Effort without better judgment.
Context repeatedly reconstructed; long unfiltered option lists; avoidable clarification; duplicated approvals; preventable errors and their repair.
LIFTED / APPLIED COLLABORATION / V0.1 · 10 NOV 2026
Better decisions.
Less wasted attention.
A working method for human–AI collaboration: reduce avoidable decision friction without giving up comprehension, independent judgment or accountability.
STATUSProvisional operating framework
EVIDENCE BASEOne longitudinal collaboration; selected traceable cases
VALIDATIONProspective effectiveness not established
01 / THE PROBLEM
Not all friction is a defect.
AI can reduce the cost of producing work while increasing the cost of understanding, selecting and verifying it.
Effort that protects a decision.
Checking an assumption; understanding a dependency; inspecting evidence; weighing alternatives; refusing an unsafe shortcut; verifying an outcome.
The objective is not less human involvement. It is more effective human attention at consequential decision boundaries.
02 / OPERATING PROCEDURE
The Lifted Decision Loop.
Five stages. Explicit responsibility. Verification before a task is treated as complete.
Define the outcome.
Clarify purpose, constraints, evidence and who has authority to decide.
Reduce the burden.
Retrieve context, narrow alternatives, distinguish facts from assumptions, expose risks.
Exercise judgment.
Approve, challenge, change direction or reject; understand consequential trade-offs.
Act and verify.
Implement a bounded change, test it, identify deviations and preserve rollback paths.
Evaluate the loop.
Compare intended and actual outcomes, classify the bottleneck and revise only with evidence.
A standing authorization can cover low-risk, reversible steps. Significant commitments require an informed human decision. Successful execution, passing tests, deployment and real-world usefulness are different outcome states.
03 / SIX PROVISIONAL PRINCIPLES
Rules that must survive testing.
Each principle is an intervention hypothesis, not a finding. Evidence that contradicts it should change the method.
01 / CONTEXTHYPOTHESIS
Context before production.
Establish the real objective, prior decisions and missing information before producing a consequential deliverable.
Challenge: Context retrieval costs more than it saves on comparable tasks.
02 / SELECTIONHYPOTHESIS
Decisions before deliverables.
Present compact, evidence-backed choices rather than requiring the human to screen elaborate proposals.
Challenge: Compression hides alternatives or reduces decision quality.
03 / AUTONOMYHYPOTHESIS
Delegation proportional to risk.
Expand AI autonomy where actions are bounded, reversible and verifiable; escalate where consequences or uncertainty rise.
Challenge: This allocation increases oversight costs or misses material risks.
04 / APPROVALHYPOTHESIS
Approval requires understanding.
Expose dependencies, trade-offs, uncertainty and rollback options before consequential authorization.
Challenge: Added review provides no meaningful comprehension or error-detection benefit.
05 / DIAGNOSISHYPOTHESIS
Correct mechanisms, not symptoms.
Determine whether a failure is local or reflects a recurring architecture, evidence or handoff problem.
Challenge: Broad diagnosis adds complexity without reducing recurrence.
06 / CLOSUREHYPOTHESIS
Verification closes execution.
Separate implemented, tested, deployed, accepted and operationally effective states.
Challenge: Additional verification consistently fails to improve reliability relative to its cost.
Cross-cutting rule: neither participant may silently replace the human's objective with an alternative success criterion. The intended result must remain visible throughout the loop.
04 / FIRST PROPOSED TEST
Can better briefs reduce total review cost?
Proposal registered here as a test design. No new trial results are claimed.
EXPERIMENT / DECISION BRIEF
Make a consequential choice inspectable before approval.
Hypothesis: a structured, risk-calibrated brief reduces human review and correction effort without impairing comprehension or material error detection.
- Setting: approximately 10–12 consequential but reversible engineering decisions in one active project.
- Comparison: use the existing review workflow and the proposed brief on reasonably comparable decisions; alternate or randomize allocation where practical.
- Brief fields: proposed change, rationale, affected systems, evidence, material risk, alternative, rollback and exact human decision required.
- Measure: initial review plus clarification, correction and rollback time; comprehension, pre-approval detection, post-approval defects and decision survival.
- Interpret: examine contrary cases and confounders; retain, revise or reject the protocol before expanding it.
A provisional continuation target is at least 20% less total human attention, with no newly missed material risks or detectable comprehension loss. This is an operational threshold, not a statistical significance claim. A severe missed risk triggers reassessment even if time was saved.
05 / ORIGINATING EVIDENCE
Real work, limited inference.
These public records motivate the method. They do not establish that the proposed interventions work.
SELECTION LOOP / OCT 2026
Scope, speed and selection
Two reported selection sessions with different conditions and positive-decision gates. Not a controlled productivity comparison.
Inspect the field experiments ↗ FIELD NOTE 154 / VERIFICATIONReview burden after delegation
Repeated implementation corrections motivated the question of whether delegation transfers rather than removes verification labor.
Read the published observation ↗ FIELD NOTE 157 / APPROVALHuman authority requires diligence
A disclosed architectural dependency was approved without adequate human review. Later rejection did not make it unauthorized.
Read the corrected attribution ↗ FIELD NOTE 158 / DIAGNOSISFrom local repairs to coherent classification
A human-driven reframing examined whether locally defensible fixes addressed the shared diagnostic model. Production improvement was not yet verified.
Read the published case ↗The H / A / L / X bottleneck taxonomy continues to classify human, AI, shared-loop and external constraints. Observations, hypotheses, interventions and verified outcomes remain separate. The private ledger remains the canonical research record; this page is a public derivative.
Improve the combined system, not merely the output of either participant.
This is a single-case, evolving method with selective visibility and no established causal effect. We will preserve counterexamples, identify measurement gaps, and version changes to rules rather than retroactively converting hypotheses into findings. Transfer to other settings must be tested, not assumed.