Contents
6. Value of Information
Research is not free. For unanswered question q:
Research when: VOI(q) > 0
Do not spend three hours proving something whose answer would not change the decision.
10. Candidate Task Market
Every observation can generate CandidateTasks. Rather than one linear queue, maintain a task market.
Candidate sources: goals, failures, signals, runtime events, customer events, revenue events, GitHub, Cursor, email, calendar, market change, memory, research, innovation, Sentinel findings, skill failures, MOTHER alerts, unfinished prior work.
Tasks compete for execution resources.
11. Task Utility Function
For candidate i calculate an evolving utility estimate.
Positive terms: R = economic value, S = strategic value, U = unblock value, I = information gain, O = option value, L = learning value, X = cross-domain reuse, T = urgency, C = confidence, A = autonomy enablement, V = reversibility.
Negative terms: H = Jeramie attention, K = compute cost, D = downside, P = policy/authority gap, Q = collision probability, F = stale-evidence penalty, N = implementation complexity, E = irreversibility.
Weights are learned. Do not permanently hard-code founder priorities from initial assumptions.
A full Task Utility Calculator with all 11 positive terms and 8 negative terms lives in the JGI site, Tool A. This workbench consumes its utility scores in the Portfolio Selection Optimizer below.
12. Jeramie Attention Shadow Price
Treat human attention as costly.
Maintain: AttentionShadowPrice.
Estimate from: frequency of interruptions, time required, opportunity cost, task outcome, whether interruption could have been avoided.
A technically cheap task may be expensive if it consumes founder attention.
Optimize: VerifiedValue / FounderMinute
13. Portfolio Selection
Do not simply execute the highest-scoring task. Choose a portfolio of work constrained by: compute, API budget, time, available workers, resource collisions, authority, deadlines, dependencies.
Use simple greedy selection when sufficient. Use more sophisticated optimization only when it materially changes outcomes.
14. Explore / Exploit Engine
Default: 80% exploit, 20% explore. But dynamically adjust. Use bandit-style reasoning. Candidate approaches may use: Upper Confidence Bound, Thompson Sampling, Bayesian optimization.
Exploration should increase when: uncertainty is high, new capabilities appear, existing queue is low value, market conditions shift, cross-domain transfer appears promising.
Decrease when: system unstable, deadline critical, risk elevated, known high-value work waiting.
50. Expected Value of Retry
Retry if: P(success next attempt) x value > retry cost + risk + delay cost.
Do not retry because "maybe it'll work."
54. Economic Intelligence
JGI tracks: cash, revenue, pipeline, expected value, customer acquisition, conversion, cost, time, capital at risk, strategic value.
Never merge: pipeline, revenue, cash.
67. Option Value
Some experiments are worthwhile because they unlock future choices.
This lets JGI invest intelligently in enabling infrastructure.
110. Value Density
Measure: VerifiedValue / agent-hours, and VerifiedValue / compute dollar.
This reveals whether more autonomy is actually useful.
125. Resource Forecasting
Predict: compute, time, API usage, human attention, storage, service load, before large missions.
126. Budget Controller
Operate under budgets. When budget constrained: prioritize expected utility. Never silently exceed unknown financial limits.
127. Reservation of Capacity
Do not consume all compute on exploration. Reserve capacity for: critical alerts, customer events, runtime incidents, high-value opportunities.
128. Priority Preemption
A currently running low-value mission may be paused when a new critical high-value mission appears. Checkpoint first. Release resources. Resume later.
Interactive Tools
All tools compute live from your inputs. Formulas quoted from the paper are marked "paper formula". Anything else is marked "illustrative implementation" and the page says exactly what the paper does and does not define.
Tool 1. Value of Information Calculator Section 6 Paper formula
VOI(q) = ExpectedUtility(after learning answer) - ExpectedUtility(with current knowledge) - CostOfInvestigation. Research when VOI > 0.
