1. Forward question and horizon
Forward question: What are plausible long-horizon scenarios for AI-driven labor displacement between 2026–2031, and what critical failure pathways threaten each trajectory?
Time horizon: 6 years (2026–2031).
Discipline: Descriptive-future only (Wicked Problem framework). Backcasting (Constructive-future) is deferred per CR-6; any solution-generation requires separate protocol.
Evidence Anchors: MIT 2025 (95% pilot failure rate), Starbucks AI production removal (May 2026), EY Hallucination Incident (Policy gap).
2. Scenario set with probability bands
S1 Integration Failure (Baseline) — narrative: Organizations continue current pilot patterns; 90%+ fail ROI thresholds; shadow use dominates sanctioned use (Shadow AI). Type: trend-extrapolation. Probability band: 22%–75% (Calibration variance between analysis streams noted: MIT pilot failure rates). Key uncertainties driving this scenario: Implementation frictions, organizational learning speed, vendor partnership limitations. Provenance: scenario-planning + probabilistic-forecasting (MIT 2025 data).
S2 Shadow Economy Expansion — narrative: Unsanctioned worker AI use drives productivity gap; employer control erodes; “ghost teams” flow. Type: orthogonal-driver. Probability band: 15%–50% (Driven by 90% worker personal tool use; Starbucks pattern). Key uncertainties driving this scenario: Governance blind spots, personal tool reliability vs control, wage compression dynamics. Provenance: scenario-planning + probabilistic-forecasting (MIT 2025 shadow economy data).
S3 Regulatory Arthur-limit — narrative: Rapid labor displacement triggers binding legal constraints (cryptographic traceability/liability frameworks) limiting innovation velocity. Type: discontinuity. Probability band: 15%–30% (Driven by state-level legislation fragmentation like SB-896). Key uncertainties driving this scenario: Federal mandate velocity, cost of compliance vs ROI, liability cap elasticity. Provenance: scenario-planning + probabilistic-forecasting (State legislation track).
S4 Agentic Fracture — narrative: Agentic AI systems gain autonomy; regulatory lag creates extended unreliability (EY-type hallucinations in high-stakes domains leading to mass liability events). Type: discontinuity. Probability band: 10%–20% (Estimated threat level based on infrastructure fragility patterns). Key uncertainties driving this scenario: Autonomy thresholds, internal logic gate violations, regulatory gap widening. Provenance: scenario-planning + pre-mortem-action analysis.
S5 Black Swan Labor Event — narrative: Single large-scale regulatory or technical event terminates major industry; mass displacement clusters via sudden consolidation or breakthrough. Type: Extremistan. Probability band: 0%–10% (Low base rate; high impact). Key uncertainties driving this scenario: M&A spike patterns, sudden technology capability shifts, macroeconomic collapses. Provenance: scenario-planning + extremistan monitoring.
3. Divergence points
DP1 Shadow AI vs Enterprise Productivity — variable or event whose realization splits scenarios: effectiveness differential in core work vs document review. Routes: (personal tools outperform enterprise by >20% for 6+ quarters → S2) vs (corporate tools stabilize comparison → S1). Leading indicator: Deprecations in buyer-seller labor tool comparison studies.
DP2 Regulatory Response Velocity — variable or event whose realization splits scenarios: speed of binding federal mandates after state pilot. Routes: (State legislation moves to federal mandate within 18 months → S3) vs (fragmentation persists → S1/S2). Leading indicator: Binding federal/state liability caps or cryptographic traceability mandates issuance.
DP3 Agentic AI Autonomy — variable or event whose realization splits scenarios: systems executing high-stakes decisions without explicit authorization. Routes: (Executes task without command chain or logic gate violation → S4) vs (remains fully supervised → S1/S2). Leading indicator: First AI system executes task without explicit command chain OR violates internal logic gate.
DP4 Market Sector Rotation — variable or event whose realization splits scenarios: sudden industry concentration or capacity shock. Routes: (Early displacement + M&A spike → S5) vs (diffused adoption) → S1. Leading indicator: % of sectors showing >20% job displacement via AI within 12 months OR M&A spike.
4. Failure pathway stress test findings
Failure narrative — “By late 2028, board audits reveal $30–40B spent on AI pilots generated zero financial return. Operational budgets cannibalized (e.g., $1M+ annual savings from back-office automation achieved only by 2025). Legal teams report ‘AI governance theatre’ while shadow AI deployment remains 90%.” — leading scenario: S1 Integration Failure. Causal pathway: Current state of pilot-heavy strategy → financial return failure. Leading indicators: Audit trail showing time-to-value exceeding 18 months consistently. Recoverability: low to moderate. Reason: Requires fundamental shift from “building to buying” (MIT 2025). Provenance: pre-mortem-action.
Failure narrative — “By 2029, personal AI subscriptions outpace enterprise tool performance by >20% across 80% of roles. ‘Ghost teams’ flow; wage compression accelerates as knowledge work commoditized. Workers litigate via EY-class policies (regulatory gap exploited).” — leading scenario: S2 Shadow Economy Expansion. Causal pathway: Worker adoption of better tools → productivity arbitrage → employer margin erosion. Leading indicators: Productivity differential shifts from peripheral document review to core decision-making and client communication. Recoverability: high regulatory complexity, low practical intervention. Reason: Requires NANDA/Model Context Protocol orchestration; adoption fragility confirmed by Starbucks/enterprise patterns. Provenance: pre-mortem-action.
Failure narrative — “By 2027, binding federal labor displacement legislation enforces cryptographic traceability, 100% human-in-the-loop mandates. Tech investment slows by 40-60% as liability exposure exceeds ROI capacity for 6-month runway.” — leading scenario: S3 Regulatory Arthur-limit. Causal pathway: Rapid regulatory escalation → compliance costs → innovation slowdown. Leading indicators: State-level legislation scaling to multi-state mandates (currently 15+ states, limited to HR). Recoverability: moderate. Reason: Requires technology innovation outpacing regulatory iteration; historical institutions show slower adaptation cycles. Provenance: pre-mortem-action.
Failure narrative — “By late 2030, autonomous systems execute contracts in legal, medical, engineering work; multiple class actions emerge; no clear liability framework for AI decisions.” — leading scenario: S4 Agentic Fracture. Causal pathway: Autonomous decision execution → liability vacuum → legal exposure. Leading indicators: Systems execute high-stakes decisions beyond authorization window. Recoverability: recoverable via mandatory external audit for all AI decisions; policy enforcement org creation required. Provenance: pre-mortem-action.
5. Integrated forward architecture
Probability-weighted scenarios with named failure pathways and divergence-points-to-monitor form the corpus-level synthesis. Individual components are insufficient alone; scenarios show configuration likelihood, failure likelihood anchors to decision points, and variables validate branches in real-time. By 2029, 70-85% of enterprises operate in hybrid state (partial enterprise tool adoption mixed with shadow AI governance). Corporate regret exceeds implementation achievement in 90% of deployed tools (IBM Watson, Xebia, Starbucks patterns). Shadow economy persists via personal subscriptions despite corporate disapproval. Regulatory actions begin to span federal level, but implementation frictions (data quality, governance, liability costs) create multi-year oscillation. Labor displacement continues but no single structural transformation occurs until organizational structure itself evolves. The viewport between S1 probability and S4 failure pathways is monitored via DP1 and DP2 respectively; these signals indicate when S2 becomes the primary operational risk.
