attention-ledger

name	attention-ledger
description	Use this to account for where AI-assisted attention actually went and what it bought - including the uncomfortable parts.

Attention Ledger

Work Classes

Classify episodes approximately into:

PRODUCT VALUE / MAINTENANCE / LEARNING / RESEARCH / INFRASTRUCTURE / AGENT-META WORK / ADMINISTRATION / REWORK

Estimate two shares per class: share of sessions and share of downstream outcomes. The gaps between those two columns are the finding.

Name work that FEELS high-leverage but historically produced little downstream benefit. Name low-visibility work with unusually high leverage. This section should be evidence-driven and possibly uncomfortable.

Explore vs Exploit

EXPLORE: alternatives, technologies, architectures, models, strategies. EXPLOIT: commit, implement, ship, polish, scale.

Measure mode switching. Hunt for:

• premature commitment before options were generated
• excessive exploration with no commitment
• architecture churn, tooling churn, model churn
• repeated redesigns of the same subsystem
• unfinished promising work

Do not call exploration waste unless outcome evidence supports it. Exploration is R&D until proven otherwise.

Risk vs Deliberation

Compare deliberation depth against reversibility and downside:

Decision type	Observed deliberation	Appropriate?
cheap + reversible	hours of comparison?	over-deliberation
expensive + irreversible	minutes?	under-deliberation

Find both rows in the real data.

Metrics Discipline

Never equate messages, code volume, commits, tokens, or session length with productivity. Use completions, regressions, correction loops, rework, deployment success, later reversals, maintenance burden. State cycle-time uncertainty honestly.

One-Line Memory

Attention follows interest. Audit whether value followed.