coulomb/assembly-economics
5
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
assembly-economics/models/ParameterGrounding.md
tegwick 8dd9e97d33 ASSEMBLY-WP-0007-T02: ecological resource classes (renewables, materials, sinks); sinks bind the generative regime; grounding test passed; low-carbon lever 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 14913@bnt-lap001
Assistant-Session: add7ef69-b901-4756-9a67-34cf1125d96f
2026-10-13 23:11:25 +02:00

14 KiB
Raw Blame History

Parameter Grounding Register

Status: v0.2 (ASSEMBLY-WP-0008). Grounded values are collected in a separate parameter set, simulations/gadm02/grounding.py (variants central and slow adoption). GADM-0.4's defaults are left unchanged, so the effect of grounding stays visible and reversible.

Status codes:

  • G grounded: a value or range is supported by external evidence.
  • D direction-grounded: evidence supports the direction or relative size, not a magnitude.
  • B bracketed: not grounded; key readings were checked across a range.
  • A assumed, not checked.
  • C calibration choice: set to reproduce a target behaviour.

The model's stocks are normalized indices. Mapping evidence onto them always needs an interpretation, which is stated in the "Mapping" column.

1. Register

Parameter Meaning Default → grounded Status Evidence Mapping
time unit one period 1 year C — implied timescales checked (GADM-0.2.md § 3.4)
dO organism capability depreciation 0.20 / year (0.15–0.35) G software depreciation over 3–10 years; business R&D capital depreciation generally above 15 % (Li & Hall 2020); AI-GPU economic life 2–6 years organism capability combines software, know-how and compute
K_Ω level and path (aO, ell, oseed) organism adoption default 0.50 at year 10 → central 0.27 at year 10 (×0.5); slow 0.10 (×0.2) G (band) about 4.6 % of tasks cost-effectively affected within 10 years (Acemoglu 2024); 15 % of tasks faster with LLMs, 47–56 % with tooling (Eloundou et al. 2023); about a third of occupations use AI for ≥ 25 % of tasks (Anthropic Economic Index, Feb 2025); AI investment ~1–2 % of GDP (2025) K_Ω = share of information tasks carried out by organisms; start 0.05; the default reaching the exposure ceiling in 10 years is too fast
aext attention extraction 0.275 → 0.58 (with grounded care load) G (order of magnitude) ATUS 2023: 5.2 h leisure/day, 2.7 h TV; social media ~2 h 20 min/day present-day attention headroom ≈ 0.5 (engagement media take about half of leisure); the 0.275 compensated for an overstated care load
care_load attention cost of informal care 0.12 → 0.03 (0.02–0.06) G AARP/NAC 2025: 51 M caregivers of people 50+ (about 1 in 5 adults), 26 h/week, up from 22 h in 2020 population average ≈ 5 h/week of care vs. ≈ 35 h/week of engagement media
dX commons obsolescence 0.04 (0.02–0.10) G knowledge depreciation 1–5 % / year, 8–9 % in the first two years (de Rassenfosse & Jaffe 2017); software faster mixed commons of knowledge, standards, software, infrastructure
lX enclosure / loss of commons 0.06 × P (≈ 3 % / year) D 16 % of popular GitHub projects abandoned, 41 % of those revived (Avelino et al. 2019); ~15 % of widely used npm packages with abandonment issues (secondhand) annual loss of 1–3 % of maintained commons
IX deliberate commons investment 0.02 A (top baseline driver) public R&D ≈ 1 % of GDP, roughly stable (Germany 1.04–1.10 %, 2020–2023); business R&D rose to 74 % of OECD GERD no defensible mapping from GDP shares to the normalized commons inflow; see reading 3
aX, rho0 conversion of selected assemblies into commons 0.30, 0.30 D open source: about $4.15 B to recreate once versus about $8.8 T demand-side value (Hoffmann, Nagle & Zhou 2024) very high reuse leverage of non-rival commons; supports high conversion, input for WP-0007
r_k power-channel self-reinforcement market 0.3; digital channels 0.45 (discovery 0.6) D displacement of industry leaders declined sharply since 2000, linked to software and intangibles (Bessen, Denk, Kim & Righi); superstar persistence from learning-by-doing technologies (Ma, Pugsley & Zimmermann) digital channels (infrastructure, discovery, membrane, lineage) reinforce more strongly than the market channel
lw learning-with-use amplification of lineage control 2.0 D learning-by-doing behind persistent superstars (Ma, Pugsley & Zimmermann) direction only
k_w wealth → market/political control 0.04 (0.02–0.20) B lobbying returns over 130 % with a tiny effect on enactment probability (Kang 2016); elites ~15× the influence of average citizens (Gilens & Page 2014), disputed (Bashir 2015) evidence pulls in both directions on magnitude
dH current-capability depreciation 0.05 G skill depreciation ≈ 4.3 % / year without use (Dinerstein, Megalokonomou & Yannelis); decline concentrated among low-use workers (PIAAC-L, Germany) —
dD, hform, ½ / ½ split durable capability 0.01; 0.15; ½ / ½ D early-education test effects fade, but adult outcomes persist (Chetty et al., Project STAR; Deming 2009: Head Start +0.23 SD, ~80 % of Perry Preschool) supports a slowly depreciating durable component; the split is not identified
lbl learning begets learning 1.0 D skill decline depends on skill use (PIAAC-L) direction only
g3s seniors' restartability weight 0.5 G re-employment of displaced long-tenured workers, Jan 2026: 72.9 % (25–54), 57.3 % (55–64), 38.6 % (65+) (BLS Displaced Worker Survey) ratio 65+ / prime age ≈ 0.53
b2s seniors' exposure to operator lock-in 1.5 D 2017: a third of people 65+ never online, smartphone ownership 42 points lower than under-65s (Pew); recent data not found gap likely narrowing over cohorts
gcare care support → senior agency 0.02 A — —
demography births, ageing, deaths 1/45; 1/20, 1/45, 1/15; 1/8 C — stationary structure ~23 / 51 / 17 / 9 %
v0, aK, n1 experimentation, capacity build, entry 2.5, 0.35, 0.30 C — keep the GADM-0.2 baseline from collapsing
all other rates — see code A — first guesses in 0.03–0.3 per year