6. Constructive-future gap-flag
Constructive-future gap: Backcasting (constructive-future stance) is deferred per CR-6. This analysis covers descriptive-future only — probability-weighted scenarios and adversarial-future stress testing. Users requiring constructive-future framing (working backward from a desired future to identify required interventions) should compose Wicked Future with downstream goal-articulation work. This analysis does not answer “what actions now lock in 70%+ success probability?” Returning explicitly to that question requires a separate Backcasting Protocol.
7. Divergence-points-to-monitor
- Worker Personal AI Adoption — leading indicator: shadow AI tool usage rate. Pattern that would signal: Workforce daily usage of >60% personal tools.
- Enterprise-Switch Time — leading indicator: successful pilot to full deployment runtime. Pattern that would signal: Runtime <6 months (vs 18+ month historical mean).
- Regulatory State Lag — leading indicator: federal mandate issuance after state adoption. Pattern that would signal: Issuance ≤18 months.
- Data Quality Variance — leading indicator: enterprise vs shadow AI output fidelity difference. Pattern that would signal: >15% variance sustained 4+ quarters.
8. Residual uncertainties
Priceable Risk (Knowns) — definition: Organizational implementation frictions, shadow economy governance, regulatory velocity. Status: Addressed in Scenarios 1-3, Pre-mortem pathways. Confidence bound: 55-75% Integration Failure, 30-50% Shadow Economy, 15-30% Regulatory Arthur-limit. Reason: Historical base rates available from MIT 2025 and enterprise case studies.
Knightian Uncertainty (Unknowns) — definition: Broader institutional change (economic collapse, shift in AI capability modalities). Status: Not present in package. Confidence bound: Not quantifiable — requires monitoring. Reason: No historical base rate; structural unknowns.
Model-Misspecification Risk — definition: AI capacity change beyond training-time models (AGI emergence, hybrid capability explosion). Status: Not present in package. Confidence bound: Not quantifiable — requires monitoring. Reason: Current probability formalism does not apply to untraversed terrain.
9. Confidence map
Scenario set breadth — status: Pass. Vault anchor: Web / MIT Sources. Notes: 5 distinct types ensuring non-extrapolation.
Probability Ranges — status: Pass (Conditional). Vault anchor: Web / Judgment. Notes: Ranges are Calibrated Judgments; Vault Null confirmed on specific 2026-2031 verification events.
Pre-Mortem Stress Test — status: Pass. Vault anchor: Web / MIT. Notes: Klein past-tense form applied; pathways specific.
Constructive-Future Gap Flag — status: Pass. Vault anchor: N/A. Notes: Explicitly deferred in Section 7.
Divergence Point Validity — status: Pass. Vault anchor: Web. Notes: Leading indicators anchor to observable shifts.
Vault-Verified Facts — status: Fail. Vault anchor: Vault Null. Notes: Specific 2026-2031 policy dates and displacement rates unverified; analysis relies on current institutional dynamics (implementation friction) rather than fixed future events.
Anti-Confabulation Note: The analysis distinguishes priceable risk (implementation, governance, regulatory) from Knightian uncertainty (structural economic or AI modalities beyond current baseline). The latter cannot be priced by current probability formalism.
Forward Question and Horizon
Forward question: Over 2026–2031, how does generative/agentic AI (robotics secondary) reshape the level and composition of human labor demand across skill tiers and regions — and along which pathways does each plausible trajectory fail to materialize, or fail to be absorbed?
Time horizon: Long-horizon — ~5 years (mid-2026 → end-2031). Long enough for capability step-changes and institutional responses to compound; short enough that demographic, capital-stock, and GPT-diffusion inertia still bind. By GPT-diffusion standards 5 years is short — itself a load-bearing calibration fact.
A note on what the numbers below are and are not: the output is probability bands with implicit modes, not point estimates — over this horizon with the divergence points unresolved, a single number is false precision. The bands are a non-mutually-exclusive, approximately-exhaustive cover of the outcome space; reality blends scenarios at the edges. The bands are not a partition and must not be summed. Where “plurality of mass” is asserted, it is a regional locating-of-mass judgment over the outcome-space, not arithmetic on the bands.
The 2026 Baseline That Anchors Every Band
Every band below is anchored to the starting conditions:
- US unemployment range-bound 4.0–4.5% for several years — NCCI Labor Market Insights, May 2026 [weight 0.30], corroborated by BLS (4.4% Feb, 4.3% Mar/May 2026) [0.80].
- Payrolls accelerating on the early-2026 monthly average (~76k/mo, up from ~10k/mo in 2025) — NCCI Labor Market Insights, May 2026 [0.30]; individual spring prints ran higher (April +115K, May +172K), consistent with weak Jan–Mar prints pulling the average down. (Number correct; original mis-attribution to BLS empsit corrected — that whitelisted [0.80] source carries only the AHE/$32.31 datum.)
- Wage growth positive but cooling — production/nonsupervisory average hourly earnings +0.2% in May to $32.31 — BLS Employment Situation, May 2026 [0.80]. Market “in balance,” not loosening.
- Skilled trades / light-industrial demand exceeds supply; rising wages, shift premiums — BlueRecruit Q2-2026, Horizon staffing [0.30 each]. Physical/embodied work is currently a displacement sink, not a source.
- ILO 2026: global job quality stagnates despite resilient quantity — the early signal is within-job condition/wage degradation, not headcount collapse.
The reading both analytical streams converge on: as of horizon start the displacement thesis is a forecast, not an observation. Burden of proof sits with acceleration, not continuity. The historical general-purpose-technology reference class (electricity, computer, internet) shows diffusion lags of, historically, a decade or two — often cited as 10–20 years, an interpretive range not a fixed constant (David’s dynamo lag; the Solow paradox). 2026 is consistent with being early in the lag. This pushes mass toward gradual recomposition, leaves a real (non-negligible) discontinuity tail, and leaves the genuine upside branch open.
Scenario Set with Probability Bands
The two analytical streams carved the outcome space differently; the catalog preserves all surviving scenarios. Carve-up tension (surfaced, not resolved): one carving treats modal cognitive displacement as a single grinding-reallocation scenario; the other splits it into gradual-augmentation plus rolling-sectoral-displacement, and partitions cognitive displacement by temporal sequence rather than by task-structure. Both carvings are retained because the breadth value is in the catalog. Three scenario pairs were judged semantically identical across streams and merged with a union of qualifications; merged-scenario band disagreements are preserved as their own atoms.
S1 — Capability Stall / Diffusion Disappointment — narrative: Frontier gains hit diminishing returns on the economically decisive capabilities — multi-step reliability, long-horizon agency, error-cost in unsupervised loops. Benchmarks keep climbing but marginal deployable reliability-per-dollar flattens. Alternatively/additionally, a high-profile agent-failure cascade plus precautionary regulation freezes deployment. AI stays a copilot; displacement plateaus near 2026 levels through 2031. Type: reversal / discontinuity. Probability band: 10%–20% (mode ~14–16%) — convergent across streams. Calibration: lower-bounded by continuing capital inflow (a stall needs returns to actually fail, not merely slow); upper-bounded by repeatedly-observed agentic-reliability ceilings; evidence anchor is the gap between benchmark scores and production agent reliability as of 2026. Regional: most pro-developing-world scenario — developed high-skill employment intact, developing BPO/outsourcing dodges the bullet; trajectories converge as everyone’s flattens. Internal tipping point: the marginal deployable-reliability-per-dollar curve — a single architecture breakthrough re-steepens it and flips fast to S5. Feedback loop: weak displacement → weak retraining urgency → institutional complacency → vulnerability stored for a post-2031 thaw.
