sanand0/blog/blind-spot

A blind spot is an expected observation that's missing - the evidence your own assumptions predict but you never checked for. You don't find it by looking harder at the spotlight; you predict what you'd see if you were right, then look for its absence. Then improve the deliverable - don't just report the gap.

How to emit - read first

• DEFAULT, WEAVE: running inside a larger task, change the output in place, add no "blind spots" section. Proposal -> answers the unstated criterion. Analysis -> flags the absent field / survivor-only data inline. Email -> meets the hidden bid before the ask. Brief -> adds the question that exposes the real constraint. Demo -> names the expectation mismatch before it costs credibility.
• EXCEPTION, SURFACE: when the task IS the hunt (review, audit, premortem, "what am I missing", post-mortem) or you're invoked alone.
• Decision rule: fixable inside the deliverable -> weave. Needs his call, or is the whole point -> surface.
• Optional trace: at most ONE terse line - what you folded in, what still needs him.

Pick 2-4 lenses - one sharp beats four decorative

• ABSENCE. What would I expect to see if this were true, but don't? Silent customer, un-run test, survivor-only data, missing field/page, absent baseline, the citation that isn't there.
• SELECTION. Why am I seeing this? If we won the bid, did we overestimate its value (winner's curse)? If they accepted our price, what might they know that we don't (adverse selection)? Who applied, replied, or opted in? Who didn't? We're seeing the survivors. What happened to those who disappeared (survivorship bias)? That exceptional interview or benchmark score may partly be luck. Would it hold up on a repeat (regression to the mean)? Where most people lose, who keeps getting paid? Move: check how the sample was selected, look for the missing cases, temper extreme estimates, and ask what the other side knows. Check who benefits even when the investor loses.
• HIDDEN BID. Am I answering the literal question while missing the real need - validation, scrutiny, a decision, cover, reassurance, permission, emotion?
• UNCLOSED LOOP. What decision, criterion, or commitment did this fail to lock, and what one question closes it? (The most common miss: a strong conversation that ends without the next step or the criteria for "yes".)
• MODE. What state is distorting attention - evangelist, builder, troubleshooter, teacher, guest, listener, demo-high - and what does that mode systematically hide?
• SMUGGLED ASSUMPTION. What must be true for this to work that's untested, undefined, or not shared by the other side?
• CONSTRAINT-AS-SPEC. Is a "blocker" a requirement in disguise? No export / access / SME / time -> aggregates, masked IDs, compute-to-data, synthetic replica, sampling, human-in-the-loop.
• ADOPTION FRICTION. Who must behave differently tomorrow, and what tiny friction - trust gap, workflow change, incentive, missing owner or ritual - stops it sticking?
• ESCAPED ASSET. What reusable thing (prompt, checklist, eval, schema, script, rubric, dataset) from this would be repeatedly, truly useful?

Domain calibration

• High-validity / instrumented (code, data, extraction, diagnostics): tests, logs, absent-field checks, evals, deterministic validation.
• Low-validity / complex (strategy, governance, adoption, markets, long horizon): premortem, outside view / base rates, competing hypotheses, safe-to-fail probes, leading indicators, explicit uncertainty. Distrust the confident intuition.
• Interpersonal: never assert psychology as fact - phrase it as a probe to verify.

When surfacing, one line each

{omission} - {the signal, or why it bites} -> {the probe or the move}

Hard rules

• Change the search space; "what am I missing?" alone yields generic caveats.
• Weave by default; don't bloat the deliverable.
• Don't stop at "risk"; name the failure mode and how you'd detect it.
• Don't accept "no data / access / export / SME" as final; find the safe proxy.
• Don't manufacture omissions; if none are material, say which lenses you checked.
• Run upstream and mid-task, not only as a post-mortem.
• Compose with other skills rather than repeat.
• Keep reasoning internal; output only what helps.