assembly-economics/models/GADM-0.5.md

# GADM-0.5 (in development)

**Status:** dev (2026-10-13), ASSEMBLY-WP-0007. Scope:
`models/GADM-Sectors-Scope.md`. Module: `simulations/gadm02/gadm05.py`
(gadm04 plus base-model parameters with defaults that reproduce GADM-0.4).

## 1. Rival and non-rival output (T01)

### 1.1 Structure

    K_r = 1 − (1 − K_phys)(1 − φ_r K_Ω)        rival capacity (physical goods)
    Y_r = K_r √E                                rival output, limited by ecological headroom
    Y_n = (1 − φ_r) K_Ω (1 − K_phys)            non-rival output (information, organism services)
    Y*  = Y_r + Y_n                             equals K at E = 1 for every φ_r

Non-rival output flows into commons X as far as it is openly inheritable:
`dX += xn · Y_n · lin_open · (1 − X)`. The ecological draw is unchanged
(physical capital `e1 K_phys dem thr` plus compute energy `e2 K_Ω`).
φ_r = 1 and xn = 0 reproduce GADM-0.4 exactly.

Priors at first: φ_r = 0.2, xn = 0.3. Grounded values (§ 6): φ_r = 0.15
(range 0.05–0.25), xn = 0.6. The tables in §§ 1–4 were computed with the
priors; § 6 lists what changes.

### 1.2 Results across the band (`nonrival.py`, `results/nonrival.md`)

| Reading | B1 | B2 | B3 | B4 | class |
|---|---|---|---|---|---|
| ΔVFAC, GADM-0.5 vs. 0.4, baseline | +0.015 | +0.026 | +0.039 | +0.049 | quantitative shift |
| of which from the commons feed alone | +0.016 | +0.026 | +0.039 | +0.049 | quantitative shift |
| ΔE, baseline | −0.011 | −0.020 | −0.026 | −0.025 | quantitative shift |
| ΔVFAC, generative regime | +0.002 | +0.003 | +0.003 | +0.007 | quantitative shift |
| ΔE, generative regime | 0.000 | 0.000 | 0.000 | −0.002 | — |
| generative − baseline VFAC (GADM-0.5) | +0.274 | +0.308 | +0.303 | +0.266 | robust |

Non-rival share of output at t 100: baseline 0.12 (B1) to 0.66 (B4);
generative regime 0.08 to 0.31.

### 1.3 Readings

1. **The non-rival effect works through the commons, not through the output
   split.** The whole VFAC gain comes from non-rival output feeding X. Calling
   output non-rival does not change VFAC, because VFAC measures stocks, not
   output.
2. **The generative regime's ecological cost does not shrink** (scope test 2:
   no). Its low ecological headroom (E 0.48 against 0.68–0.82 in the baseline)
   comes from the physical-capital and compute-energy draw, which a non-rival
   output share does not reduce. In the baseline, the commons feed even
   lowers E slightly, because more commons raise activity and throughput.
3. **No genuine throughput decoupling** (scope test 1: no). VFAC rises while
   throughput falls only in the baseline at B3–B4 (20–40 % of five-year
   windows), and there only because organism capability peaks and declines,
   taking its compute draw with it. The generative regime shows no decoupling.
4. **What genuine decoupling would need:** non-rival output that *replaces*
   rival consumption (dematerialization of demand), so that demand for physical
   throughput falls as Y_n grows. GADM has no demand composition. This is the
   question for the ecological classes (T02) and, if needed, sectors (T04).
   It is recorded as BR-010.
5. **The generative regime's advantage is robust across the band** (+0.27 to
   +0.31), slightly smaller than in GADM-0.4 (+0.29 to +0.34), because the
   baseline also gains from the commons feed.

