assembly-economics/models/AdoptionBand.md

Adoption Band — Anchors and Probe Results

Status: v0.1. GADM-0.4 with grounded power, care and attention parameters (grounding.py). The band anchors differ only in adoption parameters. Each starts at its own equilibrium, with recalibrated channels.

1. Anchors (T01)

Anchor	Adoption parameters	K_Ω at t 5 / 10 / 25 / 100	Basis
B1 conservative	aO 0.10, ell 0.166	0.09 / 0.10 / 0.12 / 0.07	Acemoglu (2024)
B2 central	aO 0.25, ell 0.415	0.20 / 0.27 / 0.28 / 0.16	between LLM exposure and the tooling ceiling (Eloundou et al. 2023)
B3 fast	aO 0.50, ell 0.83	0.39 / 0.49 / 0.46 / 0.34	tooling ceiling reached within a decade
B4 aggressive	B3 + self-enhancement selfenh 3.0	0.70 / 0.80 / 0.72 / 0.64	AI accelerating its own development (Trammell & Korinek; Davidson et al.)

Self-enhancement loop (new in GADM-0.4, default off): organism investment and learning are scaled by (1 + selfenh \cdot K_\Omega), so the more capable organisms are, the faster they improve. selfenh = 3.0 is tuned so B4 reaches about 0.8 within ten years. At selfenh ≥ 6, K_Ω saturates instead of declining, which is a different trajectory regime. That variant is not an anchor yet.

All anchors show a peak-and-decline path for K_Ω under baseline governance: rising domination and eroding demand slow organism investment after the first decade or two. This is a model property, to be checked against data (ASSEMBLY-WP-0010).

2. Probe results (T02)

Full tables: simulations/gadm02/results/band_probe.md. 36 readings: 22 robust, 7 quantitative shifts, 7 qualitative breaks.

2.1 Robust across the whole band

• Generative regime best, automated capitalism worst, generative regime lowest domination.
• Generative regime gain over baseline: +0.29 to +0.34 VFAC.
• Commons-rich economy gain over baseline: +0.10 to +0.12.
• Organism lever order lineage > membrane > attention.
• P3: a wealth tax alone barely moves domination.
• Dominant domination channel in all non-generative regimes: lineage (membrane in the attention oligopoly).

2.2 Quantitative shifts

Reading	B1 → B4	Reading
Dividend abundance gain over baseline	+0.038 → +0.013	dividends matter less as adoption accelerates
Automated capitalism loss vs. baseline	−0.099 → −0.047	largest at B1, where the baseline itself does best
Lineage lever effect	+0.164 → +0.098	the lever weakens as learning with use accelerates (P2: the critical lineage diffusion rises above the tested L = 0.8)
Membrane lever effect	+0.109 → +0.066	as above
P3 effect size	−0.018 … −0.055	small throughout

2.3 Qualitative breaks

Recorded in research/BreakRegister.md.

1. Regimes that bound domination only under conservative adoption. Dividend abundance and the commons-rich economy keep P under its 0.7 ceiling at B1, and breach it from B2 onward.
2. Thresholds crossed with faster adoption. Concentrated lineages breach the domination ceiling from B3. Attention extraction breaches the attention floor from B3, and the attention oligopoly breaches it at B4.
3. Aggressive self-enhancement changes the structure of domination. At B4, lineage control becomes the dominant domination channel even in the generative regime and its attention-extraction variant (B1–B3: market, or discovery).

2.4 Readings

1. The program's core claims hold across the band. What changes is which policies are sufficient: under conservative adoption, moderate reforms (dividends, commons) keep domination bounded. Under central to aggressive adoption, only the full generative package does.
2. This bears directly on the "Europe lags" hypothesis. A jurisdiction that adopts slowly (B1-like) can get by with moderate reforms. A fast adopter needs organism-specific governance early. Whether lagging is worse or better then depends on whether the slow adopter later imports B3/B4-speed organisms through foreign lineages. That needs open-economy structure the model does not have (ASSEMBLY-WP-0009-T04 / later).
3. B4 needs lineage governance beyond the current lever range. Under self-enhancement, even L = 0.8 and m = 0.8 leave lineage control dominant. Candidate mechanisms that GADM does not yet represent: public or commons lineages, compute commons, mandated lineage interoperability. This is a structural question, and the first candidate for "qualitatively different model" (break register entry BR-003).

