✅ Never skip Phase 4
This is the method’s number one rule: a report delivered without Phase 4 is not a STORM report. Adversarial review and checking every citation against the primary source are what separate the briefing from a well-formatted “dump.” Skipping Phase 4 turns the report into an opinion dressed up as research.
When you trust a finding, you trust it because it was checked — not because the statement sounded convincing. The flowchart below is the decision you should repeat for each finding before accepting it.
The purple diamond is the question you ask about each finding. The cyan branch (YES) is valid only when Phase 4 has run and the primary source checks out; the red branch (NO) never becomes a firm conclusion—it goes to the disputed-signal bar.
12/12 checadas · 1 fabricada · 3 corrigidas · 2 rebaixadas
✓ With Phase 4
- ✓Each citation checks out against the primary source
- ✓The banner states the actual score, without dressing it up
- ✓Confidence scores reflect verified evidence
✗ Skipping Phase 4
- ✗Numbers sound certain, but no one checked them
- ✗An invented citation passes as truth
- ✗Turns into an opinion that looks like research
🎚️ Don't bloat the panel
The temptation is to think that “more agents = better research.” That's the opposite of what the method teaches. The limit is deliberate: at most five lenses e one verifier per citation cluster. Going beyond this doesn't improve coverage—it just burns budget and adds redundant noise.
📐 Why the limit is five lenses
Each new lens is worthwhile only if it brings an angle the others don't cover. STORM's five lenses (practical, academic, skeptic, economist, historian) already cover distinct priorities. A generic sixth lens tends to repeat what the five have already said.
- •Identical lenses find the same thing multiple times.
- •More than 1 verifier per cluster doubles the work without increasing rigor.
- •If you need another angle, swap out a lens—don’t add a sixth (you learned this in 3.3).
✓ Compact dashboard
- ✓5 lenses with distinct priorities
- ✓1 verifier per citation cluster
- ✓Switch lenses when you need another angle
✗ Bloated panel
- ✗10 lenses "just to be safe"
- ✗3 verifiers checking the same citation
- ✗Cost and time go up, quality doesn’t
🔬 Real research only
Every lens and citation must point to a a real source, actually consulted. No made-up studies, figures, or URLs. The golden rule: if a number can't be verified, you downgrades or cuts — never disguises something as confirmed.
🟡 New here? — “disguise”
Disguise is presenting an unverified number with the same confidence as a confirmed one — for example, writing “the market is worth US$ 40 billion” without a primary source, as if it were fact. The honest approach is to mark it as NÃO VERIFICADO, move it to the contested-signal bar, or remove it. The reader needs to know what they can rely on.
Verified? Keep it with the score the evidence supports.
Primary source checks out → stays in the body of the report with the tag CONFIRMADO.
Partially verified? Rewrite it honestly and downgrade it.
Single or commissioned research → reattribute the source and lower the rating; tag REBAIXADO.
Can't verify it? Cut it or move it to disputed.
Without a real primary source → leave it out of the body. Tag NÃO VERIFICADO or FALSO, and reports it in the banner.
🪞 Reveal that the panel is authored
The five lenses look like five independent experts, but they were written by the same author (the same model, with the same framing). That’s why the report always reveals this: when all five lenses agree, it’s a strong hypothesis—not independent proof from the field.
⚖️ Convergence ≠ field consensus
- Authorial convergence: all 5 lenses agree → a good sign, but they all started from the same frame.
- Field consensus: independent researchers, using their own data, reach the same conclusion.
- Practical consequence: present agreement as a “robust hypothesis to check,” never as “proven.”
💡 Practical tip
Leave one sentence visible in the report: "This panel was built by a single author; agreement across lenses is a strong hypothesis, not independent consensus." Small, but it’s what keeps the briefing honest.
💸 Cost and rate limits
A run creates ~9 to 11 agents. This is expected and, better yet, it's predictable: unlike deep research, which launches dozens of identical agents, STORM has a fixed number by design. Knowing the count helps you manage your runs and avoid hitting a rate limit midway through.
Phase 1 — five lenses in parallel
5 Agents general-purpose in a single message. This number is fixed.
Phase 4b — verifiers, 1 per cluster
~4 to 6 agents, each checking a citation cluster against the primary source.
Total: ~9 to 11 agents per run
Predictable by design. The cost doesn’t unexpectedly explode like it does with a swarm of generic agents.
🧯 Manage rate limits
- Space out the runs: running several STORMs in a row in a short period is what usually hits the limit.
- Model by subagent: using cheaper lenses (Haiku/Sonnet) reduces cost—you saw this in 3.3.
- Don't inflate: the limit of 5 lenses / 1 verifier per cluster is also your cost cap.
📋 Final checklist before trusting it
Before making a decision or sharing a briefing, go through this checklist. If everything on the green side is checked and nothing on the red side appears, you can trust it. This is the practical version of the Topic 1 diamond, applied to the entire report.
✓ Ready to trust
- ✓Phase 4 ran and the banner is accurate
- ✓Each citation has a real status tag
- ✓Preprints and single studies are downgraded
- ✓The report reveals that the panel is original
- ✓You personally checked the numbers that matter most
✗ Stop and review
- ✗Any number without a primary source in the body
- ✗Blank banner or one that doesn't match the tags
- ✗Convergence presented as proven consensus
- ✗Panel inflated beyond 5 lenses
Wrapping up the course: a finding comes from a preprint that hasn’t been peer-reviewed, and Phase 4 couldn’t confirm the number. What does the method say to do?
🎉 You’ve reached the end of the course
From the theory of a single prompt to the method’s guardrails, you now know how to run, read, customize, and trust with discernment in a STORM brief. These are the principles that underpin everything:
Congratulations:
You’ve completed the STORM Research course 🎉 — now you can run it on your own topics.