assembly-economics/research/ScenarioBandMethod.md

<!--
Created 2026-10-12 at the project owner's direction. Defines how the program
handles AI adoption uncertainty: bands of scenarios instead of forecasts,
jurisdiction archetypes, robustness classification across the band, and
regular checkpoints against real data. Referenced from INTENT.md, Research
Method.
-->

# Scenario Bands, Jurisdictions and Reality Checkpoints

**Status:** method v0.1. It governs how GADM results are produced and
reported from ASSEMBLY-WP-0009 onward.

## 1. Premise

There is no historical precedent for AI adoption. Its speed comes from an
**inner self-enhancing loop** (organisms learn with use, AI accelerates its
own development, investment follows returns). Whether that loop accelerates,
stalls or breaks depends on things outside economics proper:

- **legislation:** what may be deployed, who controls membranes, what must be
  explainable;
- **adoption decisions:** by firms, public bodies and individuals;
- **sentiment:** whether the change is framed and experienced as good or bad
  for society. Sentiment then feeds back into legislation and adoption.

These differ strongly between jurisdictions, and they change over time.
Hence:

> The program does not forecast one adoption path. It works with a **band**
> of adoption paths, probes every scenario across that band, checks which
> results hold across it, and compares the band with real data at regular
> checkpoints.

## 2. The adoption band

Four anchors, all expressed through K_Ω (the share of information tasks
carried out by information organisms; 0.05 today):

| Anchor | K_Ω at 10 years | Basis | Model |
|---|---|---|---|
| **B1 conservative** | ≈ 0.10 | Acemoglu (2024): about 4.6 % of tasks cost-effectively affected within 10 years | `grounding.variant("slow")` |
| **B2 central** | ≈ 0.27 | between LLM-only exposure (15 % of tasks faster) and the tooling ceiling (Eloundou et al. 2023) | `grounding.variant("central")` |
| **B3 fast** | ≈ 0.50 | exposure ceiling with tooling (47–56 %) reached within a decade | GADM-0.4 adoption defaults |
| **B4 aggressive / self-enhancing** | → 0.8+ within 10–15 years | AI accelerating its own development: self-replicating and self-improving capital (Trammell & Korinek), automated AI research (Davidson et al.) | adds a self-enhancement loop: organism capability raises its own investment and learning rates |

B1 and B2 rest on the evidence collected in ASSEMBLY-WP-0008. B3 and B4 are
not "unlikely tails" to be discarded. B4 is the case where the inner loop
dominates, and it is the case most likely to need a *qualitatively different*
model.

## 3. Modulators: legislation and sentiment

Adoption is not exogenous. Two modulators act on the inner loop.

**Legislation** sets, per jurisdiction:

- deployment permissiveness: scales organism investment and seeding;
- membrane rights (`m`): user/community control, portability;
- explainability and provenance duties (`expl`);
- lineage rules (`L`, `lin_open`): open inheritance, diffusion;
- attention protection (`x_att`).

**Sentiment** Σ is a slow state variable. It moves toward perceived net
benefit: broad claims, agency and attention headroom versus domination,
displacement and loss of attention. Σ scales adoption (people and firms adopt
what they trust) and, with a lag, shifts legislation (tightening when Σ falls,
loosening when Σ rises).

This gives a **legitimacy loop**:

```text
adoption → outcomes → sentiment → adoption and legislation → adoption
```

The loop can stabilize or destabilize the inner self-enhancing loop. It is a
model hypothesis in its own right and must be checked against data.

## 4. Jurisdiction archetypes

Archetypes are defined by **mechanism** (parameter settings), not by
country names. INTENT.md asks that assumptions be made explicit rather than
hidden behind political labels. Real jurisdictions are mapped onto archetypes
only with policy evidence, and the mapping is revisited at every checkpoint.

| Archetype | Permissiveness | Membrane rights | Explainability | Lineage rules | Initial sentiment |
|---|---|---|---|---|---|
| **J1 rights-first / precautionary** | low | high | high | open | mixed |
| **J2 deregulated fast-adopter** | high | low | low | concentrated | positive |
| **J3 platform-led** | high | low (operator) | medium | strongly concentrated | mixed |
| **J4 state-coordinated** | high | low for users, state-held | medium | state-concentrated | managed |

**The owner's hypotheses to test (not model inputs):** Europe will lag in
adoption (J1-like), and Argentina may thrive (J2-like). The model should be
able to say under which band anchor and which sentiment dynamics "lagging"
is better or worse for VFAC, agency and attention, and the checkpoints
should show which jurisdictions actually follow which archetype.

## 5. Probing scenarios across the band

Every reading is computed for every scenario × band anchor (× archetype,
where relevant), and classified:

| Class | Criterion | Consequence |
|---|---|---|
| **robust** | same sign, same ranking across the band | can be stated without qualification |
| **quantitative shift** | same sign and ranking; magnitude changes by more than 50 % | state with the band range |
| **qualitative break** | sign flip, ranking change, a different dominant channel, a floor violation appearing, or a tipping/hysteresis regime appearing in part of the band | flag: the model or the reading may need to be **qualitatively different** in that part of the band; record in a break register |

Global sensitivity analysis is run per band anchor too: do the dominant
drivers change across the band?

## 6. Reality checkpoints

Checkpoints compare the band with observed data, per jurisdiction where data
allow.

**Cadence:**

- **quarterly (light):** adoption indicators only (AI use shares, AI
  investment, usage indices);
- **annual (full):** sensor bundles for all stocks (`metrics/MetricsCatalog.md`),
  sentiment surveys, legislation changes;
- **event-driven:** major legislation, major capability jumps, market shocks.

**Prediction register.** Before each checkpoint, the current model version
writes down **pre-registered predictions**: indicator, jurisdiction, horizon,
band range per anchor. At the checkpoint:

1. Record the observation and its source.
2. Locate it in the band. Which anchor does reality track, per jurisdiction?
3. Classify it: inside the band / outside the band (model failure).
4. Decide on model changes explicitly: parameter re-grounding through
   `models/ParameterGrounding.md`, or structural change through a workplan.
   Never re-tune silently. Earlier predictions stay in the register
   unchanged.

Over time this measures which band anchor each jurisdiction follows, whether
the legitimacy loop behaves as modeled, and whether a qualitatively different
model is needed for some jurisdictions.

## 7. How this changes ongoing work

- Every new model reading is reported across the band (at least B1, B2, B3,
  and B4 where implemented), with its robustness class.
- ASSEMBLY-WP-0007 (goods, ecology, selection) reports its tests across the
  band.
- ASSEMBLY-WP-0009 implements the band anchors, the self-enhancement loop,
  sentiment, legislation and archetypes.
- ASSEMBLY-WP-0010 runs the checkpoints and keeps the prediction register.