assembly-economics/models/GADM-0.2.md

<!--
Created 2026-10-12 under ASSEMBLY-WP-0004 (T03 decision, T04 specification,
T05/T06 findings). Implemented by simulations/gadm02/gadm02.py; the code is
the reference if the two ever disagree, and any divergence is a bug.
-->

# Generative Assembly Dynamics Model — GADM-0.2

**Status:** exploratory hypothesis generator. Not calibrated to any real
economy. All numbers are in normalized, unitless model space.

GADM-0.2 extends GADM-0.1 (`GenerativeAssemblyDynamicsModel.md`) in three
ways:

1. It specifies every flow with signs (see `GADM-CLD.md`).
2. It adds the information-organism extensions selected below.
3. It is implemented as a runnable toy simulation.

---

## 1. Decision: which organism extensions enter GADM-0.2 (T03)

Candidates from `InformationOrganismsInGADM.md` § 4:

| Candidate | Decision | Rationale |
|---|---|---|
| Split K into K_phys + K_Ω | **in** | Learning with use (R7) and lineage control (R8) are the main new loops. Without the split, they cannot be expressed. |
| Attention headroom A | **in** | Reference scarce resource of the program. Needed to test whether VFAC without A misses attention harm. |
| Organism selection channel O | **in** | Cheap to add. Needed to examine hypothesis M3 (O's quality and its contribution to P share a cause). |
| Membrane control and lineage concentration in Ṗ | **in**, as levers `m` and `L` | Main institutional levers of the organism lens. |
| Participation claims C_p | **in**, as lever `cp` | Lets scenarios test a post-work claim channel. It is treated as a share of capital return, which is a simplification (see § 6). |
| Compute C as its own stock | **deferred to 0.3** | Represented for now as a direct draw of K_Ω on E (`e2`). A separate stock only pays off once energy classes are modeled (GADM-0.4). |
| Age cohorts | **deferred to 0.3** | See `GADM-0.3-Candidates.md`. |

---

## 2. State

Twelve stocks, all in [0, 1]:

| Symbol | Code | Stock | New in 0.2 |
|---|---|---|---|
| H | `H` | Human capability | |
| X | `X` | Assembly commons | |
| K_phys | `Kp` | Physical / conventional productive capacity | split |
| K_Ω | `Ko` | Information-organism capability | ✓ |
| F | `F` | Risk capacity | |
| G | `G` | Effective agency | |
| N | `N` | Diversity & composability | |
| Q | `Q` | Epistemic capacity | |
| P | `P` | Domination concentration | |
| E | `E` | Ecological headroom | |
| S | `S` | Social cohesion | now a stock relaxing to a target |
| A | `A` | Attention headroom: unclaimed, self-directed attention | ✓ |

**Bounding convention.** Every stock follows
\(\dot x = \text{inflow}\cdot(1-x) - \text{outflow}\cdot x\), so stocks stay in
[0, 1] without clipping. Clipping is applied only as a numerical guard.

**Institutional levers** (scenario axes, all in [0, 1]):

| Lever | Meaning |
|---|---|
| `m` | membrane control held by users and communities rather than operators |
| `L` | lineage diffusion: 0 = concentrated, 1 = diffuse ownership of organism lineages |
| `lin_open` | share of lineages that are openly inheritable (commons rather than enclosure) |
| `std` | open standards and interoperability |
| `disc` | discovery concentration |
| `expl` | explainability and provenance enforcement |
| `x_att` | attention regime: 0 = protected, 1 = extraction-funded |
| `auto`, `dshare`, `cp`, `od` | automation intensity, social dividend share, participation claims share, ownership diffusion of private returns |

---

## 3. Equations

### 3.1 Auxiliaries

\[
K = 1-(1-K_{phys})(1-K_\Omega), \qquad Y^* = K\sqrt{E}
\]

Claims (GADM-0.1 § 9, extended with C_p):

\[
w_L = w_{L0}(1-auto\cdot K_\Omega),\quad c = 1-w_L,\quad
\Pi_p = c\,(1-dshare-cp)
\]
\[
B = w_L + c\,(dshare+cp) + \Pi_p\cdot od
\]

\(B\) is the **breadth of claims**: the share of output reaching the broad
population as wages, dividends, participation claims or diffused ownership.
Demand factor: \(dem = (0.4+0.6B)(0.6+0.4S)\).