2. Readings under grounded parameters

Full tables: simulations/gadm02/results/grounded_readings.md. Three variants: GADM-0.4 defaults; grounded with central adoption; grounded with slow adoption. Each starts at its own equilibrium with recalibrated channels.

Reading Defaults Grounded, central Grounded, slow
Best / worst scenario generative / automated capitalism same same
Scenario ranking ρ vs. defaults 1.00 1.00 0.98
Lowest domination generative generative generative
Lever order lineage > membrane > attention +0.111 / +0.075 / +0.031 +0.122 / +0.094 / +0.045 +0.164 / +0.109 / +0.039
P3: full wealth tax ΔP (membrane-dominated) −0.028 −0.055 −0.018
G3: senior agency gain at cp = 0.3 (care share) +0.070 (55 %) +0.069 (56 %) +0.077 (48 %)
Baseline VFAC / generative VFAC 0.321 / 0.659 0.328 / 0.658 0.396 / 0.684

Readings

  1. Every structural reading survives grounding. None of the qualitative results of GADM-0.2 to 0.4 is overturned by the evidence collected here.
  2. The default model adopts organisms too fast. It reaches half of all information tasks within ten years, beyond what even the exposure ceiling suggests. Grounded paths are 0.27 (central) or 0.10 (slow) at year 10.
  3. Governance matters more, the faster organisms are adopted. Under slow adoption, current-trend regimes do much better (baseline 0.321 → 0.396, dividend abundance 0.337 → 0.434), while the generative regime barely changes (0.659 → 0.684). The gap between current trends and the generative regime narrows from 0.34 to 0.29. Adoption speed, an empirical unknown, sets the stakes of organism governance.
  4. Commons investment is the most important ungrounded parameter. It is the top baseline driver (GADM-0.4.md § 3), but there is no defensible mapping from GDP shares to the normalized commons inflow. Its sibling evidence (open-source reuse leverage) points at the conversion of assemblies into commons rather than at investment levels.
  5. The care load was overstated by about 4×. The default was 0.12; the evidence supports 0.03. Grounding it forced a joint recalibration of attention extraction. The two calibrations had been compensating for each other.

3. Input for ASSEMBLY-WP-0007 (validation and falsification)

The evidence suggests tests that WP-0007's structural extensions should pass or could fail:

WP-0007 task Evidence Expectation / falsification test
T01 rival vs. non-rival output open-source reuse leverage (≈ 2,000× demand-side value over recreation cost); AI investment ~1–2 % of GDP but ~90 % of H1-2025 GDP growth in information processing (Furman) Non-rival output should carry most of the commons effect. If splitting Y* does not raise the commons/reuse contribution, the reuse mechanism is mis-specified.
T02 ecological classes compute energy and AI-hardware turnover (2–6-year economic life) Organism growth should bind energy and sinks before materials. If materials bind first, e2 or the class mapping is wrong.
T03 endogenous selection weights leader displacement declined since 2000 (Bessen et al.); lobbying shifts enactment probability only marginally but with high returns (Kang 2016) Selection capture should be slow and incremental, not abrupt. A model where domination flips selection weights within a few years would contradict the evidence.
all adoption-speed band (grounded central vs. slow) Every WP-0007 reading should be reported under both adoption variants (reading 3).

T01 result (2026-10-13, models/GADM-0.5.md § 1): test passed. With the split, the commons effect is carried by non-rival output (xn): it accounts for the whole VFAC gain (+0.015 to +0.049 in the baseline across B1–B4). The split did not produce throughput decoupling (BR-010). New priors phi_r 0.2 (organism output mostly information tasks; Anthropic Economic Index) and xn 0.3 (about aX) need grounding: φ from task-level usage data (L4), xn from open-licensing shares of AI-generated output.

T02 result (2026-10-13, models/GADM-0.5.md § 2): test passed. Organism growth draws most on sinks (−0.036) and renewables (−0.032), and least on materials (−0.018). New priors to ground: the class starting levels, the load shares, and material replenishment (reserve growth). Which class binds next depends on the replenishment.

4. Remaining open grounding needs

  1. Commons investment IX: needs a mapping concept, not just data.
  2. Wealth conversion k_w: evidence is contradictory.
  3. gcare, and the durable/current split.
  4. Recent (post-2020) data on age gaps in digital access.

Sources