3. Sentiment and legitimacy loop (T03)

3.1 Structure

gadm04 gains two slow stocks, switched on by sent = 1 (default off, so all earlier results are unchanged):

• Sentiment Σ (Sig) moves at rate sig_rate (≈ 5 years) toward 0.5 + sig_gain × (pb − pb₀). Perceived net benefit pb = 0.3 B + 0.3 G + 0.2 A + 0.2 (1 − P) − sig_disp × K_Ω (1 − m) is judged against its value at the start: people compare with the situation they know.
• Legislative stance Lg follows Σ with lag leg_lag (5 years).

Σ and Lg scale the organism inflow (investment, learning, seeding) by (1 + 2 sig_adopt (Σ − 0.5)) (1 + 2 leg_perm (Lg − 0.5)). A stance below neutral also tightens membrane rights m and explainability expl by leg_m, leg_expl × (0.5 − Lg). The parameters are uncalibrated; they set a mechanism, not a measured strength.

Reproduce: cd simulations/gadm02 && python3 -B legitimacy.py (results/legitimacy.md).

3.2 Results (baseline scenario, t = 100)

Anchor	K_Ω peak off → on	K_Ω t 100 off → on	Σ end	VFAC off → on	loop behaviour
B1	0.12 → 0.10	0.07 → 0.06	0.40	0.396 → 0.480	damping
B2	0.29 → 0.21	0.16 → 0.05	0.30	0.328 → 0.374	damping
B3	0.50 → 0.35	0.34 → 0.08	0.26	0.298 → 0.357	damping
B4	0.80 → 0.56	0.64 → 0.08	0.25	0.263 → 0.353	oscillating (3 reversals ≥ 0.01)

Variants: a stronger, faster response (gain 4, lag 2) also oscillates at B2 and B3. A slower legislature (lag 10) changes little, because Σ already scales adoption directly. Starting sentiment matters only for the first decade. At B4 a positive start (Σ = 0.7) gives a boom to 0.77 by year 5 and a bust to 0.11 by year 25. A negative start (Σ = 0.3) gives a lower, later peak (0.40 at year 10).

3.3 Readings

1. The legitimacy loop is a brake on adoption throughout the band. It lowers the K_Ω peak by 0.02 (B1) up to 0.24 (B4), and raises VFAC by 0.05–0.09 through more membrane rights, more explainability and less attention extraction.
2. At B4 the brake overshoots. Fast self-enhancement drives sentiment down faster than legislation follows, and the system oscillates: boom, backlash, partial recovery. The loop does not stop the inner loop smoothly. It turns a fast rise into a boom-bust cycle (BR-005).
3. The loop compresses the band. With the loop, B2–B4 end at nearly the same K_Ω (0.05–0.08) and VFAC (0.35–0.37). The adoption anchor decides the first decade; legitimacy decides the long run. This holds only if sentiment really tracks perceived benefit, which is the main empirical question for the checkpoints (L6).
4. Sentiment ends negative in every anchor, even B1. Part of this is not about organisms: the baseline drifts toward higher domination P, and Σ reads that as lost benefit. The reading "AI adoption causes backlash" is therefore not separated from "rising domination causes backlash" in this model. A decomposition of pb per term belongs to the prediction register (T06), so that observed sentiment can tell the two apart.
5. For the jurisdiction hypotheses (T04): a positive starting sentiment (J2-like) buys early adoption but not a better end state. A negative start (J1-like) costs early adoption and ends at about the same place. Under this loop, "lagging" is not worse in the long run. This depends on the closed-economy assumption (§ 2.4, reading 2).

3.4 Limitations

• Σ is a single national mood. There is no polarization, no split by cohort and no media dynamics.
• Legislation responds symmetrically and without capture. Domination P does not weaken the legislative response, although the political channel Ppol suggests that it should.
• The weights in pb and all loop parameters are not yet chosen by data. The readings are qualitative until the L6 sentiment series constrain them.

4. Jurisdiction archetypes (T04)

4.1 Settings

Archetypes are lever settings (research/ScenarioBandMethod.md § 4). Every archetype starts from the same anchor equilibrium. Its levers apply from t = 0, so differences come from policy, not from different starting states. perm (deployment permissiveness) is a new gadm04 lever, default 1. It scales the organism inflow.