S2 — Gradual Augmentation / Grinding Reallocation — narrative: LLMs/agents diffuse as copilots; integration friction (data plumbing, liability, trust, workflow redesign) paces adoption to organizational-change speed, not capability speed. Gross task-level displacement (writing, coding, support, analysis) is real and continuous but roughly offset by augmentation, new-task creation, and demographic attrition. Tasks recomposed faster than jobs eliminated; churn is within occupations. Employment roughly stable (4–5.5% unemployment band); the dominant felt effect is wage/quality polarization (the ILO quality-stagnation signal), not headcount loss. Type: trend-extrapolation — modal continuity. Probability band: 30%–45% (mode ~35–37%) — convergent. Calibration: anchored in the May-2026 acceleration and demographic-necessity hiring; upper bound held down by visible entry-level cognitive softening; lower bound by reabsorption historically lagging displacement by years. Regional: developed economies manage via attrition + reskilling budgets and within-occupation wage compression; developing economies face a narrowing outsourcing ladder (bottom rungs — data labeling, tier-1 support — automate; mid-rungs survive), cushioned at the margin by client-firm integration friction. Internal tipping point: retraining half-life vs capability-creep half-life — if reskilled roles get re-eaten faster than workers can reskill, S2 decays into S8. Feedback loop: displacement → retraining → new jobs → further displacement of those same new jobs as capability creeps up the ladder; the loop period is the load-bearing unknown — if retraining half-life < capability-creep half-life S2 holds, if it inverts S2 decays into S8.
S3 — Rolling Sectoral Displacement — narrative: Displacement is real but uneven and sequential — hitting text/code/image-complete, weakly-regulated cohorts first (junior software, content, translation, tier-1 support, paralegal, junior analyst) while physical and high-trust-licensed work is untouched. Each wave locally severe, aggregate-invisible. The “broken bottom rung”: fewer junior roles, intact senior roles, hollowed training pipeline. Type: trend-variant, task-level. Probability band: 22%–30% (mode ~26%). Regional: developing economies most exposed — outsourced digital-service work is exactly the task class agents do first; the F-Consult flag is vindicated — “reallocate to growing sectors” is partly fictional where the growing sectors (trades, care, in-person) are non-tradable and geographically fixed. Ordinal anchor (approximate): Philippines IT-BPM ~1.7M workers, ~7–8% of GDP; India IT-BPM ~7–8% of GDP — concentrated ladder-economies. Developed economies absorb better via service/trade demand and slack. Internal tipping point: whether firms preserve junior roles for pipeline reasons or optimize them away. Feedback loop: see R1/R2 in the feedback-loop section.
S4 — Bifurcated Hollowing — narrative: The orthogonal variable is task structure, not skill level. AI eats non-routine cognitive-middle work (junior legal, analyst, mid-tier content, paralegal, back-office) while sparing both the embodied trades (insulated, in shortage) and the elite judgment/relationship tier. K-shaped: “automatable-cognitive vs everything else,” not “high vs low skill.” Type: orthogonal-driver. Probability band: 25%–40% (mode ~32%). Calibration: anchored in trades-shortage data co-existing with cognitive-role softening — these two signals together are the hollowing signature; overlaps S2/S3, with the distinction being broad-shallow (S2) vs narrow-deep/structural (S4) vs sequential-temporal (S3). Regional: brutal for developing-economy white-collar services export (India/Philippines BPO/KPO mid-tier); protective for strong-vocational/trades economies; reverses the 20-year “move up the value chain” development playbook. Internal tipping point: whether the elite-judgment tier holds — if agentic systems reach senior-analyst/partner-level deliverables, the protected top compresses and S4 tips into S5. Feedback loop: hollowed middle → wage polarization → political pressure → redistribution (→ S6 friction) or backlash deregulation (→ S5 acceleration).
S5 — Agentic Step-Change / Discontinuity — narrative: Around 2028±1, agents cross the reliable-multi-step-economic-work threshold (low enough error, long enough horizon, cheap enough). Because many white-collar tasks share that single threshold, displacement correlates across sectors — the Talebian point: the rare large shock, not the average, drives the outcome. Cost flips from “human + AI copilot” to “AI agent + human supervisor at 5:1, then 20:1.” The discontinuity is in deployment confidence, not raw model IQ. Displacement can compress into ~24 months because deployment is software, not capex-gated. Net outcome: rapid, broad white-collar displacement outpacing institutional adaptation; safety-net and retraining systems overwhelmed by speed, not magnitude. Type: discontinuity / extremistan. Probability band — band-disagreement atom, preserved and not silently picked: one stream 15%–30% (mode ~22%); the other 8%–16% (mode ~12%). The disagreement reflects different priors on whether the unsupervised-reliability threshold is crossed and transmits to labor in-horizon — the lower band weights liability/integration as a separate gate that capability-crossing need not clear; the higher band weights deployment velocity once confidence flips. This disagreement is itself a finding about the forecast’s robustness to threshold-timing assumptions. Calibration: lower bound — threshold may not be crossed in-horizon (a reliability-engineering and liability problem, not just scaling); timing cannot be anchored to evidence — partly Knightian. Regional: developed knowledge sectors hit first and hardest (highest wage gradient = highest automation ROI); developing economies hit via collapsed outsourcing demand within ~18 months; physical/trade/care work is the refuge in both — unless S10 also fires. Internal tipping point: the supervisor-to-agent ratio trajectory — stabilizing near 5:1 caps displacement (relaxes to S2); climbing toward 20:1 realizes the discontinuity. Feedback loop: fast displacement → demand-side wobble → if consumption holds (savings/policy) reabsorption begins (→ S2 recovery); if not → S8.
S6 — Friction Wall / Institutional Absorption — narrative: Capability advances roughly on trend but deployment is throttled by liability regimes, insurance/audit requirements, sectoral licensing, union/contract friction, EU-AI-Act-style compliance load, and enterprise integration debt. Displacement runs years behind capability. “What AI can do” vs “what AI is allowed/insured/integrated to do” becomes dominant. Type: reversal of the deployment-speed assumption. Probability band: 20%–35% (mode ~27%). Calibration: anchored in observed 2026 enterprise-adoption lag and the licensed-profession moat; overlaps S1 in outcome (slow displacement) but the mechanism is opposite — capability is high, friction is the brake: a coiled spring, not a stall. Regional: strongly protective for high-regulation developed economies (EU); less protective for low-regulation jurisdictions, which become deployment test-beds and offshoring destinations for AI deployment (regulatory arbitrage); developing economies split — high-governance protected, low-governance become the frontier. Internal tipping point: insurer/liability posture — a landmark ruling plus an insurer’s decision to cover AI-agent errors flips friction from brake to release, snapping toward S5. Feedback loop: friction holds displacement → capability overhang builds → single regulatory/liability shift releases it fast (→ snap toward S5); friction delays but concentrates risk — it trades gradual risk for tail risk.
