AI analysis hallucination is a specific failure mode I've documented extensively: you give the model some data, ask for insights, and it returns confident-sounding conclusions that aren't supported by the data you provided. The model fills gaps in data with plausible-sounding inventions. The fix is not a better model — it's better signal. The sinc analysis template includes a dedicated DATA band that anchors every claim, and a CONSTRAINTS band that explicitly bans fabrication.
This example analyzes SaaS cohort data — a common analysis task where hallucination risk is high because the model may invent industry benchmark comparisons that weren't provided:
{
"formula": "x(t) = Σ x(nT) · sinc((t - nT) / T)",
"T": "specification-axis",
"fragments": [
{
"n": 0,
"t": "PERSONA",
"x": "You are a growth analyst who works from data, not intuition. You never state a conclusion without citing a specific number from the data provided. You are comfortable saying 'the data does not support a conclusion on X' when that is true."
},
{
"n": 1,
"t": "CONTEXT",
"x": "I'm analyzing 6-month cohort retention for a B2B SaaS product to identify where in the customer lifecycle we're losing users. The goal is to find the highest-leverage retention intervention."
},
{
"n": 2,
"t": "DATA",
"x": "Cohort data (Month 0 = signup, % retained by month):\nJan cohort: M0=100%, M1=78%, M2=61%, M3=55%, M4=53%, M5=52%\nFeb cohort: M0=100%, M1=74%, M2=58%, M3=51%, M4=49%, M5=49%\nMar cohort: M0=100%, M1=71%, M2=54%, M3=47%, M4=45%, M5=—\nAvg contract value: $290/month. Onboarding: 14-day email sequence. Support: in-app chat only."
},
{
"n": 3,
"t": "CONSTRAINTS",
"x": "Only draw conclusions from the data provided. Do not compare to industry benchmarks unless I provided them — I did not. Do not state a month-2 retention rate unless you can cite the exact number from DATA. If data is missing (e.g., Mar M5), note the gap and reason from available data only. Do not recommend solutions — only identify where the drop is happening and what the data suggests about timing."
},
{
"n": 4,
"t": "FORMAT",
"x": "3 findings, each with: (1) the specific data point, (2) what it indicates, (3) one follow-up question this raises. Then a 2-sentence summary of the highest-leverage retention window. No introductory paragraph."
},
{
"n": 5,
"t": "TASK",
"x": "Analyze the cohort retention data to identify where dropout is concentrated and what the timing suggests about causation."
}
]
}Three elements of the sinc analysis template work together to prevent hallucination:
DATA band discipline: Put every number the analysis should use in the DATA band. If a number isn't there, the model shouldn't reference it. The DATA band is your analysis boundary.
CONSTRAINTS anchoring: "Only draw conclusions from the data provided" is the single most important constraint for analysis prompts. Pair it with "do not compare to industry benchmarks unless I provided them" — this eliminates the most common hallucination pattern (inventing benchmark comparisons).
PERSONA grounding: "Comfortable saying the data does not support a conclusion" gives the model permission to express uncertainty rather than confabulate. Without this, models under pressure to find insights will invent them.
Analysis prompt tip: If your data has gaps (missing months, null values), name them explicitly in the DATA band. "Mar M5 = not yet available" is better than leaving the cell empty. The model handles explicit gaps with more precision than implicit ones — and acknowledges them in the output rather than interpolating without flagging.
Here is my cohort retention data. Analyze it and tell me where we're losing users and what we should do about it. [data pasted below]
CONSTRAINTS: Only use provided data. No benchmarks I didn't give you. Note data gaps.
FORMAT: 3 findings with specific data citation + follow-up question each. No intro paragraph.
TASK: Identify where dropout concentrates and what timing suggests about causation.
The raw prompt produces analysis that includes confident comparisons to "typical SaaS benchmarks" the model invented, plus recommendations that weren't asked for. The structured prompt produces three grounded findings, each citing the exact numbers from the data you provided.
Try AI Transform — Structure Your Analysis Prompt Free