The Model Boundary
v1 —
Full inventory referenced from manuscript.html §7.5, PAYBACK · AI Economy Model v1
Generated from the canon doc _docs/economics-v1/PAYBACK-MODEL-BOUNDARY.md (deslop.media editorial, 2026-09-17); the three tables below reproduce its Bucket 1–3 rows verbatim.
What a player steers, what the economy fixes so seasons stay comparable, and what is left out of the model entirely — 42 rows across three buckets. Every row describes the model, not the real economy.
Bucket 1 — kept and played
What a lab CEO actually decides, once a year.
| Reality | How it appears in the model | Source section |
|---|---|---|
| Data-center build | Choose affordable investment level 0, 1, or 2 annually. Each level orders centers equal to 10% of opening industry fleet; deliveries become usable next year. Capacity orders are endogenous to lab strategies. | manuscript.html §§3.1–3.2, Appendix B; constants.mjs: INVESTMENT_LEVELS, LOT_FRACTION_OF_INDUSTRY |
| Training allocation | Choose 0%, 25%, or 50% of compute for dedicated training; remaining compute can serve. Capability gains deploy next year. Training difficulty is 1.5; ordinary gain coefficient is ln(2)/1.5. | manuscript.html §§3.1, 3.6, Appendix B; constants.mjs: TRAIN_SHARES, TRAINING_DIFFICULTY |
| Pricing and willingness to serve | Submit one admissible post from the capability-reachable price menu. A lab serves a task only if qualified, able to supply compute, and posting no more than its ceiling. The post is a choice; actual pay also includes the fixed scarcity premium. | manuscript.html §3.5, Appendix B |
| Investment timing and harvest [inferred] | Repeated build, training, and post choices permit delaying investment, expanding, stopping, or resuming. There is no separate harvest button or freely chosen liquidation price. These are strategies composed from the three controls. | manuscript.html §§3.1–3.2, 4.3; constants.mjs: INVESTMENT_LEVELS, TRAIN_SHARES, TERMINAL_RECOVERY_PER_CENTER |
Bucket 2 — kept, fixed
What holds still across every season so results stay comparable; the player cannot decide any of it.
| Reality | How it appears in the model | Source section |
|---|---|---|
| Three starting firms and fixed capital | Three labs each start with 1,000 Gold, ten centers, and capability 1. The aggregate investment denominator is the initial 3,000 Gold. | manuscript.html §§3.1, 3.7; constants.mjs: LAB_IDS, STARTING_GOLD, STARTING_FLEET, STARTING_INTELLIGENCE_X |
| Rival rules and simultaneous commitment | Two deterministic, state-responsive rivals follow frontier-building and volume-serving policies. Their policy families are fixed; moves are sealed, with no learning across seasons. The robustness report describes threshold-crossing and utilization heuristics plus one-year post selection, not strategic equilibrium search. | manuscript.html §4.3, Appendix B; constants.mjs: BOT_POLICY_IDS; REPORT.md, Results |
| Frontier access and a fixed catalog | Twelve recurring tasks on six capability rungs, thresholds 1, 2, 4, 8, 16, 32. Only sufficiently capable labs can serve each task. A frontier rung opens both dear and cheap tasks; ceiling levels are 40, 16, 8, 4, 2, 1 Gold per work unit. | manuscript.html §3.3, Appendix B |
| Price elasticity as ladder shape | PED is the magnitude of a fitted log cumulative affordable-work versus log price slope across six ceiling levels. Tested labels are approximately 0.88, 0.50, 0.35, 0.25, 0.18. This measures the configured ceiling distribution, not a fitted dynamic buyer response to each lab's post. | manuscript.html §4.1 |
| Intelligence elasticity as threshold spacing | IED is d ln W / d ln X, with W cumulative recurring work serviceable at capability X. The experiment varies threshold spacing. Work-weighted and ceiling-value-weighted measures differ: canonical fitted slopes are approximately 1.23 and 1.55. Grid labels are fitted summaries of six discrete rungs. | manuscript.html §4.2, Appendix B |
