major general-reference and wire-service domains at a comparable weight. (Domain-agnosticism audit §5)
  2. Remove ANTI_LAZINESS_RIDER from research judge/synthesis prompts. Verbatim coding-agent boilerplate (“stubs,” “TODO blocks,” “implement ALL requested logic”) injected into the research answer synthesis and quality-judging prompts — nonsensical noise for “summarize the causes of the French Revolution,” and a plausible source of code-like completeness framing bleeding into prose synthesis. (Domain-agnosticism audit §6)
  3. Give the post-hoc citation audit pass its own numbered implementation task. Stage 1’s design spec named it (competitive-landscape doc’s #1 differentiation finding), but it was prose-only inside a plan document, not a numbered task — and prose-only items are exactly what falls off an execution list under time pressure, per the divergence audit’s own pattern finding. This item is P0 not because it’s cheap, but because failing to give it a task number is what caused it to be skipped once already.

P1 — High impact, directly closes Stage 1’s real gaps

Section titled “P1 — High impact, directly closes Stage 1’s real gaps”
  1. Independent-source corroboration counting as the universal trust fallback. The highest-leverage recommendation from the cross-domain methods survey: needs no external API, and is the single technique that makes trust scoring work for history, journalism, current events, and general research — exactly the domains TrustScorer currently can’t reach. Feeds directly into the GUI’s new “Confirmed by N independent sources” citation indicator.
  2. Domain-gate TrustScorer’s OpenAlex call. Skip the per-hit title search for hits whose domain is clearly non-academic — closes the wasteful-and-fragile issue found in the domain-agnosticism audit (§1) and is a prerequisite for corroboration-counting to be the primary signal for non-academic hits rather than a fallback competing with a noisy academic lookup.
  3. Wire ClaimVerdict.self_consistency into the confidence gate. Computed and stored since Stage 1, never consumed — genuinely live gap, not stale documentation, confirmed by direct code read. Also: rename or clearly disambiguate from orchestrator_policy.rs’s pre-existing, unrelated self_consistency field before wiring, to avoid the naming collision the divergence audit flagged.
  4. Bring ResearchView up to the GUI standard already set by ClaimsView/NoveltyEvidencePanel. Headline verdict banner + expandable claim accordion (validated design, see the GUI representation doc), reusing VerdictBadge and the signal-grid pattern rather than inventing new components. This is the single most user-visible item in this whole synthesis — today’s ResearchView shows literally nothing beyond raw markdown.
  5. Complete NoveltyEvidenceBundle population. Confirmed still empty despite the lexical-dedup work landing — the schema this item was meant to feed was never actually fed. Once populated, the GUI work is already solved (NoveltyEvidencePanel’s existing signal grid is directly reusable, per the GUI doc).

P2 — Real value, moderate effort

Section titled “P2 — Real value, moderate effort”
  1. Evidence-tier tagging per claim, decoupled from source-tier (cross-domain methods survey technique 3, generalizing the historical primary/secondary/tertiary distinction too) — an LLM-classifiable field, no new API needed.
  2. CourtListener (legal) and ClinicalTrials.gov/openFDA (medical) validity/status checks, wired alongside Crossref using the same “check if this record has been superseded” pattern, both free/no-auth APIs.
  3. Domain/outlet-level reputation scoring for non-academic sources (E-E-A-T-style LLM rubric) as the non-academic analog to OpenAlex venue reputation.
  4. “Disputed narrative” report framing for genuinely contested/corroboration-thin topics — render competing credible positions with their own citation sets instead of forcing single-answer synthesis, surfaced in the GUI as “Contested — N credible perspectives” on the headline banner.
  5. Wire WorthinessSignalsV2 into the real SCIENTIA finding-promotion pipeline (confirmed still zero callers) — but only after item 4’s corroboration-counting and item 11’s domain-reputation scoring land, so non-academic sources don’t all get force-bucketed into WorthinessProfile::Social regardless of actual quality, per the domain-agnosticism audit’s finding.

P3 — Lower urgency, still real

Section titled “P3 — Lower urgency, still real”
  1. Recency-decay weighting keyed to topic volatility (cross-domain methods survey technique 11).
  2. Fact-check-org verdict lookup for checkable political/viral claims (technique 8) — no free unified API, needs targeted WebFetch, keep scoped narrow.
  3. All ten Stage-1 P2-P4 carryovers confirmed still untouched: cross-encoder reranking, synthesis_max_tokens bump (still hardcoded 1200), semantic result caching, durable async jobs, per-hop observability, Reflexion-style self-revision, an internal eval harness, skills packaging, GUI settings-search indexing for API keys.

The meta-finding: how Stage 2’s plan must be written differently

Section titled “The meta-finding: how Stage 2’s plan must be written differently”

The implementation-divergence audit’s most important finding isn’t any single dropped feature — it’s the pattern: nearly every Stage 1 code-review fix was “the logic was applied to only one of N equivalent call sites” or “a cross-cutting invariant wasn’t audited across sibling paths,” not wrong algorithms. This directly shapes how Stage 2’s implementation plan must differ from Stage 1’s:

  • Every “wire X into Y” task must include a grep-verified inventory of every call/construction site X needs to reach, established while writing the plan — not discovered during code review.
  • Cross-cutting invariants get their own audit task (e.g. “does the corroboration-counting fallback apply everywhere TrustScorer is called, not just the dominant site”) rather than being scoped implicitly inside one feature task.
  • Every design-spec item gets a numbered task, even ones that “sound like one function” — P0-3’s citation-audit near-miss is the concrete proof this matters.
Section titled “Recommended next step”

This synthesis is the complete input for Stage 2’s implementation plan. Per the brainstorming skill’s terminal-state rule, translate items #1-#8 (P0/P1) into a concrete plan via the writing-plans skill next — applying the meta-finding above as a hard constraint on how each task is scoped, not just what it covers.