The method fails in predictable ways. Most of them feel like success.
That's the difficulty. A contaminated session doesn't produce obviously bad output — it produces confident, coherent, mutually reinforcing output that reads exactly like a good result. The failures worth knowing are the ones with no error message.
| Failure | What it looks like | Severity | The fix |
|---|---|---|---|
| False consensus | All three agree, quickly and confidently, on something your own experience says is hard. | High | Treat unanimity as a prompt to ask "what would have to be true for this to be wrong?" — then check that specific thing externally. Consensus is not verification. |
| Contaminated briefing | You paste one platform's answer into another to "get its take." Their outputs are now dependent. | High | Independence is only available on the first pass. Run all three cold, then compare. Cross-referencing afterward is fine and useful — before is fatal. |
| Loaded brief | Three agreeable, well-argued responses that all endorse the direction you'd already chosen. | High | Reread the brief before dispatching. If your preferred answer is visible in the question, rewrite it so any answer is genuinely possible. |
| Summarized source | You paste your own précis of a document instead of the document, so all three reason about your reading. | High | Upload the file itself to each platform. Whatever your summary missed, three platforms will now miss identically — and call it agreement. |
| Unverified unanimity | The claims all three agreed on go straight into the synthesis unchecked, because agreement felt like checking. | High | Route agreed-upon factual claims through a grounded source anyway. These are the least-checked claims in most sessions. |
| Delegated synthesis | You hand all three responses to a fourth model and ask it to merge them into a conclusion. | High | That is not synthesis, it is a fourth opinion with extra steps — and it reintroduces exactly the single-voice problem the method exists to solve. The conductor synthesizes. |
| Anchoring | You read and start evaluating the first response before the other two arrive. Everything after gets graded against it. | Medium | Read all three before judging any. If you must wait on a slow deep-research run, wait — don't fill the time by forming a view. |
| Time-skew misread | A disagreement about recent facts gets adjudicated as if it were about substance. | Medium | Ask each platform what date its answer reflects and whether it retrieved or recalled. A sourced fresh answer beats a confident stale one regardless of tone. |
| Deadlock | Genuine conflict you cannot adjudicate, and each additional round restates rather than resolves. | Medium | Reframe, decompose to the specific hinge, ground the factual part, then accept unresolvability as a finding or escalate to a human specialist. |
| Session drift | Two sessions later you're answering a subtly different question than the one you started with, and you can't say when it changed. | Medium | Write the question down at the start and reread it before synthesis. A running context document for anything spanning more than one sitting. |
| Context contamination | An old thread's assumptions bleed into a new question because you reused the conversation. | Medium | Fresh thread per session, on every platform. Name sessions identically across platforms so you can find them without reusing them. |
| Session fatigue | You notice yourself accepting the first reasonable answer, skipping the comparison, seeking permission rather than information. | Medium | This is a conductor problem, not a platform problem. Stop, log where you are, return when you have the attention to do it properly. |
| The session that won't close | Round after round, each producing smaller refinements you keep treating as progress. | Medium | Diagnose which it is: genuinely unresolvable for now, avoidance of the decision, or the ensemble's ceiling reached. All three mean stop. |
| Format mismatch | Three responses in three shapes — a table, an essay, a bulleted list — that resist comparison. | Low | Specify the output format in the brief. This is cheap to prevent and tedious to fix afterward. |
| Ceremony | Full ensemble convened for a question one platform could have answered, burning attention you'll want later. | Low | Route first. Not every question deserves the ensemble, and knowing when to step back is part of the discipline. |
The three that account for most bad sessions
If you only internalize three lines from this table, make them these — they are upstream of nearly everything else:
- Never let one platform see another's answer before it writes its own. The independence is the entire mechanism. Everything downstream is worthless without it.
- Never let agreement stand in for verification. Correlated error means unanimity is a weak signal, and it is weakest exactly where models are strongest.
- Never write a brief whose preferred answer is visible. You will get it back three times and mistake it for corroboration.
A member that has gone flat
Sometimes a platform that used to carry its seat stops contributing anything the others don't. Before concluding it has declined, check three things in order: whether you've been assigning it the wrong mode, whether your prompts to it have gotten lazier than your prompts to the others, and whether the platform itself has changed under you. Models get updated without announcement, and a capability you relied on can move.
If it still underperforms after those checks, reconsider its role. The methodology endures; the roster may not.
When a platform is simply unavailable
Outages happen mid-session. For moderate stakes, proceed with a reduced ensemble and note in your log that you did. For high stakes where that seat's mode is essential — particularly grounding — delay rather than substituting silently. A two-platform session recorded as a three-platform session is a record you'll trust more than you should six months from now.
Notice that almost none of these are platform failures. They are conductor failures — shortcuts in briefing, sequencing, verification, and attention. The platforms are rarely the weak link in a bad session. The process around them is.