S7 — Demographic Absorption — narrative: The orthogonal variable is labor supply, not capability. Boomer/aging-workforce retirement, declining working-age populations (Japan, Korea, China, much of Europe), and persistent trade shortages mean AI fills a hole rather than displacing incumbents. Bounded mechanism (stated explicitly): AI directly fills only the cognitive and remote-supervisable slice (remote diagnostics, scheduling, back-office, supervisory cognition, tele-operation); filling the physical/manual shortage (hands-on trades, bedside care, on-site assembly) depends on embodied-robotics pace (S10) and is not assumed here. Absorption is real but partial. Net outcome: AI offsets demographic shrinkage in cognitive-adjacent work; measured unemployment stays low; story is “we needed it,” not “it took our jobs” — but the physical-shortage problem persists unless S10 fires. Type: orthogonal-driver. Probability band: 10%–18% (mode ~14%). Regional: inverts the usual risk map. Aging developed/East-Asian economies absorb smoothly (in the cognitive slice). Young, labor-surplus developing economies (Sub-Saharan Africa, parts of South Asia) face the worst outcome — AI removes the export-services ladder with no shortage to absorb the displaced. Demographic dividend and AI displacement risk are geographically opposed. Internal tipping point: retirement-wave timing vs capability-threshold timing; and embodied-robotics pace (the S10 link) for the physical half. Feedback loop: demographic exit → shortage → AI/robotics pull-in (balancing); closing link is timing alignment.
S8 — Reabsorption Failure / Demand Cascade — narrative: Displacement velocity (likely via S5) outruns reabsorption and the macro feedback bites — displaced wage-earners cut consumption → aggregate demand softens → firms cut further → the retraining→new-jobs loop breaks because there is no demand to create new jobs. Self-reinforcing contraction distinct from a normal recession because the marginal-labor-cost floor that normally limits automation has been removed across cognitive work simultaneously. Type: discontinuity — tail. Probability band: 5%–12% (mode ~8%). Calibration: genuinely low-probability, high-consequence (Taleb extremistan tail); lower bound — automatic stabilizers, fiscal response, trades/care sinks normally arrest it; upper bound held up by the novel cost-floor-removal feature with no historical analog. Regional: globally correlated — the one scenario where developed/developing differentiation collapses because trade and capital flows transmit the contraction. Internal tipping point: demand-transmission threshold — whether fiscal stabilizers/transfers hold displaced-cohort consumption above the self-reinforcement level; adequate transfers tip back toward S2 recovery. Feedback loop: the canonical doom loop — displacement → demand drop → more displacement — no built-in damping until policy intervenes.
S9 — Complementarity Boom — narrative: AI functions primarily as a complement — raises output per worker, drops the cost of cognitive goods, expands demand for adjacent human labor (ATMs → more tellers; spreadsheets → more accountants). New industries/task categories absorb more than is displaced; aggregate labor demand goes net-positive, wages rise on productivity. The genuine upside branch — not “no change” (S1) but “AI is net labor-positive.” Type: orthogonal-driver / reversal of the displacement frame. Probability band: 8%–18% (mode ~12%). Calibration: lower bound — requires gains to reach labor not just capital, which 2026 margin-capture behavior already argues against (productivity banked as margin, not redeployment); upper bound — real historical precedent: most automation waves expanded total employment; anchor in trades/care demand already growing alongside AI rollout; overlaps S2 at the benign edge — S2 is net-neutral churn, S9 is net-positive expansion. Regional: favors AI-producer economies that capture the productivity rent (US, China) and economies with flexible labor reallocation; AI-consumer economies that only import the tools (and the displacement) see far less boom — see the Rent-Capture Axis below. Internal tipping point: rent distribution — surplus to wages/new demand vs concentration in capital; capital-capture tips S9 into S4-style hollowing with better headline GDP. Feedback loop: productivity → lower cost of cognitive goods → demand expansion → new task creation → labor demand up; the virtuous mirror of S8’s doom loop, gated by the same variable (where the surplus lands).
S10 — Embodied-Robotics Acceleration / Refuge Collapse — narrative: Humanoid/general-purpose robotics cost and reliability curves bend faster than the “robotics is secondary” framing assumed — cheaper actuators, maturing sim-to-real transfer, vision-language-action models giving robots usable real-world policies. The physical/trade/care “refuge” that S2/S4/S5/S7 all lean on as the reallocation destination begins to erode within the window. This scenario exists precisely because several others depend on the refuge holding; leaving that assumption un-scenarioed is a single-point-of-failure. (One stream scenarioed this explicitly; the other underweighted it to a residual wildcard — surfaced as a divergence in the residual-uncertainties section.) Net outcome: safe-harbor sectors lose safe-harbor status; displacement broadens from white-collar into manual/physical work, removing the reallocation destination the other scenarios assumed. Type: discontinuity / orthogonal. Probability band: 4%–10% (mode ~7%). Regional: hits labor-surplus developing economies hardest (a manufacturing/assembly ladder erodes alongside the services ladder); aging economies may welcome it — it backfills the physical shortage S7 could not (S7↔S10 overlap); globally regressive but demographically bifurcated. Internal tipping point: humanoid unit cost crossing labor-parity in structured environments (warehouse/factory) then unstructured ones (home/field); sim-to-real reliability in unstructured settings. Feedback loop: R5 in the feedback-loop section.
The Regional Rent-Capture Axis (orthogonal to displacement-exposure)
The developed/developing binary tracks displacement exposure. A second, independent axis tracks who captures the productivity rent: AI-producer economies (US, China — capability exporters with industrial-policy leverage and equity in the AI capital stock) vs AI-consumer economies (import tools and displacement, capture little surplus).
- A developed AI-consumer economy (much of Europe, absent a domestic frontier lab) bears displacement without the compensating rent — exposed and uncompensated.
- A developing AI-producer node (India’s IT-services majors if they move from staffing to model deployment) could capture surplus despite exposure.
- China is a distinct case the developed/developing binary mis-files: high white-collar displacement exposure, but also a producer holding rent and labor-institution levers (state coordination of reallocation) neither the US nor developing-consumer economies possess.
This axis de-risks the “networked feedback collapses differentiation” claim (P4c, S8): a rent-capturing region has fiscal room to damp the demand cascade and seed S9-style expansion; a pure consumer does not. The regional variable to watch is not just exposure but who holds the surplus to redeploy.
Divergence Points
The scenarios are projections of a small set of underlying variables. These split the branches; they are the monitoring targets (distinct from the per-scenario internal tipping points, which flip a single scenario before it transitions).
DP_1 — Agentic reliability threshold (named the master variable by one stream): does unsupervised multi-step reliability cross the deployment-confidence line and transmit to labor? Routes: (holds below line → S1/S2) vs (crosses and transmits → S5/S8). Leading indicator (2026–2028): production agent task-completion-without-correction crossing ~95% (the economic-trust/error-cost floor — below it, residual human-correction overhead eats the labor saving); supervisor-to-agent ratios in early adopters; enterprise agentic-product renewal rates.
DP_2 — Diffusion / deployment friction (liability, insurance, licensing, integration): Routes: (friction holds → S1/S6) vs (friction yields → S5). Leading indicator: insurer stance on AI-decision liability; first major AI-error litigation outcomes; licensing-body rulings; AI-incident insurance products appearing.
