Research platform

Platform Evidence Research — Orchestrator

Pipeline skill — turns owned-account analytics (engagement, follower growth, per-post performance) into a sourced per-platform evidence base that social, SEO, short-form, and evaluation skills consume to ground recommendations in measured reality instead of intuition. Operates on the operator's own accounts (not market trends).

Core Question: "What does our own platform evidence actually say — and which recommendations does it support?"

Why this skill exists, evidence-not-intuition doctrine, credential-free posture, distinction from research-shortform, when NOT to use: references/playbook.md [PLAYBOOK].

---

Critical Gates — Read First

Non-negotiable before dispatching any agent (5 evidence-source types + thresholds: references/evidence-protocol.md [PROCEDURE]):

1. No invented metrics. Every number traces to owned_analytics / public_metrics / manual_export / forum_observation / prior_eval with a measured_at date. Unsourced = fabricated.
2. Source-type labeled per datum. A public-page view count is not owned analytics; an estimate is not a measurement. Never present a guess or benchmark as instrumented owned data.
3. Evidence availability is platform-specific — declare it honestly. Each per-platform section declares RICH / MODERATE / CONSTRAINED. TikTok and Instagram default CONSTRAINED — never invent retention-curve or demographic depth a platform doesn't expose.
4. Missing evidence is a gap, never a fabrication. Platform with no evidence → NO_EVIDENCE flag + "what to export" note. All platforms empty → NEEDS_CONTEXT.
5. Metrics decay — two freshness windows. metrics_window_date (30d refresh / 60d warn) governs performance numbers; algorithm_context_date (90d refresh / 180d warn) governs platform-mechanic context. Stale evidence is flagged, never silently aged.

Quality Gate

Critic verifies before delivery (all 5 PASS required, max 2 rewrite cycles; full rubrics in references/scoring-rubrics.md):

• Every metric traces to a tagged evidence source with measured_at
• Every datum labeled owned_analytics / public_metrics / manual_export / forum_observation / prior_eval — no blank, no public/benchmark mislabeled as owned
• Every per-platform section declares MEASURED / PARTIAL / NO_EVIDENCE matching its actual evidence count + source mix
• Every recommendation names platform, evidence source, freshness window, confidence — generic fails
• NO_EVIDENCE platforms carry no recommendations — only a "what to export" gap note

---

Before Starting

Per references/_shared/before-starting-check.md [PLAYBOOK]:

0. Mode resolution per references/_shared/mode-resolver.md. --fast collapses to single-pass intake + synthesis with critic skipped — but the 5 Critical Gates STILL enforce (safety supersedes --fast). Cold Start still fires when platform scope or account ownership is unresolved. Session execution profile (single-vs-multi): inherit per references/_shared/execution-policy.md.
1. Read implementation-roadmap/canonical-paths.md if present — verify output path.
2. Read .forsvn/index/manifest.json — check for prior platform-evidence artifact (warm-start) and for .forsvn/loops/*/evals/ artifacts (a prior_eval source the operator may not have mentioned).
3. Run Pre-Dispatch per references/procedures/pre-dispatch.md [PROCEDURE] — needed dimensions, evidence-intake prompt, warm/cold start, write-back map.

Artifact	Source	Required?
.forsvn/loops/*/evals/*.md	run-pipeline / evaluate-*	Optional — becomes a prior_eval evidence source
research/icp-research.md	research-icp	Optional — audience grounding for reading segments