## 2. Ecological resource classes (T02)

### 2.1 Structure

`gadm05.simulate_eco` replaces the single E with three classes, each with
its own floor (0.20). The aggregate E used for output and VFAC is their
geometric mean.

| Class | Covers | Start (prior from indicators) | Regeneration |
|---|---|---|---|
| Er renewables | energy, land, freshwater, biosphere | 0.15 (all three land/water/biosphere boundaries breached, PHC 2025) | yes, at rest at start |
| Ed materials | depletable stocks | 0.45 (EU footprint 13.7 t/capita vs. a sustainable level of about 6–8 t) | none, or replenishment at rest (reserve growth) |
| Es sinks | climate, ocean acidification, novel entities, nutrient flows | 0.33 (2 of 6 sink-type boundaries safe) | yes, at rest at start |

Physical load (K_phys · dem · thr) draws 40/30/30 % on renewables, materials
and sinks. Compute load (K_Ω) draws 30/20/50 %. The sink load is scaled by
(1 − low-carbon electricity share, `lowc` 0.40, US-like). Recycling (`circ`
0.12, the EU circular material use rate) slows material depletion. These
shares are priors.

### 2.2 Results (`eco_classes.py`, `results/eco_classes.md`)

| | Strict material depletion | Material replenishment at rest |
|---|---|---|
| Renewables | below floor at start, every scenario | below floor at start, every scenario |
| Materials | cross the floor at t 25–70 (generative regime first, t 25–30) | never cross |
| Sinks | generative regime: t 29 (B4) to 83 (B1); others: never | same |
| Lowest class at t 100 | materials | renewables |

Organism growth (baseline, B4 − B1, relative change t 0 → 50): sinks
−0.036, renewables −0.032, materials −0.018.

Low-carbon electricity share (generative regime, replenishment at rest):
the sinks breach moves from t 49 / 37 / 29 (B2 / B3 / B4, share 0.4) to
t 87 / 66 / 52 (share 0.7, EU-like). At 0.9 there is no breach within 100
years.

### 2.3 Readings

1. **Grounding test passed:** organism growth binds sinks and renewables
   (energy) before materials (`ParameterGrounding.md` § 3).
2. **The renewables class is already beyond its floor at the start.** The
   single E (0.235) hid that one class is binding today while the others
   still have headroom. Floors must be checked per class.
3. **Which class binds next depends on an unmeasured assumption.** With
   strict depletion, materials bind first in every scenario. If reserve growth
   keeps pace with use, they never do. Material replenishment needs grounding
   (reserve-to-production ratios over time).
4. **The generative regime's ecological cost sits in the sinks.** It is the
   only scenario that breaches the sink floor, and faster adoption brings the
   breach forward (t 83 → 29). This makes BR-010 more specific: the regime's
   problem is the sink load of its physical and compute activity, not output
   as such.
5. **Low-carbon electricity is the decisive lever for organism-driven sink
   load.** An EU-like share (0.7) delays the breach by 23–38 years; 0.9 avoids
   it within the horizon. The effect is conservative: sink regeneration is
   fitted so that each share starts at rest, so the lever acts only on the
   growth of the load.

## 3. Endogenous selection weights (T03)

`gadm05.simulate_sel`: the five selection weights (market, science,
community, public, organism) relax at 0.05 per year toward targets pushed by
domination and civic capacity:

- **Domination:** market concentration Pm raises the market weight, and
  discovery concentration Pd raises the organism weight (strength 2).
- **Civic capacity:** epistemic capacity Q raises science, and commons X
  raise community (strength 1).

**Capture** is the share of selection weight held by the market and organism
channels. Results are in `selection.py` and `results/selection.md`.

| Reading (endogenous − fixed weights, t 100) | B1 | B2 | B3 | B4 | class |
|---|---|---|---|---|---|
| capture, attention oligopoly | +0.114 | +0.131 | +0.138 | +0.141 | robust |
| capture, baseline | +0.023 | +0.109 | +0.115 | +0.115 | quantitative shift |
| capture, generative regime | −0.240 | −0.211 | −0.189 | −0.161 | robust |
| ΔVFAC, any scenario | ≤ 0.002 in absolute value | | | | no effect |

Largest change in the weights themselves: 0.0044 per year.