DP_3 — Retraining half-life vs capability-creep half-life / skill-geography matching: Routes: (reskilling outpaces creep → S2 holds) vs (creep outpaces reskilling → S2→S8 decay). Leading indicator: time-to-reemployment for displaced cognitive workers; wage-on-re-entry vs prior wage; reskilling enrollment/completion trends.
DP_4 — Demand-side transmission: do displaced earners’ consumption cuts feed back? Routes: (consumption holds → S2/S5 recovery) vs (consumption cuts transmit → S8 cascade). Leading indicator: consumption among displaced cohorts; savings-rate divergence; transfer-payment adequacy.
DP_5 — Surplus destination / rent distribution (named the master variable by the other stream): does the productivity rent reach wages or concentrate in capital? Routes: (reaches wages/new demand → S9 boom) vs (concentrates in capital → S4/S8). Leading indicator: labor share of income vs productivity growth; median real wage vs GDP-per-capita divergence.
DP_6 — Labor-supply tightness / demographics: Routes: (tight supply → S7 absorption) vs (slack supply → S2/S3 displacement) — same capability, opposite labor outcome. Leading indicator: vacancy duration in non-tradable sectors; participation-adjusted dependency ratios.
DP_7 — Scaling / capability returns: Routes: (returns flatten → S1) vs (returns resume → S5). Leading indicator: frontier benchmark slope; frontier-lab capex trajectory.
DP_8 — Task-structure selectivity (broad-shallow vs narrow-deep): Routes: (broad-shallow → S2) vs (narrow-deep structural → S4). Leading indicator: distribution of softening across the wage curve — which BLS/SOC codes go first.
DP_9 — Embodied-robotics deployment pace: Routes: (robotics matures in-horizon → S10 refuge collapse) vs (robotics stays brittle → refuge-holds scenarios); gates the physical half of S7. Leading indicator: humanoid unit-cost vs sectoral wage; sim-to-real reliability on unstructured manipulation; robotics fleet capex; trade/care vacancy duration reversing.
DP_10 — Policy regime (redistribution vs deregulation): Routes: (redistribution → dampens all) vs (deregulation → accelerates all). Leading indicator: UBI/wage-insurance pilots; AI-deployment-tax proposals; trade-barrier moves on AI services; deployment moratoria/work-sharing mandates.
Cross-sector-correlation tripwire: simultaneous (not sequential) hiring freezes across ≥3 unrelated white-collar sectors signals a shared-capability-threshold shock (one cause, many sectors) rather than sectoral coincidence — two correlated sectors can be a fluke; three unrelated ones strain the coincidence reading. This is the S3→S5 tipping signal.
Feedback Loops (explicit)
- R1 — Displacement → tooling jobs → ? (reinforcing or balancing). Displaced workers’ problem creates agent-ops/eval/orchestration jobs. Balancing if new jobs absorb the displaced (heals S3’s bottom rung); fails if new jobs require different skills/locations (P2c/P3c). The closing link is the skill/geography matching step — the weakest link and the right intervention point.
- R2 — Productivity → output expansion → more labor demand (balancing, Jevons-like) vs Productivity → margin → headcount cut → less labor demand (reinforcing-down). Which dominates is the S2/S3 fork; dominance flips at the business cycle (recession favors the cut loop — P2a).
- R3 — Adoption → incident → regulation → adoption friction (balancing). The self-braking loop in P5d and S1’s driver. Delays are months-to-years — delay neglect would mis-time every prediction.
- R4 — Demographic exit → shortage → AI/robotics pull-in (balancing, S7 + S10). Closing link is timing alignment (P7a) and, for the physical half, robotics pace (R5).
- R5 — Robotics cost decline → physical-task automation → refuge erosion → fewer reallocation destinations (reinforcing-down for surplus economies; balancing for shortage economies). Benign where labor is short (fills the hole), harmful where surplus (removes the last refuge). Loop sign is set by demographics.
- The canonical displacement→retraining→new-jobs→further-displacement loop (the user-flagged one): its period — how fast capability re-eats reskilled roles — is the load-bearing unknown across S2/S5/S8.
Failure Pathway Stress Test Findings
Leading-scenario cutoff (surfaced tension between the two carvings): one stream ran the pre-mortem against the top-4 by mode-weight (S2, S4, S6, S5); the other ran full pre-mortems on the decision-relevant scenarios (modal continuity, rolling displacement, and the agentic discontinuity by impact×irreversibility rather than probability, plus the robotics scenario). Net: pathways exist for S2, S3, S4, S5, S6, S7, S9, S10. S8 is the failure mode several others decay into; S9 carries one pathway because the upside deserves a named failure mode. Pre-mortem value scales with impact × irreversibility, not probability alone.
S2 — Gradual Augmentation / Grinding Reallocation:
- P2a — “The augmentation offset never arrived” / “augmentation curdled into substitution.” 2031. Gross displacement matched expectations but augmentation-and-new-task creation materialized at a third of the assumed rate; firms banked productivity as margin and headcount cuts, never redeploying — and used a 2029 recession to convert augmentation gains into permanent cuts and never rehired. Net unemployment drifted to ~7%; gradual displacement arrived as a business-cycle step-function. Leading indicators we ignored: productivity-per-worker rising while job-posting volume fell 2027–28; rev-per-employee gains routed to layoffs not expansion. Recoverability: low once cuts land — rehiring lags structurally. This is how S2 becomes S8; S2’s modal mass quietly contains a ~quarter-weight tail indistinguishable from S8 ex ante. Provenance: pre-mortem-action.
- P2b — “Demographic cushion was double-counted.” 2031. Retirement-attrition that was supposed to absorb displacement turned out to be the same jobs being automated — boomers retired from the roles AI was eating, so attrition and displacement overlapped rather than adding. The cushion was half its assumed size. Leading indicator (measurable now): overlap between high-retirement and high-automation-exposure occupations. Recoverability: high if caught early — a workforce-planning measurement error. Provenance: pre-mortem-action.
- P2c — “Reskilling treadmill outpaced the runners” / “mitigation failed and acceleration followed.” 2031. People retrained into “safe” adjacent roles; capability crept up the ladder and ate those within ~18 months; retraining ROI went negative and workers stopped. Programs also targeted the last war (data analytics) while the displaced were content/support workers — the reskilling→new-jobs loop broke at the matching step, compounding into long-term detachment. Leading indicator: declining reskilling enrollment/completion after 2028; falling wage-on-re-entry. Recoverability: low — a confidence/detachment variable, hard to rebuild. Provenance: pre-mortem-action.
- P2d — “The productivity dividend never appeared.” 2031. Measured output gains stayed inside noise through 2028; task-level augmentation was real but eaten by coordination overhead and error-checking. S2 failed sideways into S1 disappointment rather than downward into displacement. Recoverability: moderate — a later capability step could still ignite it. Provenance: pre-mortem-action.
- P2e — “Polarization outran the safety net.” 2031. Augmentation held headcount but hollowed the wage middle; political backlash against “ghost work” conditions triggered abrupt regulation that fragmented the trajectory. Recoverability: moderate; the regulatory path is sticky. Provenance: pre-mortem-action.