Experimentation (GADM-0.1 § 5, plus a Q-gated organism term, hypothesis M2):

\[
V = \min\!\left(1,\; v_0 H^{.5}X^{.5}F^{.3}G^{.4}N^{.3}Q^{.3}(1-P)^{.5} + v_o K_\Omega Q\right)
\]

Selection (GADM-0.1 § 6, plus organism channel O weighted by prevalence):

\[
M=\sqrt{A\cdot dem},\quad O = Q(1-disc)
\]
\[
\sigma = \frac{w_m M + w_s Q + w_c N(1-P) + w_p S + w_o K_\Omega O}{w_m+w_s+w_c+w_p+w_o K_\Omega},
\qquad V_{sel} = V\sigma
\]

Reuse: \(\rho = \min(1, \rho_0 + \rho_O K_\Omega\, lin\_open)\).

### 3.2 Stocks

\[
\begin{aligned}
\dot X &= (a_X\rho V_{sel} + I_X)(1-X) - (\delta_X + l_X P + l_{Xo}K_\Omega(1-lin\_open))X\\
\dot K_{phys} &= a_K V_{sel}\,dem\,(1-K_{phys}) - \delta_K K_{phys}\\
\dot K_\Omega &= (a_O V_{sel}\,dem + \ell K_\Omega\,dem\,Q + o_{seed})(1-K_\Omega) - \delta_O K_\Omega\\
\dot H &= (h_0(\tfrac{S+B}{2}) + h_A A + h_T X)(1-H) - (\delta_H + desk\cdot K_\Omega(1-m))H\\
\dot F &= (f_0 + f_Y Y^* B)(1-F) - \delta_F F\\
\dot G &= (g_1B + g_2X(1-P) + g_3F + g_4K_\Omega m)(1-G) - (\delta_G + b_1P + b_2K_\Omega(1-m))G\\
\dot N &= (n_1V(1-P) + n_2X\,std)(1-N) - (\delta_N + n_3P + n_4K_\Omega(1-L))N\\
\dot Q &= (q_0 + q_1X + q_oK_\Omega\,expl)(1-Q) - (\delta_Q + q_2P + q_xK_\Omega(1-expl) + q_3\max(0,V-Q))Q\\
\dot P &= (p_1V_{sel}(1-B) + p_2\,disc + p_3K_\Omega(1-m) + p_4K_\Omega(1-L))(1-P) - (p_5QS + p_6X\,std + p_7B)P\\
\dot E &= e_{reg}(1-E) - (e_1K_{phys}\,dem\,thr + e_2K_\Omega)E\\
\dot S &= s_{rate}\left(G^{.4}(1-P)^{.3}B^{.2}A^{.1} - S\right)\\
\dot A &= a_{rec}(0.4+0.6B)(1-A) - a_{ext}\,x\_att\,(0.2+K_\Omega)A
\end{aligned}
\]

Parameter values are in `simulations/gadm02/gadm02.py` (`PARAMS`, `INIT`).

### 3.3 Diagnostics

\[
VFAC = \left(H\,X\,G\,N\,Q\,S\,E\,(1-P)\,A\right)^{1/9},\qquad
VFAC_{0.1} = \left(H\,X\,G\,N\,Q\,S\,E\,(1-P)\right)^{1/8}
\]

These use equal exponents. VFAC is 0 if any factor is 0. As in GADM-0.1, VFAC
is a research instrument, not a welfare function.

Constitutional floors, checked separately: \(G \ge 0.2,\; E \ge 0.2,\; P \le 0.7,\; A \ge 0.2\).

### 3.4 Calibration

**Time unit:** one period = one year (ASSEMBLY-WP-0005-T01), so all rates are
per year and the horizon is 100 years. The implied timescales are mostly
plausible: commons obsolescence about 25 years (`dX`), physical capital about
20 years (`dK`), cohesion adjustment about 5 years (`srate`). Organism
capability, at about 17 years (`dO`), is the main exception (§ 5.1).

