LIFTED / FIELD EXPERIMENTS / THE SELECTION LOOP
LONGITUDINAL STUDY · TWO RUNS RECORDED
How does human judgment turn AI-generated possibilities into actions worth pursuing—and which of those choices endure?
This evolving experiment follows the journey from candidate generation through human classification, execution and subsequent survival. Run 001 (15 in 75) and Run 002 (20 in 3) have different conditions; the numerals describe historical sessions, not limits on what can be tested.
01 / THE MEASURED BASELINE
The initial figures come from the human participant's report. They capture review and promotion. Task outcomes and financial returns remain for later observation.
The strongest fifteen actions in the EDRP library were processed.
Self-reported session total; individual action durations remain unmeasured.
Two survived selection to enter the actual weekly workflow.
75 minutes divided by 15 decisions. Individual review times remain unmeasured.
Selection attention allocated across two promoted threads. Execution cost and demonstrated value will be measured separately.
2 of 15 reviewed actions entered the weekly workflow.
Classification guardrail. Thirteen actions stayed outside the weekly workflow (86.7%). That total contains distinct dispositions: explicit rejections, decisions closed after learning, and deferrals. A subsequent attempt to compile the recorded thread decisions recovered an incomplete tally. The explicit rejection rate awaits verification.
02 / THE RAPID SCREEN
Twenty narrowly scoped website-refinement actions were read and classified directly by the human. The record comes from the participant's report, contemporaneous private observation, and already approved Field Note 147 ↗.
Short, website-specific opportunities screened without an additional AI evaluation pass.
Approximate participant-reported duration, not independent stopwatch timing.
Ten rejected (50%); three deferred (15%); seven accepted for execution (35%).
03 / COMPARATIVE MEASURES
Reported durations and decision counts permit a quantitative description. The two sessions use different selection gates: promotion to weekly workflow versus immediate acceptance of a bounded action.
| Measure | 15 in 75 / EDRP | 20 in 3 / Website actions |
|---|---|---|
| Actions examined | 15 | 20 |
| Human attention | 75 min reported | ≈3 min reported |
| Classification throughput | 0.2 actions/min | ≈6.7 actions/min |
| Decisions | 2 promoted; 13 not promoted | 7 accepted; 3 deferred; 10 rejected |
| Acceptance / promotion | 13.3% promoted to ongoing work | 35% accepted immediately |
| Selection attention / positive decision | 37.5 min per promoted thread | ≈0.43 min per accepted action |
| Downstream evidence | 2 selected for weekly work; later results unknown | Execution, survival and realized effects unknown |
Arithmetic: 20 ÷ 3 ≈ 6.7 classifications per minute, versus 15 ÷ 75 = 0.2, or roughly 33 times the nominal throughput. The attention-per-positive-decision figures are 75 ÷ 2 = 37.5 and 3 ÷ 7 ≈ 0.43 minutes. The positive-decision stages are different, so their ratio is descriptive rather than a standardized gain in selection quality.
The participant rejected ten, deferred three, and selected seven. He described the seven as successful against his prospective criterion: anticipated usefulness per human minute, assessed using contextual mental simulation.
The participant attributed part of the difference to the tighter generating prompt. Candidate type, venture coverage, review modality and decision stage also changed. Causal attribution remains open.
Implementation, later abandonment, further human time, spawned threads and externally observed consequences still require follow-up. The perceived usefulness numerator was not numerically measured.
The human's own success gate is recorded as stated. Selection was complete in the participant's judgment; no completed work products are asserted. Evidence: private Lifted cycle dated 18 October 2026 and approved Entry 147. All duration-derived measures inherit approximate timing uncertainty.
04 / THE CONVERSION PATH
The original EDRP investigation generated an action library intended to support up to 100 worthwhile prompts. It returned 70, ranking the first 15 highest.
Fifty-five actions remain outside this session’s review. The two promoted threads reflect selection priority; their downstream results are open for observation.
05 / WHAT THE PARTICIPANTS OBSERVED
The exercise revealed a difference between compressing the wording of a task and reducing the effort required to judge its output.
The human reviewed fifteen candidate execution threads within seventy-five minutes. Some were rejected immediately. Others were entertained long enough to expose useful facts, reveal an inadequate premise, or refine the opportunity. Many were closed or set aside after review.
By the end, only two threads entered the ordinary weekly workflow. Their future was deliberately left undecided: either might finish in hours, or generate many new branches of investigation. Continuation established priority. Results remain to be observed.
The human objected to describing the other thirteen as wasted effort. Every considered option required judgment; even a well-founded rejection can reduce uncertainty and prevent later commitments. The expense was real, and its value extends beyond the threads that survived.
“Rejected, those entertained all contribute to that expense.”
Human participant, 18 October 2026. The relevant cost is the full 75 minutes, including attention spent on proposals subsequently rejected or deferred.
Research ruled out weak opportunities, corrected assumptions or narrowed the viable path. Its information could contribute even when the action was rejected.
Some assignments introduced classification, comparative modeling or extensive investigation beyond the practical means and immediate decision of the exercise.
Templates, scoring systems and operating frameworks risked converting simple opportunities into additional administrative systems requiring human inspection.