Archetype	perm	m	expl	L / lin_open	x_att	other	initial Σ
J1 rights-first / precautionary	0.6	0.7	0.7	0.6 / 0.6	0.3	std 0.6	0.45
J2 deregulated fast-adopter	1.5	0.2	0.2	0.2 / 0.2	0.7	—	0.65
J3 platform-led	1.3	0.1	0.4	0.1 / 0.1	0.8	disc 0.6	0.5
J4 state-coordinated	1.3	0.2	0.4	0.2 / 0.2	0.5	managed sentiment: sig_gain 0.5, leg_lag 10	0.6

Lever magnitudes are choices that express the method's low/medium/high table, not estimates. Reproduce: cd simulations/gadm02 && python3 -B jurisdictions.py (results/jurisdictions.md).

4.2 Results

J1 has the highest VFAC at t 100 at every anchor, with and without the legitimacy loop. It is the only archetype that keeps domination below the P floor (P ≤ 0.65 against 0.79–0.89 for J2–J4).

J1 − J2 (positive = lagging does better):

Anchor	Loop	Δ mean Y* 0–25	Δ mean Y* 0–100	Δ VFAC t 100	Δ G	Δ A
B1	off / on	+0.000 / +0.012	+0.195 / +0.205	+0.32 / +0.29	+0.24 / +0.21	+0.14 / +0.06
B2	off / on	−0.024 / +0.024	+0.167 / +0.247	+0.32 / +0.27	+0.23 / +0.20	+0.15 / +0.03
B3	off / on	−0.038 / +0.047	+0.074 / +0.263	+0.25 / +0.23	+0.16 / +0.15	+0.19 / +0.03
B4	off / on	−0.044 / +0.069	−0.017 / +0.271	+0.18 / +0.19	+0.13 / +0.12	+0.17 / +0.02

Decomposition (each lever regime at perm 0.6 versus 1.5, rest unchanged): lagging by itself always costs early output (mean Y* 0–25 lower by 0.01 to 0.10) and raises VFAC by 0.01–0.13. Under J2 levers with the legitimacy loop on, speed hardly matters (|Δ VFAC| ≤ 0.01), because backlash slows adoption anyway.

4.3 Readings

1. The archetype gap comes mainly from rights, not from speed. The J1–J2 VFAC gap is 0.18–0.32. The speed component alone is 0.01–0.13. The owner's question "is lagging better or worse?" therefore splits: lagging as such costs output early and helps agency and attention a little. What decides the long run is membrane rights, open lineages and attention protection.
2. "Europe lags" (J1-like): the model says lagging costs early output when there is no backlash (B2–B4, loop off: −0.02 to −0.04 mean Y*). It does not make the EU worse off in the long run at any anchor. This rests on the rights levers and on the closed-economy assumption: a J1 jurisdiction that imports concentrated foreign lineages does not get J1's L and lin_open.
3. "Argentina may thrive" (J2-like): a J2 regime is richer early only without the legitimacy loop and only at B2–B4 (+0.02 to +0.04 mean Y* 0–25). With the loop, the positive start turns into backlash (Σ 0.65 → 0.21–0.29). J2 then ends poorer than J1 at every anchor and breaks the domination floor. "Thriving" in the model requires sentiment that stays positive, which requires broad claims and contained domination. That is what checkpoint indicators L6 (sentiment) and F2 (concentration) can test.
4. Managed sentiment (J4) sustains adoption but not VFAC. With the loop on, J4 keeps the highest K_Ω (0.38 at B4, t 100 against 0.08 for J2), because its weak and slow sentiment response avoids backlash. Its domination stays at P ≈ 0.82, and VFAC is among the lowest.
5. Long-run prosperity follows domination. Y* falls in J2–J4 because domination lowers effective demand and selected value in GADM. The reading "deregulation is poorer in the long run" depends on that mechanism (GADM-0.2 § auxiliaries) and should be read as conditional on it.

4.4 Mapping real jurisdictions (provisional)

Policy evidence as known up to 2025. The L7 legislation collection at each checkpoint updates it, and the mapping is revisited at every full checkpoint.