Artifact Contract

• Path: docs/forsvn/artifacts/research-research-platform-<YYYY-MM-DD>-<slug>.md (flat v2 grammar; one artifact per account scope + platform set)
• Lifecycle: pipeline (regenerable, point-in-time; refresh governed by the two freshness windows). Evidence is re-measured on re-run, never amended in place.
• Frontmatter fields: skill, type, status, date, stack (=research), review_surface (=md), account_scope, platforms_analyzed, metrics_window_date, algorithm_context_date, evidence_sources_logged, coverage_per_platform. Full schema: references/format-conventions.md [PROCEDURE]. v2 baseline: references/_shared/artifact-contract-template.md.
• Required body sections (cross-stack contract, in order): TL;DR · Evidence Base · Per-Platform Evidence · Cross-Platform Comparison · Recommendations · Missing Evidence & How to Close It · Open Risks & Caveats · What This Evidence Doesn't Cover
• Side effects (on PASS or done_with_concerns): write artifact → run bun scripts/manifest-sync.ts → experience write-back (account scope + platforms → docs/forsvn/experience/content.md).
• Consumed by: write-social, optimize-seo, research-shortform, evaluate-content, evaluate-shortform, publish-social — each reads Per-Platform Evidence + Recommendations to ground decisions in measured performance.
• Cross-stack OUTPUT contract: frontmatter schema + 8 body sections + per-datum source-type tags + MEASURED / PARTIAL / NO_EVIDENCE flag + recommendation 4-part attribution (platform / source / window / confidence) are load-bearing — schema changes require atomic update of consumers.

---

Agent Manifest

5 agents across 2 layers (evidence-intake × N + benchmark in parallel; synthesis → recommendation → critic sequential). Full table + dispatch sequence + --fast behavior: references/agent-manifest.md [PROCEDURE]. Mechanics (spawn, critic routing, single-agent fallback, chain position, skill deference): references/procedures/dispatch-mechanics.md [PROCEDURE].

Single route — full Layer 1 + Layer 2 sequence runs every time. Evidence depth varies by what the operator supplies, not by route choice.

---

Anti-Patterns

references/anti-patterns.md [ANTI-PATTERN] — fabricated metrics, source-type laundering, fake precision on CONSTRAINED platforms, coverage-flag inflation, generic recommendations, stale-evidence masking, critic-loop overrun, cross-stack contract drift. Re-read before any output ships.

Durable Rules (protected)

<!-- SLOW_UPDATE_START --> <!-- No pinned rules yet. Populate via the slow-update workflow (see references/slow-update-fence.md). Each pinned rule must (a) be procedural not instance-specific, (b) be earned from a regression or critic-flagged failure, (c) cite the artifact / decision record that justified pinning. --> <!-- SLOW_UPDATE_END -->

Completion Status

Skill returns one of:

• DONE — all 5 critic rubrics PASS within ≤2 cycles; every in-scope platform carries a coverage flag (MEASURED, PARTIAL, or honest NO_EVIDENCE).
• DONE_WITH_CONCERNS — critic loop cap reached; remaining failures are surfaceable as warnings (e.g., one platform PARTIAL, one source past freshness). Concerns pinned at the top of the artifact.
• BLOCKED — WebSearch / WebFetch blocked when a benchmark pull was required; or supplied exports unreadable. State what's needed.
• NEEDS_CONTEXT — no evidence supplied for any platform and none retrievable; recommend what to export and which evaluate-* loop could feed prior_eval data.

Next Step

After delivery: feed Per-Platform Evidence + Recommendations into write-social / optimize-seo / research-shortform / evaluate-content / evaluate-shortform / publish-social. Refresh when metrics_window_date > 30d or algorithm_context_date > 90d.

---

Worked Example

Full run — 3-platform evidence base with mixed coverage flags traced through the critic gate: references/examples/platform-evidence-walkthrough.md [EXAMPLE].

References

• references/playbook.md, evidence-protocol.md, format-conventions.md, scoring-rubrics.md, anti-patterns.md
• references/agent-manifest.md, references/procedures/{pre-dispatch, dispatch-mechanics}.md
• references/_shared/{before-starting-check, mode-resolver, pre-dispatch-protocol, confidence-labeling}.md
• references/platforms/ — per-platform evidence schemas (x, linkedin, tiktok, youtube, instagram): availability tier, exposed metrics, definitions
• 5 sub-agents in agents/; critic-agent.md holds the canonical 5-rubric gate
• research-skills/CLAUDE.md — stack conventions (Pre-Dispatch, Complexity Routing, Multi-Agent)