There is no empirical calibration. Three coefficients were adjusted so that
the **baseline** stays near its initial state, instead of trivially collapsing
or saturating:

- `v0` = 2.5: the seven-factor product is about 0.14 at initial values.
- `aK` = 0.35 and `n1` = 0.30: otherwise K_phys and N decay to about 0.12.

Everything else is a first guess, chosen so that rates fall roughly in 0.03–0.3
per period.

---

## 4. Scenario results (T05)

Full table: `simulations/gadm02/results/summary.md`. Scenario definitions:
`scenarios/`. Final values at t = 100:

| Scenario | VFAC | VFAC_0.1 | P | G | A | B | Notes |
|---|---|---|---|---|---|---|---|
| Baseline | 0.46 | 0.44 | 0.52 | 0.44 | 0.65 | 0.59 | slow erosion of X and N, rising P |
| Automated capitalism | 0.32 | 0.29 | **0.73** | 0.28 | 0.58 | 0.33 | P ceiling breached in 69 % of periods; demand and cohesion fall |
| Dividend abundance | 0.45 | 0.43 | 0.54 | 0.45 | 0.62 | 0.69 | demand stabilized, but VFAC ≈ baseline |
| Universal venture access | 0.49 | 0.47 | 0.52 | 0.57 | 0.64 | 0.58 | agency up via restartability |
| Commons-rich | 0.54 | 0.54 | 0.48 | 0.47 | 0.62 | 0.57 | X 0.27 → 0.66; second best overall |
| Attention oligopoly | 0.37 | 0.36 | 0.69 | 0.38 | 0.55 | 0.61 | P approaches the ceiling; N collapses to 0.17 |
| Generative regime | **0.65** | 0.64 | 0.31 | 0.64 | 0.78 | 0.70 | stable; lowest ecological headroom (E 0.48) |
| — with concentrated lineages | 0.53 | 0.50 | 0.50 | 0.46 | 0.79 | 0.71 | loses about two thirds of the gain over baseline |
| — with attention extraction | 0.58 | 0.59 | 0.41 | 0.60 | 0.45 | 0.71 | VFAC_0.1 understates the loss |

### Readings

These describe the model's behavior under its assumed signs. They are
candidate hypotheses, not claims about economies.

1. **Abundance without breadth fails** (automated capitalism). Output Y* is
   close to the baseline (0.50 vs. 0.53), but domination crosses its ceiling
   and claims breadth B drops from 0.59 to 0.33. This reproduces GADM-0.1's primary question in
   model form.
2. **Dividends stabilize demand, not agency.** Dividend abundance raises B
   from 0.59 to 0.69, but G, P and VFAC barely move. In this model, claims
   alone do not counter domination or commons erosion.
3. **Reuse matters through large shifts, not interaction.** The
   commons-rich regime is the strongest single-lever regime, but `rho0` alone
   is not a top driver (rank 20 of 80 one-at-a-time). The original reading,
   "reuse works as a bundle", was **refuted** by the global analysis (§ 5.1):
   the bundle shows almost no interaction. The scenario's effect comes from
   moving I_X, ρ_0 and `lin_open` by 100–200 %, far outside local
   sensitivity ranges.
4. **Organism governance dominates the generative regime.** Concentrating
   lineages (`L`, `lin_open`, `m` low) costs about two thirds of the
   generative regime's VFAC gain over baseline (0.65 → 0.53, baseline 0.46).
   An extraction-funded attention regime costs about 40 % (0.65 → 0.575).
5. **VFAC without attention understates attention harm.** Moving from the
   generative regime to its attention-extraction variant, VFAC falls by 0.075,
   but VFAC_0.1 falls by only 0.042 (0.636 → 0.594). In the extraction
   variant, VFAC_0.1 even ranks above VFAC, the reverse of every other
   scenario. This supports including A, and also the Goodhart warning in
   `InformationOrganismsInGADM.md` § 4.1.
6. **The generative regime has an ecological cost.** It has the lowest E of
   all regimes. It stays above the floor, but `ereg` ranks third in its
   sensitivity. Abundance needs throughput decoupling (GADM-0.1 § 14) to stay
   generative.