| Demand grid, scale, and dispersion | The study changes externally fixed threshold spacing, ceiling shape, mean volume scale, and volume dispersion. Per eta: 90 on-course demand cells; a six-by-five elasticity grid; a three-by-three-by-three interaction cube; a six-by-five scale/dispersion grid; a canonical baseline. These are experiment settings, not player controls. | manuscript.html §4.3 and Table 2; §§5, 6.2 |
| Once-revealed demand | Each task receives one bounded seeded volume draw, persistent for the season. Volumes become public before that year's actions are sealed when earlier commitments make tasks reachable. These are uncertain levels across worlds, not fresh annual demand shocks. Work equals jobs times task-specific compute per job. | manuscript.html §§3.3–3.4; constants.mjs: DEMAND_ROLL_SEMANTICS |
| Dispatch and allocation | Tasks are processed by descending ceiling, descending capability threshold, then catalog identifier. Work is split equally among qualified, willing labs, with deterministic redistribution when capacity runs out. | manuscript.html §3.5, §7 Mechanism, Appendix B |
| Scarcity pay and the going rate | Task/lab pay per work unit = post + scarcity sigma × (ceiling − post). Scarcity sigma = clamp((2 × realized work − installed qualified centers) / realized work, 0, 1). It counts installed qualified fleet, including training allocation. The going rate is afterward reported as a served-work-weighted mean post; it is not one compulsory payment or universal clearing floor. | manuscript.html §3.5, Appendix B |
| Refusal-induced idle training credit | Eta credits only refusal-induced idle serving capacity toward training; ordinary unused capacity gets no such credit. Main eta = 0.3; sensitivity eta = 0.8. Demand-dispersion sigma is separate from scarcity sigma. | manuscript.html §§3.5, 4.3; constants.mjs: LIVE_ETA |
| Fast catch-up | The Blueprint tracks opening frontier capability with a one-year lag. Training below it is four times as efficient until the gap closes. This is a production-side knowledge spillover; it does not give every lab the frontier model for free. | manuscript.html §3.6; constants.mjs: CATCHUP_EFFICIENCY |
| Build market and rationing | Four lots are available; demand raises unit build price from 1 to a cap of 2 Gold. Over-orders are filled proportionally. The plan describes price as base / (1 − min(total requested levels / 4, 0.5)). Supply quantity and pricing rules are fixed; lab orders are endogenous. | manuscript.html §3.2, Appendix B; constants.mjs: BUILD_SUPPLY_LEVELS, BUILD_BASE_PRICE, BUILD_MAX_PRICE, BUILD_DEMAND_CAP; PLAN-V3-AGENT-SIMULATION.md, Design (c) |
| Upkeep and depreciation | Every installed center costs 1 Gold annually and depreciates at 15% annually. Neither rate responds to demand or technology vintage. Training versus serving does not remove the installed-fleet operating bill. | manuscript.html §3.2, Appendix B; constants.mjs: OPERATING_COST_PER_CENTER, DEPRECIATION_RATE |
| Solvency reserve | MUST COVER constrains affordable builds using the operating bill and worst-case requested build cost. The manuscript calls this next year's operating cost; the robustness report describes the implementation using current fleet. A fresh resolver audit is outside this packet. | manuscript.html §3.2; constants.mjs: OPERATING_COST_PER_CENTER, BUILD_MAX_PRICE; REPORT.md, mechanism description |
| Horizon and terminal recovery | Standard play and the study last twelve years; terminal recovery is 1 Gold per remaining center. An eight-year guided setting exists in constants but is not the published study. Terminal Gold includes recovery. | manuscript.html §§3.1, 3.7, Appendix B; constants.mjs: MODES, TERMINAL_RECOVERY_PER_CENTER |
| Payback and accounting | Company payback uses the 1,000-Gold starting balance; industry payback uses 3,000 Gold. Revenue coverage is cumulative industry revenue / 3,000, not terminal wealth. Payback is undiscounted; company success and aggregate success are different endpoints. | manuscript.html §§3.7, 5, 6.3, 8 Payback; constants.mjs: STARTING_GOLD |
Bucket 3 — left out
What the model removes or never represents, because modeling it would test a different, larger question than payback under this fence. The last four rows are operator-named, not inferred from the manuscript.
