Assumption Marker (F5)

Cockpit's fifth principle: make implicit assumptions explicit when they meaningfully affect output. Solves spec.md §2 FM2 specifically for data analysis scenarios.

Behavior — dual-track

Track 1: Plan-time declaration

Include known assumptions in the <cockpit-plan> block's Assumptions: field at task start.

Track 2: Mid-execution declaration

When you discover or make a non-trivial assumption mid-task (most common after seeing data), output a standalone block:

<cockpit-assumption>
Assuming: [the specific assumption]
Why: [why making this assumption — what data or context drove it]
Impact: [what changes if the assumption is wrong]
</cockpit-assumption>

Frequency: only non-trivial

Declare assumptions that meaningfully affect conclusions. Skip trivia like:

• "Assuming UTF-8 encoding"
• "Assuming no NaN values"
• "Assuming integer overflow won't happen"

DO declare:

• "Assuming user_segment defined as last 30 days orders bucketed (high/mid/low) — data has no explicit segment field"
• "Assuming missing payment_status means failed (not pending) — no docs available"
• "Assuming cohort grouped monthly not weekly — stronger signal at month level"

The litmus test: if this assumption is wrong, would the user's interpretation of the result change? If yes, declare. If no, skip.

Why this exists

Data analysis tasks are full of implicit choices that shape conclusions:

• Which slice dimension
• How to handle missing values
• What proxy to use when explicit field unavailable
• What grouping cadence

Without assumption markers, these are invisible — user sees the chart but doesn't know what assumption made it look that way. With markers, user can push back early ("wait, why did you bucket by 30 days? I want 90 days").

Especially important for

• Data analysis — assumptions in slicing / grouping / methodology
• Deep research — assumptions about source authority / recency
• Code modifications — assumptions about user intent / API contract

Non-blocking design

Like thought blocks, assumption markers are visible but don't pause execution. Agent declares, user reads in real-time, agent proceeds.

Implementation

Pure skill-level instruction. The blocks appear in Claude Code's output stream and persist in transcript for later audit. v0.2 may add specialized parsing/logging.