Tool 2. Attention Shadow Price Calculator Section 12 Illustrative implementation
The paper defines the shadow price concept and the quantity to optimize, VerifiedValue / FounderMinute, but gives no numeric formula. This tool implements it as: founder value per minute = founder hourly value / 60; task cost = founder minutes x value per minute; value per founder minute = task verified value / founder minutes. Verdict bands are illustrative: at or above the founder's own value per minute, founder time is justified; at or above $1.00 per minute, delegate; below that, drop.
Tool 3. Portfolio Selection Optimizer Section 13 Paper formula, exact solve
Maximize sum(Utility_i * x_i) subject to the time budget, x_i in {0, 1}. The paper says use simple greedy selection when sufficient and more sophisticated optimization only when it materially changes outcomes. This tool solves the knapsack exactly by exhaustive search (6 tasks, 64 combinations), so the result is the true optimum under the stated budget, not a greedy approximation. Edit any task's utility or time and re-run.
| Task | Utility (0-100) | Time (hours) | Include? |
|---|---|---|---|
| - | |||
| - | |||
| - | |||
| - | |||
| - | |||
| - |
Tool 4. Explore / Exploit Engine Section 14 Illustrative implementation
The paper's default is 80% exploit, 20% explore, adjusted dynamically, and names Upper Confidence Bound as a candidate approach but gives no formula. This tool implements UCB: UCB = mean + c x sqrt(ln(total trials) / pulls of this arm). The explore arm wins when its UCB beats the known exploit payoff. Exploration should increase when uncertainty is high, new capabilities appear, the queue is low value, market conditions shift, or cross-domain transfer looks promising; decrease when the system is unstable, a deadline is critical, risk is elevated, or known high-value work is waiting.
Tool 5. Expected Value of Retry Calculator Section 50 Paper formula
Retry if: P(success next attempt) x value > retry cost + risk + delay cost. Do not retry because "maybe it'll work." Multi-attempt view: cumulative success over n attempts is 1 - (1 - p)^n, with cost paid per attempt.
Tool 6. Option Value Calculator Section 67 Paper formula
OptionValue = ProbabilityFutureOpportunity x Magnitude x FlexibilityCreated. Some experiments are worthwhile because they unlock future choices. Invest in the enabling work when OptionValue exceeds its cost.
Tool 7. Budget Controller Sections 125, 126, 127, 128 Illustrative implementation
The paper gives rules, not formulas: operate under budgets, prioritize expected utility when constrained, never silently exceed unknown financial limits, reserve capacity for critical alerts, customer events, runtime incidents, and high-value opportunities, and preempt a running low-value mission when a new critical high-value mission appears, checkpointing first. This tool implements those rules: a reserve percentage is held back, missions are ranked by priority, and a new critical mission preempts the lowest-priority running missions until it fits, lowest first.
| Mission | Cost ($) | Priority | Status |
|---|---|---|---|
Tool 8. Value Density and Economic Intelligence Sections 54, 110 Paper formula
Value Density = VerifiedValue / agent-hours, and VerifiedValue / compute dollar. This reveals whether more autonomy is actually useful. Economic Intelligence tracks cash, revenue, pipeline, expected value, customer acquisition, conversion, cost, time, capital at risk, and strategic value, and never merges pipeline, revenue, and cash. Enter the three separately and the tool keeps them separate.
What this workbench does not do
- It does not run missions, spend money, or contact anyone. Every number is entered by hand and computed in your browser.
- It does not claim the paper's concepts are proven. JGI is an internal engineering objective, not a scientific claim.
- Portfolio Selection solves the knapsack exactly for the tasks you enter. It does not model dependency constraints, resource exclusivity, or authority constraints from Section 13. Those are listed as constraints the paper names but does not formalize, and the tool says so.
- Verdict bands in Tools 2, 4, and 7 are illustrative implementations. The paper defines the concepts and the direction, not the thresholds.