---

## 5. Sensitivity (T06)

Full tables: `simulations/gadm02/results/sensitivity.md`
(one-at-a-time, ±20 %, effect on final VFAC).

| Rank | Baseline | Generative regime |
|---|---|---|
| 1 | `dO` organism depreciation (+) | `m` membrane control (+) |
| 2 | `ell` organism learning (−) | `L` lineage diffusion (+) |
| 3 | `p7` claims breadth checking P (+) | `ereg` ecological regeneration (+) |
| 4 | `wL0` labor share (+) | `expl` explainability (+) |
| 5 | `IX` commons investment (+) | `v0` experimentation (+20 % → −) |

Readings:

- **Under baseline governance, faster organism learning lowers VFAC.** Higher
  `ell` and lower `dO` both reduce VFAC: K_Ω grows under operator-controlled
  membranes (`m` = 0.3) and concentrated lineages, and this feeds P, lock-in
  and deskilling. In the generative regime, the same organism growth is
  beneficial, and the governance levers `m` and `L` become the top drivers.
  This is the model form of hypothesis M3. Caveat: it follows directly from
  the assumed moderation of links #39–#41 by `m` and `L`. The model shows that
  these assumptions matter most, not that they are true.
- **More experimentation can hurt** in the generative regime (+20 % `v0` →
  lower VFAC). This happens through the epistemic brake B3 (link #33, a
  contested link).
- One-at-a-time analysis cannot detect interactions. The global analysis
  in § 5.1 checks for them.

### 5.1 Global sensitivity (ASSEMBLY-WP-0005-T01)

Full tables: `simulations/gadm02/results/gsa.md`. Method: Morris screening
over all 80 parameters (20 trajectories), then Sobol indices for the top 10
plus the reuse bundle. Parameters vary ±50 %, N = 256, with bootstrap 95 %
intervals. S1 uses the Saltelli (2010) estimator with centred outputs; ST
uses the Jansen estimator. The uncentred S1 estimators were tried first and
were unusable (VFAC's mean is large relative to its spread).

| | Baseline | Generative regime |
|---|---|---|
| Top ST | `dO` 0.34, `p7` 0.13, `ell` 0.10, `wL0` 0.07, `IX` 0.07 | `m` 0.27, `L` 0.13, `ereg` 0.13, `p7` 0.07, `expl` 0.06 |
| ΣS1 (analysed set) | 1.04 | 1.06 |
| Reuse bundle ΣS1 / ΣST | 0.09 / 0.11 | 0.08 / 0.09 |

Readings:

1. **The model is nearly additive** within ±50 %. ST ≈ S1 for every analysed
   parameter, and ΣS1 ≈ 1. Interactions are small, so the one-at-a-time
   rankings in § 5 are confirmed.
2. **The "reuse as a bundle" reading is refuted.** The bundle explains about
   10 % of variance, with negligible interaction (ΣST − ΣS1 ≈ 0.01–0.02).
   Reading 3 in § 4 has been corrected.
3. **Organism governance remains the dominant lever in the generative
   regime.** Membrane control `m` alone explains roughly 27–32 % of variance,
   and lineage diffusion `L` a further 13–16 %. Reading 4 in § 4 holds
   globally.
4. **Calibration priority: `dO`.** Organism-capability depreciation explains
   about a third of baseline variance. At 0.06 per year it implies a 17-year
   lifetime, which is plausibly too long for AI systems. It should be
   calibrated first.

---

## 6. Known limitations

- No units, no calibration. Comparisons between scenarios are meaningful;
  absolute values are not.
- Every structural sign comes from `GADM-CLD.md`. The contested links (#33,
  #40, #41, #46, #49) drive several readings.
- Participation claims `cp` are a fixed share of capital return. A real
  mechanism would tie them to contributions (attention, judgment, governance),
  which is WS-14 work.
- A single population with no cohorts, sectors or regions.
- Attention dynamics are weak: in the attention oligopoly, A falls only to
  0.55. `aext` may be underestimated. *Resolved in GADM-0.3* (`GADM-0.3.md`
  § 4.1: `aext` recalibrated to 0.275 against time-use evidence).
- Euler integration, dt = 0.1. No stiffness issues were observed at these
  rates.

## 7. Reproduce

```bash
python3 -m unittest discover -s simulations/gadm02
cd simulations/gadm02 && python3 run_scenarios.py && python3 sensitivity.py
cd simulations/gadm02 && python3 gsa.py          # ~5 min; --quick for a smoke run
```