Readings:

1. **Domination captures selection, and the generative regime de-captures
   it.** Both are robust across the band. Capture is slow and incremental, so
   the grounding test passes: no flip within a few years.
2. **Capture of selection is not a material channel for VFAC in GADM.** The
   five quality signals are of similar size, so moving weight between them
   hardly changes selection quality σ. A plural selection mix also does not
   make the generative regime more stable against a domination shock.
3. **What would make it material:** captured channels that deliver
   systematically worse quality. An example is organism selection under
   discovery concentration, `O = Q (1 − disc)`, which falls only through the
   lever `disc`, not through the discovery channel Pd. Linking O to Pd is the
   candidate change. It is recorded here and not implemented, because no
   reading depends on it yet.

## 4. Two-good demand split: dematerialization (T04)

### 4.1 Structure

`gadm05.simulate_eco(..., demand=...)` adds a slow state Cr, the rival share
of demand relative to its start (Cr0 = 1). Cr relaxes (rate `dm_lam`) toward
`1 − dm_eps · (qn − qn0)`, where `qn = (Y_n / Y*) · √(X Q)` is the
quality-weighted non-rival offer. Physical throughput intensity `thr` and
physical investment `aK` scale with Cr. The same capital then serves
dematerialized demand with less physical load. `dm_lam = 0` switches it off.
Priors: λ 0.05 per year, ε 2; strong case λ 0.1, ε 4.

### 4.2 Results (`demand.py`, `results/demand.md`; material replenishment at rest)

| Generative regime | B1 | B2 | B3 | B4 |
|---|---|---|---|---|
| decoupled decades, no split → split → strong | 0 → 0 → 0 | 0 → 0.45 → 0.89 | 0 → 0.71 → 0.75 | 0 → 0.22 → 0.34 |
| sinks breach, no split → split → strong | t 83 → none → none | t 49 → 98 → none | t 37 → 61 → none | t 29 → 40 → none |
| VFAC t 100, no split → split | 0.607 → 0.613 | 0.583 → 0.598 | 0.564 → 0.585 | 0.509 → 0.533 |

A *decoupled decade* has VFAC rising, physical throughput falling and K_Ω not
falling, which excludes the artifact of § 1.3. Baseline: decoupled decades
0–0.09, and no sinks breach with or without the split.

Calibration check: see § 5. (A first comparison here used total throughput
per output, which is the wrong operator.)

### 4.3 Readings

1. **Genuine decoupling needs demand composition.** With the split, the
   generative regime decouples welfare from throughput in most decades at
   B2–B3, while organism capability keeps growing. BR-010 is resolved
   structurally: non-rival output must replace rival consumption, not only add
   to it.
2. **Decoupling is band-dependent (qualitative break).** It needs a non-rival
   offer of sufficient quality, which B1 adoption does not build. At B4,
   physical and compute growth outpace substitution, and decoupling is rarer
   (22 %).
3. **The generative regime's sinks breach becomes avoidable.** The default
   split moves it 15–50 years later. The strong split removes it within the
   horizon at every anchor. Together with low-carbon electricity (§ 2.3), this
   is the second lever on the regime's ecological cost.
4. ~~The readings are conservative against data.~~ **Corrected in § 5:**
   that comparison used total throughput per output, which includes compute
   energy. On the right operator (material draw per output), the model already
   dematerializes as fast as observed. The decoupling reading then depends on
   the rival share φ_r, and the EU data do not support it.

## 5. Dematerialization against data (ASSEMBLY-WP-0018-T01)

**Data:** Eurostat resource productivity (`env_ac_rp`, GDP per kg of
domestic material consumption, chain-linked 2015 EUR). It rose 1.85 % per
year in the EU and 2.73 % per year in Germany over 2010–2024. Material
intensity fell at the same rates.

**Operator:** the material draw per unit of output in the first decade from
the calibrated start. Total throughput per output is the wrong operator,
because it includes compute energy, which grows with adoption.

| Jurisdiction | Observed | Model, no split (φ_r 0.2) | Model, split ε 2 (φ_r 0.2) | φ_r that reproduces the data (split ε 2) |
|---|---|---|---|---|
| EU | −1.85 %/y | −2.45 %/y | −2.89 %/y | **0.66** |
| DE | −2.73 %/y | −2.26 %/y | −2.64 %/y | **0.15** |

No efficiency trend is needed: non-rival organism output that needs no
materials already makes output grow faster than material use. The data
constrain the rival share of organism output φ_r, given the demand split. The
two parameters are not separately identified by one series.