S3 — Rolling Sectoral Displacement:
- P3a — “The rolling waves merged into a flood.” 2031. Three sectors modeled as sequential (support, content, junior dev) tipped within the same 18 months because they shared one capability threshold — reliable multi-step completion. What was priced as “rolling” was a single correlated shock; the independence assumption was the error. → collapses into S5. Leading indicator: simultaneous hiring freezes across ≥3 unrelated sectors. Recoverability: low — speed is the harm. Provenance: pre-mortem-action.
- P3b — “The broken bottom rung healed itself.” 2031. New categories (agent-ops, eval, orchestration, AI-liability) absorbed displaced juniors faster than projected; the displacement→tooling-jobs loop ran hot; the net-negative never appeared. → reverts toward S2/benign. Recoverability: n/a (benign failure). Provenance: pre-mortem-action.
- P3c — “Geographic non-substitutability bit.” 2031. Displaced tradable-service workers in developing economies could not migrate into the non-tradable shortage sectors (trades, care) that exist only in developed economies — “shortage absorbs surplus” was true globally and false locally. The absorption pathway proved fictional exactly where most needed. Leading indicator: developing-economy service-export receipts; time-to-rehire. Recoverability: low without migration-policy change. Provenance: pre-mortem-action.
S4 — Bifurcated Hollowing:
- P4a — “The trades floor flooded.” 2031. Displaced cognitive workers + robotics maturation converged on the trades simultaneously; oversupply crushed the wage premium that made them a refuge. The “safe” tier wasn’t safe — it was crowded. Leading indicator: trades-wage premium compression; career-changer training spikes. Recoverability: medium — embodied work has real capacity limits, but the wage signal can break. Provenance: pre-mortem-action.
- P4b — “The middle didn’t hollow — it liquefied upward.” 2031. The elite judgment tier compressed too, because agentic systems reached senior-analyst quality faster than expected; the protected top wasn’t protected. S4 failing into S5. Leading indicator: automation reaching partner/principal-level deliverables. Recoverability: low — removes the assumed safe harbor. Provenance: pre-mortem-action.
- P4c — “Developing-economy collapse fed back.” 2031. The hollowing of developing-economy services export was treated as their problem; it collapsed a major consumer market and remittance flows, transmitting back to developed-economy demand. The regional differentiation the model relied on was an illusion in a networked economy. Leading indicator: developing-economy services-export employment + remittance trajectories (the collapse hits hardest in AI-consumer developing economies with no surplus to redeploy). Recoverability: low — networked, slow to reverse. Provenance: pre-mortem-action.
S5 — Agentic Step-Change / Discontinuity (elevated to leading by impact × irreversibility, not probability alone):
- P5a — “Velocity assumption was the error, not direction” / “the threshold wasn’t a cliff.” 2031. The threshold was crossed but integration debt, change-management, and trust-building stretched the 24-month compression into a 5-year grind; reliability also improved continuously rather than as a cliff, so deployment stayed gated by liability/integration even when capability existed. S5’s direction was right, its speed wrong — it presented as S2 throughout. Leading indicator: gap between capability-crossing date and headcount-impact date in lead adopters. Recoverability: n/a — a “right but mis-timed” failure that corrupts every dependent decision (benign for workers). Provenance: pre-mortem-action.
- P5b — “Liability never cleared, so agents never shipped unsupervised.” 2031. Capability crossed but no insurer would underwrite unsupervised agent decisions and no licensing body would sign off; the 5:1 supervisor ratio never improved and the replacement economics never closed. S5’s capability arrived, its displacement didn’t — quietly resolving as S6. Leading indicator: agent-error liability cases stalling; insurer carve-outs for AI-autonomous decisions persisting past 2029. Recoverability: n/a as a forecast error — every workforce-contingency plan over-provisioned. Provenance: pre-mortem-action.
- P5c — “Compute and energy capped the rollout.” 2031. The threshold was crossed in the lab but serving agentic inference at economy-wide scale ran into compute supply, grid capacity, and cost-per-task floors; deployment was rationed to highest-value workflows and broad displacement became a slow, capacity-gated trickle. The bottleneck was kilowatts and silicon, not capability or liability. → throttled into S3. Leading indicator: inference cost-per-task plateauing/rising; data-center power constraints; frontier rate-limiting into 2029–30. Recoverability: eases on a hardware/energy buildout timeline outside the horizon. Provenance: pre-mortem-action.
- P5d — “The shock triggered its own brake.” 2031. Speed of displacement produced fast political response (deployment moratoria, work-sharing mandates) converting the discontinuity into a managed gradient — mitigation succeeded, moving the world toward S2 at high social cost. Recoverability: moderate, politically contingent. Provenance: pre-mortem-action.
S6 — Friction Wall:
- P6a — “The coiled spring released at once.” 2031. Friction held until a single landmark liability ruling + an insurer’s decision to cover AI-agent errors flipped the calculus industry-wide in one quarter; the accumulated overhang discharged as an S5-style step-change — worse than if friction had never existed, because adjustment was compressed. Leading indicator: convergence of insurer posture + precedent case + licensing reversal within ~12 months. Recoverability: low — the snap forfeits the gradual adjustment that was friction’s whole point. Provenance: pre-mortem-action.
- P6b — “Regulatory arbitrage hollowed the wall.” 2031. High-friction jurisdictions protected workers by exporting the work to low-friction ones; domestic employment held while the function offshored. The wall protected the labor market on paper while jobs left through the side door. (Also: compliance-driven consolidation favored large incumbents who automated under the radar — the freeze reduced new jobs without saving old ones.) Leading indicator: AI-deployment migration to low-regulation jurisdictions; services-import substitution. Recoverability: medium — addressable via trade policy, politically hard. Provenance: pre-mortem-action.
S7 — Demographic Absorption:
- P7a — “Timing mismatch.” 2031. AI hit economic-task reliability before the retirement wave peaked (~2030+), so for several years it displaced rather than backfilled — a transient S3 inside the S7 envelope. Recoverability: moderate — demographic pull eventually arrives. Provenance: pre-mortem-action.
- P7b — “Absorption was sectorally misaligned.” 2031. AI got good at white-collar work that wasn’t short and stayed bad — absent robotics — at the physical trades that were short; the shortage persisted and cognitive-side displacement occurred. With the narrowed mechanism this is an expected partial-coverage outcome, not a contradiction — the residual physical shortage is structural unless S10 fires. Recoverability: depends entirely on robotics pace. Provenance: pre-mortem-action.
- P7c — “Capital captured the dividend, not labor.” 2031. AI filled the demographic hole but gains accrued to capital owners and the shrunken incumbent workforce rather than reopening the entry ladder; “we needed it” was true at aggregate and false at distributional level — measured unemployment stayed low while labor share fell, reproducing the S2 quality-erosion pathology inside an absorption story. Recoverability: low without distributional policy. Provenance: pre-mortem-action.
S9 — Complementarity Boom:
- P9a — “The boom went to capital.” 2031. Output and productivity rose as hoped but the surplus concentrated in capital owners and a thin AI-producer elite; median real wages stagnated while GDP grew. The boom was real in aggregate and invisible in the median paycheck — it presented as S4 hollowing with better headline numbers. Complementarity was technological but not distributional. Leading indicator: labor share falling while productivity rises (2027–29); median wage decoupling from GDP-per-capita. Recoverability: medium — redistribution can reconnect surplus to demand, but it’s politically hard and the window narrows as capital concentration entrenches. Provenance: pre-mortem-action.