Observed loop bottleneck: AI → workflow. The AI initially optimized for producing a rich, copy-ready library, but insufficiently filtered proposals for downstream adjudication cost. The human became the final complexity filter. This interpretation remains provisional, pending evidence from comparable sessions.
A second failure occurred in measurement. When asked to compile the first fifteen decisions, the assistant failed to recover recorded rejection labels and mistakenly treated incomplete records as a zero-rejection tally. The human corrected it. Recording an outcome and being able to recover it across threads are different capabilities.
06 / REPEATABLE FIELD PROTOCOL
The original fifteen-action, seventy-five-minute format remains available for matched replications. Other runs can vary the candidate pool, scope and time budget when those conditions are stated explicitly. Comparison must distinguish like-for-like replication from exploratory contrasts.
Start from a documented venture investigation. Use actual operating maturity and established decisions to eliminate duplicates, premature infrastructure, and work whose value depends on unavailable evidence. Record the pool size, shortlist procedure and major changes in selection rules.
State the intended sample size and timebox for each run, or explicitly declare an untimed review. Record actual elapsed human minutes and actions classified. Preserve the original 15-action, 75-minute frame only for sessions intended as matched replications.
Promote to weekly execution; reject as unsuitable or lower value; defer for an explicit trigger; or close after useful research resolves the immediate question. Record a short reason and any meaningful uncertainty resolved. Preserve original IDs.
Count review time for every action, including rejected and entertained candidates. Separate session time from earlier AI research or execution time and later human involvement. Record interventions needed to untangle excessive complexity or unclear outputs.
At a declared follow-up point, count confirmed deliverables, closures, newly spawned threads and added human minutes. Use verified evidence from inspected threads. Evaluate new branches by their contribution to justified decisions or consequential results.
Report the relevant decision rates and attention per selected action alongside productive rejections, uncertainty reduced and actual downstream value. Document changes in venture, task difficulty, AI model or protocol that limit comparability. An improved session may promote fewer actions while improving decision quality.
07 / INTERPRETATION RULES
Metrics make the cognitive cost visible. Interpreting them requires decision quality, productive rejections and subsequent results.
| Measure | Definition | Boundary |
|---|---|---|
| Review time / action | Total human review minutes ÷ number adjudicated. | Individual review-time differences remain unmeasured. |
| Attention / survivor | Full session attention ÷ actions promoted. | Defined where promotions occurred. Measures selection attention; lifetime costs and returns belong to follow-up. |
| Promotion rate | Promoted ÷ adjudicated. | Higher rates may reflect lower selection standards; decision quality remains essential. |
| Rejection rate | Explicitly rejected ÷ adjudicated. | Track unpromoted and explicitly rejected actions separately, using recoverable decisions. |
| Productive rejection | Rejected or closed action with a documented useful insight or avoided commitment. | Count knowledge gains supported by concrete decision records. |
| Downstream yield | Verified useful results, justified decisions and viable follow-on work from promoted actions. | Measure spawned threads through verified downstream consequences. |
| Human follow-through | Human attention after promotion, tracked separately from selection time. | Measure later attention separately from the initial 75 minutes. |
Both runs rely on participant-reported time. The twenty-action session supplies a second observation with a different task and selection method, rather than a controlled comparator. Item-level timing and downstream observations remain outstanding.
08 / LONGITUDINAL RECORD
The two October observations remain distinct. Future runs can test whether the apparent screening advantage persists, and whether selected work continues to earn attention.
A commercial intelligence investigation of EDRP, a free industrial equipment registry with a separately proposed Representation service, produced 70 bounded execution prompts. Its fifteen highest-ranked actions were reviewed in a single 75-minute period.
Result: a usable attention baseline and a concrete proposed intervention—better preselection of proportionate actions. Evidence for improved performance and economic consequences will emerge from later observations.
From twenty concise website-refinement actions, the human classified ten as rejected, three as deferred and seven as immediately worth executing. The user reviewed the list directly, bringing context from the ten-venture portfolio to each quick decision.
Interpretation: narrower instructions appear to have made the decisions more mentally tractable. The observation cannot isolate prompt scope from differences in candidate design or review context. The already-authorized Field Note 147 ↗ documents the original account.
Future runs may repeat the 15-in-75 frame for comparability or vary the scope and timebox. Declare the design, measure real attention and document deviations.
Pending observation. Both rapid closures and extensive downstream branching will be assessed by their consequences.
09 / EVIDENCE & PROVENANCE
This public-safe account draws on the participant's observations and the AI-generated action library while preserving the privacy of the conversation and analytical ledger.
Reported facts: fifteen action threads reviewed in seventy-five minutes; two promoted; many dismissals grounded in excessive depth or complexity. Derived arithmetic: five minutes per action, 37.5 minutes per survivor, 13.3% promotion. Inference: upstream AI task selection may be shifted toward practical sufficiency to reduce downstream cognitive cost.
In the baseline, individual dispositions and surviving IDs await verification. In the rapid screen, aggregate dispositions (10 rejected, 3 deferred, 7 accepted) are recorded, while individual action IDs and downstream execution outcomes await verification. Neither duration has item-level timing.
The private Lifted ledger remains the analytical source of truth for its own records. This publicly authorized testing surface complements the Field Notes; the private ledger retains its separate analytical authority. Future updates should distinguish observation dates from publication dates and retain historical results unchanged.