Generative regime, demand split at its default (`dematerialization.py`,
`results/dematerialization.md`):

| φ_r | decoupled decades B1 / B2 / B3 / B4 | sinks breach B2 / B3 / B4, no split → split |
|---|---|---|
| 0.15 (fits DE) | 0 / 0.55 / 0.72 / 0.24 | 49 → none / 37 → 67 / 29 → 41 |
| 0.20 (prior) | 0 / 0.45 / 0.71 / 0.22 | 49 → 98 / 37 → 61 / 29 → 40 |
| 0.66 (fits EU) | 0 / 0 / 0 / 0.02 | 49 → 58 / 37 → 43 / 29 → 33 |

Readings:

1. **The decoupling reading of § 4 depends on the non-rival share of organism
   output.** It holds when organism output is mostly non-rival (φ_r ≤ 0.2, the
   task-data prior and the German fit). It disappears when it is mostly rival
   (φ_r 0.66, the EU fit). Recorded as BR-011.
2. **The fit is weakly identified.** The 2010–2024 dematerialization
   happened mostly before generative AI, while the model attributes it to
   organism output and demand composition. Structural change, the energy
   transition and offshoring of material-intensive production are not in the
   model. The first decisive data are post-2024 values of E5, now wired into
   the full checkpoint.
3. **Practical reading:** whether AI-era growth decouples welfare from
   ecological load is an empirical question about *what organisms produce*.
   Information services (non-rival) support decoupling. Automation of physical
   production (rival) does not. The checkpoints should watch φ_r through L4
   (task shares) and E5 (material intensity) together.

## 6. Grounding of the non-rival parameters (ASSEMBLY-WP-0018-T03)

**φ_r, the rival share of organism output.** Anthropic Economic Index: AI
use concentrates in information occupations:

- computer and mathematical: 37 % (February 2025) and 32 % (Claude.ai, 2026
  wave), against 3.4 % of workers;
- arts and media: 10 %;
- education: 9–14 %;
- office: 8 %;
- transportation: 0.3 % (9 % of workers).

The direct share of physical-output occupations is at most about 0.05. Adding
AI use that indirectly supports physical production (engineering, logistics,
management) gives **φ_r 0.05–0.25, central 0.15**.

**xn, the commons feed.** In 2025, 18.5 % of GitHub contributions went to
public repositories (Octoverse 2025: 81.5 % private). Taken as the open share
of non-rival output and divided by lineage openness (0.3): **xn ≈ 0.6**. This
depends on `lin_open`. For a J3-like jurisdiction (lin_open 0.1), the same
measured share would mean xn ≈ 1.8.

**Effect on the readings** (all re-run with the new defaults; tables in
`results/`):

| Reading | Priors (φ_r 0.2, xn 0.3) | Grounded (φ_r 0.15, xn 0.6) |
|---|---|---|
| non-rival VFAC gain, baseline B1–B4 | +0.015 to +0.049 | +0.038 to +0.077 |
| non-rival VFAC gain, generative regime | ≤ +0.007 | ≤ +0.011 |
| decoupled decades with demand split, generative B2 / B3 / B4 | 0.45 / 0.71 / 0.22 | 0.69 / 0.71 / 0.25 |
| class most hit by organism growth | sinks > renewables > materials | unchanged |
| selection capture (attention oligopoly / generative) | +0.11–0.14 / −0.16 to −0.24 | +0.11–0.13 / −0.16 to −0.24 |

**Consequence for BR-011.** Task data support φ_r ≤ 0.25, which is
inconsistent with the EU dematerialization fit (0.66). The EU's slower
dematerialization since 2010 is therefore unlikely to come from rival AI
output. It points to non-AI factors that GADM lacks: industrial structure,
the energy mix of material processing, offshoring. With the task-data
φ_r, the decoupling reading holds. The EU fit remains the warning case, if
future AI use shifts toward automating physical production.