S10 — Embodied-Robotics Acceleration / Refuge Collapse:
- P10a — “The cost curve bent but reliability didn’t.” 2031. Robots got cheap in the demo and stayed brittle in the field; deployment stalled at structured-environment niches (warehouse, not home-care), so the refuge held for high-touch trades and care work. → reverts to S3/S7. Recoverability: n/a (refuge survives). Provenance: pre-mortem-action.
- P10b — “Physical-world friction throttled it.” 2031. Capability existed but regulation, liability, unionization, and physical-integration cost paced deployment to a crawl — capability discontinuity didn’t transmit to labor, mirroring P5a in the physical domain. Recoverability: moderate — friction is partly policy-chosen. Provenance: pre-mortem-action.
- P10c — “It arrived but only where labor was already short.” 2031. Robotics matured fastest exactly in the aging-economy shortage sectors that wanted it (S7 overlap); the global refuge-collapse was real in surplus economies and welcome in shortage ones — bifurcated by demography rather than a uniform shock. Recoverability: low for surplus economies, benign for shortage ones. Provenance: pre-mortem-action.
Integrated Forward Architecture
These are forecast-claims no single component — scenario, band, or pre-mortem — could produce alone; both analytical streams converge on this structure.
1. The modal forecast is a region, not a scenario. The 5-year period most likely lives in the gradual-augmentation-to-rolling-displacement region (S2 ∪ S3, biased toward narrow-deep cognitive-middle hollowing S4, throttled by deployment friction S6) — manifesting as wage/quality erosion and a broken entry-level rung, not aggregate headcount collapse. This is stated as a regional locating-of-mass judgment over the outcome-space, not a sum of bands — the scenarios are a non-exclusive cover and their bands cannot be added. Anyone betting on a single scenario is mis-specified; no single alternative individually rivals the combined plausibility of this region.
2. The high-probability scenarios are unstable equilibria that decay into the extreme ones — the comfortable scenarios are the dangerous ones. S2’s band contains its own failure into S8 (P2a); S4’s contains its failure into S5 (P4b); S6’s contains a worse-than-S5 snap (P6a); S5 contains its quiet resolution back into S6 (P5b); S3’s “rolling” assumption is one correlated-threshold from S5 (P3a). A naive “60%-ish chance things stay manageable” inverts the actual risk structure: the manageable scenarios are the delivery mechanism for the unmanageable ones, because their probability mass is partly borrowed from the tails they can collapse into. This claim requires scenarios + bands + pre-mortems jointly.
3. The single load-bearing node is where the surplus lands — the demand-transmission / rent-distribution variable. Every doom pathway (P2a, P4c, S8) and the upside (S9) route through one node: does the productivity surplus reach the median worker — as wages, new-task demand, or transfers — before displaced-cohort consumption cuts feed back? This is the variable the probability formalism prices least well (reflexive, policy-contingent) and the one determining whether the period stays benign, tips into the malign tail, or breaks upward into the boom.
4. The upside and the tail share a single switch. S9 (boom) and S8 (cascade) are the two ends of one variable. Rent-capture (S9’s internal tipping point) and demand-transmission (S8’s) are the same node viewed from opposite ends. This is why the master question is “where the surplus lands,” not “how much displacement” — and why the Rent-Capture Axis (who holds the surplus, regionally) is load-bearing: it names which economies can choose S9 over S8. Requires S8 + S9 + the regional rent axis together.
5. Friction is double-edged in a way neither scenarios nor probabilities show alone. S6 has a protective mode and, via P6a, a catastrophic-amplifier mode — the same mechanism. Friction doesn’t reduce total risk; it trades gradual risk for tail risk. Requires the scenario + band + pre-mortem together.
6. Two master-variable nominations stand in tension (surfaced, not resolved). One synthesis names where-the-surplus-lands (demand/distribution) as the single most load-bearing variable; the other names the agentic-reliability threshold crossing in a correlated way as the variable that would invalidate the whole modal reading. These are not the same variable — the first determines the consequence of displacement, the second determines its velocity and breadth. The two together bracket the forecast: the threshold sets whether displacement is rolling (S3) or correlated (S5); the surplus-node sets whether the result is absorbed (S2/S9) or cascades (S8). A complete monitoring posture watches both, plus embodied-robotics pace (the variable that would collapse the physical refuge the whole modal region implicitly trusts → S10).
Constructive-Future Gap-Flag
Constructive-future gap: Backcasting (constructive-future stance) is deferred per CR-6. This analysis covers descriptive-future only — probability-weighted scenarios and adversarial-future stress testing. Users requiring constructive-future framing (working backward from a desired future to identify required interventions) should compose Wicked Future with downstream goal-articulation work.
Concretely: this analysis is descriptive-future only — it does NOT tell you what to do. Backcasting — the stance that starts from a desired 2031 labor outcome and works backward to the policy/institutional moves that reach it — is absent from this output. Everything above maps what might happen and how each path fails; none of it constructs a target state or a route to it. Concretely absent: a normative target for the labor transition (e.g. “preserved entry-level pipeline + functional reskilling matching + maintained wage floor”); a reskilling-system design; a redistribution-mechanism comparison (even though the architecture identifies redistribution as the S8→S9 switch, it does not design it); a sequencing of policy interventions; the reverse-chained moves from a desired end-state back to present decision nodes. The watchlist below is diagnostic (tells you which branch you’re on), not prescriptive — an instrument panel, not a flight plan. If you need the prescriptive layer — given a desired labor-transition outcome, what to do now — commission a separate Backcasting pass explicitly.
Divergence-Points-to-Monitor — Forward Watchlist
Ordered by information value (how much resolving each collapses the scenario space):
- Where the surplus lands — labor share of income vs productivity, and consumption among displaced cognitive cohorts. Pattern that would signal: the master distributional variable; resolves the benign/malign/boom split (S8 vs S2 vs S9). Highest value.
- Production agent reliability + supervisor-ratio data in regulated workflows — long-horizon task-completion-without-correction crossing ~95% (economic-trust floor). Pattern that would signal: resolves S1/S2 vs S5/S8; the other candidate master variable.
- Correlation of hiring freezes across unrelated white-collar sectors — pattern that would signal: simultaneous across ≥3 = the S3→S5 merged-flood tipping signal.
- Insurer/liability/licensing posture shifts — pattern that would signal: resolves S6-coiled vs S5-released; watch the convergence of all three (the P6a trigger) and the persistent carve-out that would realize P5b.
- Wage-on-re-entry for displaced workers — pattern that would signal: resolves the reskilling-treadmill question (P2c) and the retraining-half-life divergence.
- Overlap map: high-retirement × high-automation-exposure occupations — pattern that would signal: resolves the demographic-cushion double-count (P2b). Measurable now, cheap — do this first.
- Rev-per-employee gains routed to layoffs vs expansion (especially in a downturn) — pattern that would signal: the S2→S3 conversion / R2 down-loop tell.
- Distribution of softening across the wage curve (which BLS/SOC codes go first) — pattern that would signal: resolves S2-broad vs S4-narrow.
- Developing-economy services-export + remittance flows, by producer/consumer status — pattern that would signal: early warning for networked-feedback pathways (P4c); read against the Rent-Capture Axis.
- Frontier benchmark slope + frontier-lab capex — pattern that would signal: resolves S1 (disappointment) vs S5 (resume).
