Part VII - The Indicator Dashboard#
Contents
Baseline readings as of mid-2026.
A forecast that cannot be checked before its horizon is not a forecast. This part converts the claims in Parts I–VI into observables with thresholds - things that can be read off the world on a quarterly cadence, each attached to the specific claim it would confirm or break.
The discipline: every indicator here names a trigger value, and every trigger names what to revise. An indicator that moves without changing anything you believe was decoration.
The three families#
| Family | Answers | Cadence | |
|---|---|---|---|
| A | Substrate | Are the physical and financial ceilings binding on schedule? | Quarterly |
| B | Diffusion | Is the technology actually reaching into the economy? | Quarterly to annual |
| C | Governance | Which regulatory architecture are we going to get, and when? | Event-driven |
The families are not independent, and reading them jointly is most of the skill. A binding A-family constraint shows up later as a B-family slowdown, so a B stall with A slack points at institutions, while a B stall with A tight points at physics and finance - opposite revisions from the same B reading. And a C-family event reprices both: the post-incident window can relax A constraints (permitting waivers) while tightening B (deployment moratoria) simultaneously.
The five that matter most#
If only five things can be tracked, track these. Each is the leading edge of a different claim the corpus depends on.
| Indicator | Baseline (mid-2026) | Watch for | Breaks |
|---|---|---|---|
| Entry-level : senior posting ratio, knowledge professions | Depressed; entry-level ~7% of large-tech hires | Failure to recover when aggregate white-collar hiring does | Game 4 - settles the substitution-vs-cyclical split |
| Internal research cycle time per validated experiment at frontier labs | Not disclosed; infer from release cadence and paper throughput | Sustained >2× compression and physical/financial governors not binding | Uncertainty 1 - the RSI loop closing |
| AI revenue run-rate vs. capex | ~$60B (2025) revenue vs. ~$700–725B 2026 hyperscaler capex guide | Revenue CAGR falling below ~40% for two consecutive years | Compute and Capital - the correction |
| Teleoperation-to-autonomy ratio in deployed robots | Overwhelmingly teleoperated or scripted | Autonomy share crossing 50% on unstructured tasks | Robotics - the 2030s swing variable |
| Grid interconnection queue length | 5–7 yrs; ~410GW CAISO, ~380GW MISO | Any market halving its typical wait | Energy, Uncertainty 2 |
Why these five lead#
Each was chosen for when it moves, not only for what it measures. The posting ratio moves at the hiring decision, quarters before employment stocks show anything. Cycle time moves inside labs before any release demonstrates it. The revenue-to-capex gap moves in guidance and filings a year or more before a write-down forces the same fact into the open. The teleoperation ratio moves at deployment, years before robotics revenue is large enough to surface in sector statistics. And queue length moves at the application stage, roughly the length of the queue itself ahead of power actually binding. A dashboard of lagging indicators confirms; this one is built to warn. The cost accepted in exchange is a higher false-alarm rate - early-moving series are noisier than the stocks they anticipate, which is why each threshold requires persistence (two quarters, two consecutive years) rather than a single print.
Also on the short list (rounds 5–8)#
| Indicator | Family | Breaks |
|---|---|---|
| AI liability as published insurance line + reinsurer capacity | C7 | Uncertainty 6; deployment frontier |
| Leading-edge wafer share outside Taiwan | A6 | Uncertainty 4 |
| Cognitive vs physical services inflation gap | B7 | Prices two-economy claim |
| Capability gap: verifiable vs unverifiable domains | B8 | Uncertainty 5 |
| Labor share of tax revenue | B10 | Fiscal squeeze |
| Companion use × social contact; junior identity | B11 | Meaning welfare channel |
| Capture-time provenance at scale | C8 | Game 5 / media enclosure vs public verification |
| Warehouse / freight / specialty harvest autonomy | B12 | Robotics physical fork |
| Taiwan gray-zone (freight insurance, license denials, delivery slip) | U4 | Soft fail of fab continuity without crisis headline |
How to read a dashboard without fooling yourself#
Four failure modes, all of which this document is exposed to:
- Confirmation drift. An indicator chosen after the fact to support a held view tests nothing. Every threshold here was set before the reading was taken; when one is revised, the revision goes in the log, not silently into the table.
- Measurement capture. Once an indicator becomes salient, the thing being measured starts being managed. Benchmark scores went this way already. Prefer indicators that are expensive to fake - capital flows, physical deployments, queue positions - over self-reported ones.
- The perception gap. The METR result in Game 4 means self-reported productivity is unreliable in a known direction. Weight instrumented and revealed-preference measures far above surveys.
- Absence of movement is information. The most common real outcome is that an indicator sits still for three years. That is a result, and it usually favors the institutional-friction case over both the acceleration and the collapse case.
Scoring cadence#
- Quarterly - read families A and B, update baselines, note any threshold crossed.
- Annually - re-score every probability in Part V against the dashboard and record the delta. Probabilities that never move are not being updated; they are being defended.
- On event - any salient incident triggers a full re-read of family C within the 18-month window when the architecture gets written.