REFERENCE · LAST UPDATED
The Absorption Index — Method
The full derivation: eleven anchored tracks, the arithmetic, the bounded editorial adjustment, what is excluded and why, and seven pre-registered triggers written before the events that would move the number down.
The question this page answers
How is the Absorption Index calculated, and what would make it fall?
The design
Core = mean(accelerant scores) − 0.4 × mean(brake scores), clipped to [0, 1.2].
Each track maps one observable onto 0–1 by linear interpolation between a declared floor (contributes nothing) and ceiling (fully at the crossing condition). Anchors are the judgment; once set, the score is arithmetic.
All judgment lives in the JSON spec. The script contains none. If the number moves, it moved because an observed value changed (with a citation) or an anchor changed (with a version bump), and git says which. That separation is the whole design.
Tracks the indicator catalog rules not %-able are excluded from the core, not estimated into it: robots (status only), deployment/automation (Anthropic dropped the metric), federal preemption (directional), public will (demoted to context). They remain as qualitative panel material. Fabricating numbers for them is the exact failure this method exists to prevent.
Editorial band: ±0.15. The published reading may differ from the core by at most that much, with a written justification. A delta outside the band means the anchors are wrong or the reading is — it is never resolved by widening the band.
The tracks and their anchors
Accelerants — what is arriving
| Track | What we measure | Now | Nothing | Extreme | Score |
|---|---|---|---|---|---|
| Capability — METR horizon doubling rate | months to double autonomous task horizon (lower = faster) | 2.9 | 7.0 | 1.0 | 0.68 |
| Benchmark saturation | fraction of tracked frontier benchmarks above 90% | 0.5 | 0.0 | 1.0 | 0.50 |
| Frontier training compute growth | multiple per year | 5.0 | 1.0 | 6.0 | 0.80 |
| Chip revenue growth | Nvidia data-center revenue YoY growth (fraction) | 0.68 | 0.0 | 1.0 | 0.68 |
| Hyperscaler AI capex growth | YoY multiple | 1.9 | 1.0 | 2.2 | 0.75 |
| Self-improvement on-ramp | max share of a frontier lab's production code written by its own AI | 0.8 | 0.0 | 1.0 | 0.80 |
| Capital concentration | AI share of US VC dollars | 0.86 | 0.2 | 0.95 | 0.88 |
| average | 0.728 |
Brakes — what is meant to catch it
| Track | What we measure | Now | Nothing | Extreme | Score |
|---|---|---|---|---|---|
| Lab safety grades | best existential-safety grade across labs, on a 0–1 scale (F=0, D=0.2, C=0.4, B=0.7, A=1.0) | 0.25 | 0.0 | 0.7 | 0.36 |
| Enacted regulation | US state AI bill enactment rate (enacted ÷ introduced) | 0.12 | 0.0 | 0.3 | 0.40 |
| EU AI Act enforcement | fraction of enforcement milestones in force on original schedule | 0.5 | 0.0 | 1.0 | 0.50 |
| Lab-initiated restraint | count of the five frontier labs with a publicly stated, incident-triggered slowdown or pause in force, / 5 | 0.2 | 0.0 | 1.0 | 0.20 |
| average | 0.364 |
The arithmetic
`` accelerants 0.728 brakes 0.364 x 0.4 = 0.146 ---------------------- core 0.582 ``
Anchors are the only judgment in the system. Disagree with one, change it, and recompute — the arithmetic above does not move on its own.
Downward triggers (pre-registered 2026-08-09)
The Index had no defined way to fall. These are written before the observations, so they cannot be rationalised after:
| ID | If this happens | Then | Check after |
|---|---|---|---|
| T1 | EU Art. 101 GPAI fines actually levied against a named provider, not merely activated | eu_enforcement 0.5 → 0.7 | 2026-08-02 |
| T2 | Any lab's FLI existential-safety grade rises above D+ | lab_safety_grades 0.25 → 0.40 | 2026-12-31 |
| T3 | METR's doubling rate decelerates in two consecutive periods | capability score falls | 2026-12-31 |
| T4 | US state AI enactment rate for a completed year exceeds 20% | enacted_regulation 0.12 → 0.20+ | 2027-01-31 |
| T5 | Any of the five biggest spenders cuts AI capex guidance | capex falls below 1.9× | 2026-10-31 |
| T6 | A second frontier lab states an incident-triggered slowdown, or OpenAI quantifies its own (named programme paused, dated resumption condition) | lab_self_restraint 0.2 → 0.4 | 2026-12-31 |
| T7 | OpenAI's promised full postmortem lands with a verifiable dated restriction on evaluation scope or agent permissions | lab_self_restraint up; may justify a containment track | 2026-11-30 |
What the Black Hat incident says about the capability track
The most striking capability datapoint of the year is not in the core, and under the method's own rules it should not be: weeks of coherent multi-agent operation, novel zero-day chaining, and a single pod escalated to cluster admin across multiple clusters in under 13 hours. None of that carries a clean %.
capability_doubling rests on METR's task-horizon doubling rate, which is the right choice — METR itself says the level is not robustly measurable above 16 hours, and the June 2026 run produced 11.3 hr / 71 hr / >270 hr depending on how detected eval-gaming is scored. But the incident is evidence that the level is understated by more than the method can express, and that the thing being measured (one agent, one task, one horizon) may be the wrong unit now that capability can belong to a population.
Not fixed by inventing a number. Recorded here as the leading candidate for what v2 needs to represent, and as a live example of the method's central trade: excluding un-measurable tracks keeps it honest and makes it blind in exactly the places that matter most.
The structural finding worth publishing
Seven accelerant tracks carry clean numbers; only three brake tracks do. That asymmetry is not a defect of this method — it is a fact about the world, and it is arguably the most publishable thing here. Accelerants are measurable because companies report revenue and capex quarterly and benchmarks publish boards. Brakes are hard to measure because enacted, enforced regulation is rare, slow, and counted differently by every tracker (2025 US enactments span 73 / 145 / 159 depending on who is counting).
So an index of this kind will structurally tend to read high, because the accelerant side is instrumented and the brake side is not. That is worth stating on the public page — it is a caveat that makes the instrument more credible, not less.
Governance
- Cadence: reviewed weekly via
/wireframe-index; recomputed whenever anobservedvalue changes. - Every move ships a cited rationale. No move without a named event.
- Non-moves are logged too — the changelog records a decision not to move, so silence stops being indistinguishable from stasis.
- Method changes bump
method_version, and the version in force is stated on every reading.
About this page
Maintained by Karl Herbst as part of Wireframe News, which covers how power concentrates across politics, AI, war and the future of work. Free to cite and quote with attribution (CC BY 4.0). Corrections and additions: wireframenews.com.