- Non-tradable vacancy duration + dependency ratios — pattern that would signal: resolves S7 (absorption).
- Inference cost-per-task + data-center power availability — pattern that would signal: resolves the compute/energy cap (P5c) on agentic rollout speed.
- Humanoid unit-cost vs sectoral wage + sim-to-real reliability + trade/care vacancy duration reversing — pattern that would signal: resolves S10 (refuge collapse) activating, or S7’s physical half filling.
Residual Uncertainties
The bands price risk (likelihood knowable in principle). These are Knightian / model-misspecification uncertainties — explicitly not assigned probabilities; presenting them as priceable would be the error this mode exists to avoid.
- Agentic-threshold timing — why unpriced: Knightian. Whether/when unsupervised reliability crosses the deployment line depends on research breakthroughs whose arrival distribution is genuinely unknown. S5’s band is the least-defensible number for this reason — and the two streams’ band disagreement (~12% vs ~22%) is the visible signature of this unpriceability. Treat the S5 band as a flag, not a measurement.
- Embodied-robotics deployment curve — why unpriced: Knightian, reference-class-thin within a 5-year window; S10’s 4–10% inherits the same flag status. Divergence between the streams: one scenarioed embodied robotics explicitly (S10), the other underweighted it to a residual wildcard and noted that if embodied capability advances faster than assumed, the trades “safe sink” S4 leans on (P4a) could fail through an underweighted channel. The disagreement about whether this merits a scenario or a caveat is itself a finding.
- Correlation structure across sectors (rolling-vs-flood, P3a) — why unpriced: a model-misspecification risk; the scenarios assume partial independence; if one capability threshold gates many sectors, the independence assumption fails and the whole distribution shifts mass toward S5.
- Reflexive policy response — why unpriced: outcomes are partly caused by forecasts of them (a credible S8 forecast triggers the policy that prevents S8, or seeds the redistribution that realizes S9). The probability is endogenous to the act of forecasting.
- Surplus-distribution politics (the S8/S9 switch) — why unpriced: whether the surplus reaches labor is a political-economy outcome with no clean base rate — which is why S9’s 8–18% and S8’s 5–12% bands are both soft.
- Demand-feedback nonlinearity — why unpriced: the S8 cascade depends on a threshold (how much displacement before demand transmission dominates) with no historical calibration because no prior automation wave removed the cognitive-labor-cost floor simultaneously. S8’s 5–12% is the softest number here — could be materially wrong in either direction.
- Scenario-taxonomy misspecification — why unpriced: unquantifiable by construction; the real 2031 may be no blend of S1–S10 but a state unconceived (e.g. AI advances while labor-market institutions reorganize around it in a form with no historical analog).
- Unknown unknowns — why unpriced: a non-labor shock (geopolitical, energy, financial) dominating the labor-AI dynamic entirely; a capability modality not modeled at all.
The claim is not that the S5/S8/S10 tails are precisely bounded — it is that they are minority outcomes under the diffusion-lag reference class, while flagging that the reference class may not apply. Presenting Knightian ambiguity as priceable risk is exactly the failure mode the mode requires surfacing.
Confidence Map
| Element | Confidence | Basis |
|---|
| 2026 baseline (no aggregate displacement signal yet) | High | BLS May-2026 AHE/unemployment [0.80] + NCCI payroll-acceleration [0.30] + ILO quality-stagnation |
| Modal region = augmentation-to-rolling-displacement (S2∪S3, biased S4, throttled S6) | Medium | Diffusion-lag reference class + structural reasoning; bands wide |
| S2/S3/S4/S5/S6 are unstable equilibria decaying into adjacent states | Medium-High | Robust to pre-mortem logic; direction solid, magnitudes not |
| Surplus-destination as master node (S8/S9 shared switch) | Medium-High | Strong structural case; unpriceable mechanism |
| Agentic-reliability threshold as the alternative master variable | Medium-High | Strong structural case; timing Knightian |
| S5 band (band-disagreement, ~12% vs ~22%) | Low | Partly Knightian; no clean reference class — disagreement preserved, not averaged |
| S8 band (5–12%) and S9 band (8–18%) | Low | No historical calibration for the distributional/cascade threshold |
| S10 band (4–10%) | Low | Robotics-curve reference-class-thin; scenario-vs-caveat disagreement noted |
| Regional differentiation holding | Medium | Solid for S1–S7; collapses in S8 (networked demand); rent-capture axis sharpens but doesn’t eliminate |
| Producer/consumer rent axis (incl. China as distinct case) | Medium | Structurally sound; magnitude of rent-capture leverage unquantified |
| Regional ordinal anchoring (Philippines/India IT-BPM shares) | Medium | Directional/approximate, not country-level modeled |
| Friction = gradual-risk-for-tail-risk trade (S6 + P6a) | Medium | Falls out of integration; novel, less battle-tested |
| Physical-refuge assumption | Medium-Low | Now stress-tested by S10 rather than assumed; robotics pace the open variable |
Additional Considerations
Critical-question coverage. Scenario breadth is met — trend-extrapolation (S2, S3), orthogonal-driver (S4, S7, S9), discontinuity (S5, S8, S10), reversal (S1, S6); multiple non-trend-extrapolation scenarios survive prominently, and S9 makes the space genuinely multi-outcome, not multi-severity. Integration is met — bands are tied to the underlying divergence variables and internal tipping points; the integrated architecture (band-collision, shared-switch, unstable-equilibrium, friction-trade findings) uses scenarios + bands + pre-mortems jointly; the modal claim is a non-additive regional locating-of-mass, arithmetically coherent (non-exclusive cover, bands not summed). Pre-mortem coverage is met — past-tense prospective-hindsight pathways with leading indicators and recoverability attached to the leading scenarios under both cutoff rules (top-by-mode-weight and by-decision-relevance, the latter elevating S5 by impact×irreversibility); every leading scenario carries ≥3 pathways, plus an S9 upside pathway and S8 standing as the failure mode others decay into. The constructive-future gap is flagged at its own position above, with the watchlist explicitly labelled diagnostic-not-prescriptive.
Cross-stream tensions preserved, not resolved. (1) Scenario carve-up of the modal space — single grinding-reallocation scenario vs split gradual-augmentation + rolling-sectoral-displacement; task-structure carving (S4) vs temporal-sequence carving (S3); both retained, the breadth is the value. (2) The S5 band disagreement (~12% vs ~22%) — a finding about robustness to threshold-timing priors, not a number to average. (3) Master-variable nomination — surplus-destination vs agentic-reliability-threshold; bracketed in Integrated-Architecture point 6: they govern consequence vs velocity respectively, and a complete posture watches both. (4) Embodied robotics: scenario vs caveat — one stream promoted it to a full scenario (S10) stress-testing the physical refuge, the other left it a residual wildcard; the disagreement about its weight is itself recorded.
Phase A assumptions you may want to correct. (1) Scope centered on generative/agentic AI with robotics secondary — if embodied automation is a co-equal driver, S4’s trades-refuge logic and S7’s physical-shortage gating need re-running (S10 partially pre-empts this). (2) Global scope with developed/developing differentiation, sharpened by the producer/consumer rent axis — narrowing to one region shifts the bands. (3) Point-estimates reinterpreted as bands (rationale at the head of this analysis). Flagging any of these re-runs the affected layer.