Jurisdiction	Nearest archetype	Evidence	Open questions
EU	J1, with moderate perm	AI Act (Reg. 2024/1689) in force since Aug 2024, prohibitions from Feb 2025, general-purpose AI duties from Aug 2025; DMA; Data Act applicable from Sep 2025. Digital-omnibus proposals (late 2025) to delay parts of the high-risk regime point to drift toward higher perm.	lineage imports: EU adoption runs largely on non-EU lineages, so effective L and lin_open are lower than its rules suggest
US	J3 (market structure) with J2 federal stance	Jan 2025 executive order revoking EO 14110; AI Action Plan (Jul 2025); no federal AI statute. Several state laws (e.g. Colorado, California) apply rights levers locally.	federal–state split: one archetype per level may be needed
Argentina	J2 by stance	Deregulation programme (DNU 70/2023, Ley Bases 2024 with the RIGI investment regime); no comprehensive AI law as of 2025.	adoption level unknown (L3 manual); AR adopts foreign lineages, so its L and lin_open are set abroad; initial sentiment unmeasured (L6)
(reference) China	J4	state-coordinated deployment, algorithm registration and generative-AI rules	not a tracked jurisdiction; anchors J4 only

Mapping update 2026-10-13 (review, L7): the EU's drift toward higher permissiveness is now law: the AI Omnibus (Reg. 2026/1744, in force 27 July 2026) delays Annex III high-risk duties to December 2027, while transparency and GPAI enforcement started in August 2026. The EU stays J1 for rights levers, with perm moving up. The US federal stance moved further toward J2 (EO 14365, December 2025: preemption of state AI laws), but no federal statute exists, and state rights levers (California SB 53, Colorado) remain. US: J3 market structure with a J2 federal stance, unchanged as a label but stronger. Argentina: J2 confirmed. No AI law; the executive promotes AI-operated companies ("Sociedad Automatizada" bill, May 2026). Sources: checkpoints/2026/2026-10-13-review.md.

4.5 Limitations

• Closed economy: no lineage import, trade or capital flows between jurisdictions. This is the largest gap for both hypotheses.
• One lever setting per archetype. No lever trajectories (for example the EU loosening over time).
• The output advantage of fast adoption is short-lived in the model partly by construction (domination feeds back on demand); a model with export-led gains from organisms could reverse reading 3.

5. Global sensitivity per band anchor (T05)

gsa.py --model gadm04 --anchor Bn runs Morris screening and Sobol indices for final VFAC around each anchor (its parameters, equilibrium start and recalibrated channel counters). The legitimacy-loop and permissiveness parameters are held fixed. gsa_band.py compares the four runs (results/gsa_band.md; per anchor results/gsa_gadm04_B1..B4.md). About 15 minutes per anchor with --jobs 8.

Scenario	B1	B2	B3	B4
Baseline: top 3 by ST	dO 0.54, aO 0.15, q0 0.11	IX 0.26, dO 0.17, lX 0.12	IX 0.26, dO 0.21, lX 0.09	dO 0.21, IX 0.21, ell 0.17
Generative regime: top 3 by ST	dO 0.26, m 0.16, r_lin 0.16	q0 0.24, dO 0.16, IX 0.13	expl 0.21, IX 0.16, q0 0.13	expl 0.27, IX 0.16, std 0.08

Readings

1. The dominant drivers change across the band. At B1, organism parameters dominate: depreciation dO alone has ST 0.54. How fast AI capability depreciates matters most when there is little of it. From B2 on, commons investment IX leads the baseline. At B4, learning ell and attention (arec, x_att) enter the top five.
2. Under fast adoption, governance levers lead the generative regime. Explainability expl is the largest driver at B3 and B4 (ST 0.21, 0.27). Membrane rights m rank first in Morris screening at every anchor. Lineage self-reinforcement r_lin matters only at B1–B2.
3. Attention protection gains weight with speed. x_att moves from outside the Morris top 15 (B1) to rank 5 (B4, baseline).
4. Investment speed aO matters only at B1. At faster anchors, adoption is limited by learning and depreciation, not by investment.
5. Policy relevance. The calibrated present lies between B2 and B3 (models/Calibration.md § 3). In that range, the leading drivers are policy-relevant: commons investment IX and explainability expl. The most uncertain technical parameter is depreciation dO, which grounding (WP-0008) has narrowed to 0.15–0.35.

Caveat: these runs vary parameters around the uncalibrated anchors, with domination starting at P ≈ 0.37. The calibrated start has P ≈ 0.6; GSA around it is ASSEMBLY-WP-0015-T04.