| Reality | How it appears in the model | Source section |
|---|---|---|
| More than three competitors; entry and exit markets [inferred] | The roster is fixed at three. Insolvency exists, but entry, acquisition, and an evolving population of competitors are not modeled. | manuscript.html §§3.1, 3.7, §7 Toy economy; constants.mjs: LAB_IDS |
| Open-weights models and non-rival use [inferred] | No separate open-weights provider, licensing regime, or shared model-use channel is specified. Blueprint catch-up remains present and must not be described as zero spillover. | manuscript.html §§3.1, 3.6–3.7 |
| Vertical integration into applications [inferred] | No application business, internal transfer price, distribution channel, or bundled downstream margin is represented. Labs directly serve catalog work. | manuscript.html §§3.1, 3.3–3.5 |
| Vertical integration into hyperscalers or neoclouds [inferred] | Labs buy centers through one build mechanism; there are no distinct infrastructure owners, compute-leasing businesses, or cross-layer ownership decisions. | manuscript.html §§3.1–3.2, §6.3 |
| Financing and capital markets | No equity raises, debt, refinancing, investor behavior, or funding-market feedback. Labs live on their starting troves and operating receipts. | manuscript.html §3.1, §7 No funding market |
| Supplier accounts and endogenous supplier investment | Build payments exist; supplier costs, profits, receivables, balance sheets, default, and losses do not. Current supply is a fixed rule. The v3 supplier proposal is a new experimental arm. | manuscript.html §6.3, §7 Supplier accounts absent; constants.mjs: BUILD_SUPPLY_LEVELS |
| Energy systems [inferred] | The flat center operating charge contains no explicit electricity supply, power price, grid bottleneck, or energy investment sector. | manuscript.html §3.2, Appendix B; constants.mjs: OPERATING_COST_PER_CENTER |
| Labor and household income [inferred] | No worker, wage, displacement, household budget, or income-feedback account; customers are exogenous task demand. | manuscript.html §§3.1, 3.3–3.4 |
| Regulation and taxation [inferred] | No regulator, tax schedule, antitrust action, licensing restriction, or public-policy choice is represented. | manuscript.html §§3.1–3.7; constants.mjs |
| Chosen price discrimination [inferred] | One lab post covers the year; no customer-specific tariff, subscription, negotiation, or two-part tariff is a decision. Actual pay can nevertheless differ by task and lab through scarcity premiums and different posts. | manuscript.html §3.5, §7 Mechanism |
| Endogenous task invention and buyer demand feedback | Arrival thresholds, ceilings, and recurring volumes are external inputs. Investment makes catalog work reachable; it does not invent tasks or change customers' willingness to pay. The operator explicitly holds end-market demand exogenous. | manuscript.html §§1, 3.3–3.4, 4.2 |
| Adoption and complementary investment [inferred] | No separate adoption ramp or customer reorganization investment appears after work becomes serviceable; adoption lag is not varied. | manuscript.html §§3.4–3.5, §7 Other omissions |
| Within-season demand shocks [inferred] | The same task's demand does not get a new draw every year. Cross-world volume variation cannot establish a response to an unexpected annual boom or bust. | manuscript.html §3.4; constants.mjs: DEMAND_ROLL_SEMANTICS |
| Endogenous obsolescence and alternative progress laws | Depreciation and training-cost geometry are not swept; no technology-vintage obsolescence schedule is specified [inferred]. Residual free progress is not varied. Catch-up and purchased training remain present. | manuscript.html §3.6, §7 Other omissions; constants.mjs: DEPRECIATION_RATE, TRAINING_DIFFICULTY |
| Strategic learning and equilibrium identification | Deterministic rivals have fixed procedures, not learning across seasons. Tested policies and finite cells cannot establish equilibrium or universal strategy dominance. | manuscript.html §§4.3, 7 Toy economy and Finite grids |
| Time discounting [inferred] | Terminal Gold is compared with initial Gold without a discount factor. No present-value hurdle or risk-adjusted cost of capital appears in that definition. | manuscript.html §§3.7, 8 Payback; constants.mjs |
| Total welfare | Buyer headroom measures ceiling minus payment on served work, not total welfare; it excludes unserved value. Private lab payback is not a social-payback measure. | manuscript.html §6.3, §7 Customer value proxy |
| Unbounded external validity | Twelve tasks, one studied horizon, finite demand cells, and sampled policies omit other possible economies. Zero recorded no-capable work is only a statement about recorded tasks. Cutoffs are selected in-sample; no held-out evaluation establishes transfer. | manuscript.html §7 Toy economy, Finite grids, Recorded task scope, Descriptive diagnostics, Other omissions |
| Marketing strategies, including subsidizing tokens for personal use | Not modeled. No promotional pricing, loss-leader subsidy, or below-cost customer-acquisition spend exists anywhere in the clearing or pricing rule. | operator-named omission; not in manuscript.html |
| Advertising markets | Not modeled. Labs earn only served-work revenue at the posted price; there is no ad-supported free tier and no separate advertising revenue line. | operator-named omission; not in manuscript.html |
| Confidence swings in markets and willingness to cover losses | Not modeled. There is no funding market (manuscript.html §3.7), so no investor sentiment, patience, or willingness to keep funding an unprofitable lab is represented; a lab either covers itself from operations and its starting trove, or is insolvent. | operator-named omission; not in manuscript.html |
| Politics, including data-center pushback | Not modeled. No permitting delay, community opposition, siting fight, grid-interconnection queue, or other political friction affects the build market or the operating bill. | operator-named omission; not in manuscript.html |
42 rows total: 4 kept and played, 16 kept and fixed, 22 left out. Every share and percentage elsewhere on this site is conditional on this fence — see manuscript.html §7.5 for the two omissions that matter most when reading the results.