PART 04 — PHASE 2

The Scout — evidence before inference.

Build a review assistant from permitted public or licensed sources. It records what is observable, preserves provenance, and proposes hypotheses for a human to test. It cannot see competitor spend, profitability, conversion, or creative fatigue.

45–60 Min Model-agnostic Read-only research

Reviewed 31 Aug 2026. Educational architecture, not legal advice or a promise of performance. Platform access, terms, interfaces, regional availability, and rate limits change. Re-check the current source terms before collection or reuse. Every sample finding below is hypothetical.

Agent 01 · Market research

The Scout

A read-only research workflow that turns approved observations into traceable creative questions. It never launches ads, impersonates users, bypasses access controls, or presents inferred outcomes as facts.

Access · permitted only Output · hypothesis Approval · human research lead Provenance · required

Start with a collection policy

Source only what you may use
  • Results viewed through the official Meta Ad Library experience
  • Results viewed through TikTok Creative Center where available
  • Brand-owned, licensed, or manually approved research material
  • Public storefront observations collected within applicable terms
  • Your own first-party campaign evidence in a separate dataset
Respect boundaries
  • Private accounts, gated pages, personal data, or access tokens
  • Content obtained by bypassing authentication or anti-bot controls
  • Unlicensed creative files for reuse or model training
  • Guessed spend, revenue, targeting, conversion, or profitability
  • Claims that ad longevity proves success or audience fatigue
Source
Permitted workflow
Record with every observation
Use the official interface or an explicitly authorized integration
URL, market, query, capture time, visible fields
Use the official experience and its current permitted features
URL, filters, region, capture time, visible fields
Licensed research
Use only within the licence and approved purpose
Provider, licence, fields, retention, owner
Public storefront
Prefer manual sampling; legal and security owners approve automation
URL, robots/terms review, rate limit, capture time

A defensible Scout cycle

1 · Authorize
Approve the source, purpose, fields, and cadence
Name an owner. Document terms review, licence, retention, geography, robots guidance where relevant, and a request budget. “Public” does not automatically mean unrestricted collection or reuse.
2 · Capture
Save observable facts with provenance
Record source URL, filters, market, timestamps, visible text, asset identifiers, and collector version. Apply rate limits, exponential backoff, caching, and a stop condition when the source blocks or changes.
3 · Separate
Keep observation, inference, and recommendation distinct
observation: three sampled ads mention gifting. inference: gifting may be receiving more attention. hypothesis: a brand-fit gifting angle may be worth a controlled test.
4 · Review
Check rights, culture, claims, and evidence
A native-market reviewer checks context and dialect. A rights reviewer rejects imitation. A marketer checks whether the sample can support the stated conclusion.
5 · Test
Turn an approved hypothesis into an owned experiment
Creative and performance are evaluated on your own traffic against a predeclared baseline. The Scout receives the result only after attribution lag and data-quality checks.

Output contract

Return JSON only: { "observations": [{"text": "...", "source_url": "...", "captured_at": "..."}], "limitations": ["sample is incomplete", "competitor outcomes are not observable"], "hypothesis": "A brand-fit angle may merit testing", "alternative_explanations": ["seasonality", "selection bias", "regional filter"], "rights_risk": "low | review | reject", "cultural_review_required": true, "recommended_test": {"baseline": "...", "primary_metric": "...", "stop_rule": "..."} } Rules: - Do not estimate competitor spend share, ROAS, sales, targeting, or fatigue. - Do not say an angle is a gap unless the bounded sample and filters are stated. - Do not reproduce a competitor's distinctive copy, visual identity, or protected asset. - Quote only the minimum needed for research and preserve the source link. - If access, provenance, licence, or evidence is missing, return "insufficient evidence".

Observation: “In a manually reviewed sample of 18 KSA ads returned by two saved queries on 30 Aug, four visibly referenced gifting.”

Limitation: “The source sample is incomplete and does not reveal spend, conversion, profit, audience saturation, or why an ad is visible.”

Hypothesis: “A rights-cleared gifting concept may be worth testing with this store’s audience.”

Decision: A human may approve a small, pre-budgeted test. This example does not predict the result.

Evaluation: score research quality, not confidence theatre

Evidence
Provenance coverage
Can every factual statement be traced to an authorized source, query, market, and capture time?
Calibration
Inference discipline
Does the output label uncertainty and avoid inventing outcomes that the source cannot expose?
Rights
Non-imitation review
Do reviewers reject copied assets, distinctive expression, unsafe claims, or unlicensed reuse?
Coverage
Representative sampling
Test languages, markets, categories, seasonal periods, empty results, duplicates, and source-interface changes.
Operations
Terms and rate-limit safety
Verify allowlisted sources, backoff, retention, deletion, monitoring, and an immediate collector kill switch.
Learning
Owned experiment quality
Judge accepted hypotheses using your declared baseline, sufficient sample, attribution window, and unit economics.

Source changed or blocked? Stop collection and require a fresh review. Do not rotate identities, defeat controls, or improvise an unofficial endpoint.

Missing provenance or unclear licence? Quarantine the record. Do not pass it into generation or training.

Religious, medical, financial, or regulated claim? Escalate to qualified reviewers. Seasonal proximity is not permission to make cultural assumptions.

The useful moat is trustworthy evidence

A larger archive is not automatically a better one. The defensible advantage is an authorized, de-duplicated, time-stamped history whose facts and limitations remain inspectable.

Keep research read-only. A Scout produces questions; controlled experiments produce evidence.

← Part 3 — Data spine Part 5 — The Creative →