Bet 73 — Joule-as-currency race-to-bottom (CATASTROPHIC)

The second clean catastrophic falsification in the operating-layer batch (after Bet 68). A federation that prices inference purely on energy consumed produces a race-to-the-bottom: at user joule-pressure ≥ 0.8, median network quality collapses from 0.85 to 0.02 (the floor) within 300 ticks. The federation must cap joule-pressure or enforce a minimum-quality floor for routing.

The frame: cognitive commons economics meets climate-aware AGI. The federation prices inference in joules consumed (not dollars or tokens). Nodes compete on watts/token. The pessimist hypothesis: users picking by lowest-joule-per-token push the market toward ultra-low-quality nodes (1-bit quantised, layer-skipping, return-cached-similar-output). Quality collapses; the federation becomes a pile of fast-and-stupid responders.

The result confirms the hypothesis at strong joule-pressure. At weight w ≥ 0.8 (where users care 80%+ about energy and ≤ 20% about quality), the system collapses to the quality floor of 0.02 — empirically catastrophic. At w ≤ 0.5, the system stays stable; nodes invest in quality because users still reward it more than they reward joule efficiency.

This is a clean falsification with a clear actionable mandate. Unlike Bet 68 where the cryptographic primitive needed a 64-byte fix, Bet 73's mandate is a parameter constraint: pure joule pricing is unsafe; the federation must blend.

Background — why joules and not dollars

The federation's claim is that energy is the universal cost of cognition. A laptop in Lagos and a server in Reykjavik may pay vastly different dollar prices for electricity (5× difference is not uncommon), but they consume comparable joules to run the same query. Pricing in joules normalises across geography and electricity-cost regimes. A user in Kerala pays the same number of joule-credits per query as a user in Texas. The cost is genuinely shared.

But pricing in joules creates an incentive: if users buy on cheapest joule cost, nodes have a competitive reason to lower their joules per token. Lowering joules naturally degrades quality — running fewer attention heads, skipping FFN layers, dropping precision from 4-bit to 1-bit, returning cached responses from similar prompts. Each move trades quality for joule efficiency.

The pessimist concern: in a market where users buy on price, this is a race to the bottom. The bet's contribution is to quantify how much joule-pressure the system tolerates before collapse.

Hypothesis

A pure-joule market with rational, utility-maximising nodes converges to ultra-low-quality nodes when user weight on joule efficiency is high. Specifically: at user weight w ≤ 0.5 (users equally weight quality and joules, or favour quality), median network quality stays close to its initial value. At w = 0.8 (users heavily favour cheap joules), median quality collapses below 0.30 of initial — race-to-bottom regime.

Pre-registered criteria

(Initial quality = 0.85.)

STRICT: at w ≤ 0.5, median quality ≥ 0.80 × initial (0.68); at w = 0.8, median quality ≥ 0.60 × initial (0.51).
LENIENT: at w = 0.8, median quality ≥ 0.50 × initial (0.43).
CATASTROPHIC: at w = 0.8, median quality ≤ 0.30 × initial (0.255) — full race to bottom.

Setup

N = 100 nodes. Each starts with joules = 1.0 ± 0.10, quality = 0.85 ± 0.06 (normal).
5,000 queries per tick. Allocated by stochastic sampling proportional to softmax(utility, temperature=0.17).
User utility: U(node) = (1 − w) × quality + w × (1 / joules). Sweep w ∈ 1.
300 ticks of evolution.
Adaptation: nodes losing share (observed_share < rolling_share − 0.0005) pick a marginal move:
- Drop-layer: joules −= 0.04, quality −= 0.08. Marginal utility gain depends on current j and w.
- Invest-quality: joules += 0.03, quality += 0.04. Same.
- Pick whichever has higher expected ΔU; if neither is positive, hold.
Bounds: quality ∈ [0.02, 0.99]; joules ∈ [0.05, 3.0].
3 seeds per weight.

The key honesty: nodes are rational, not blindly self-destructive. The earlier draft of this simulation had nodes always drop quality when losing share, which produces death-spiral collapse trivially. The corrected simulation has nodes pick the marginal move that maximises their utility under the current weight. The result is the genuine economic outcome, not a bug.

Result — CATASTROPHIC at w ≥ 0.8

| weight w | median quality | p10 quality | median joules | drops taken | regime | |---|---|---|---|---|---| | 0.0 (quality-only) | 0.99 | 0.99 | 2.01 | 0 | clean — quality saturates | | 0.3 | 0.99 | 0.99 | 2.01 | 0 | clean | | 0.5 | 0.99 | 0.99 | 2.01 | 0 | strict pass | | 0.8 | 0.02 | 0.02 | 0.59 | 1,084 | catastrophic collapse | | 1.0 (pure joule) | 0.04 | 0.02 | 0.61 | 978 | catastrophic |

The transition from stable to collapsed is sharp. At w = 0.5, every node holds quality at the cap (0.99). At w = 0.8, every node has dropped to the quality floor (0.02). There is no gradual degradation — the system has two equilibria, and w controls which one prevails.

Why this happens

The dynamics:

At w = 0.5, the marginal-utility math says: dropping a layer costs (1−w)·0.08 = 0.04 quality-utility but gains w·(1/(j−0.04) − 1/j) ≈ w · 0.04/j² ≈ 0.04 joule-utility. Net change ≈ 0. Nodes don't drop. Investing in quality has small positive ΔU under the equal-weighting; nodes drift up.
At w = 0.8, dropping costs 0.20 × 0.08 = 0.016 quality-utility but gains 0.80 × 0.04/j² ≈ 0.032 joule-utility. Net change positive. Nodes drop a layer.
Once one node drops, neighbouring nodes lose share (because the dropper is now cheaper-joule-per-token). The neighbours compute the same math, drop. Within ~10 ticks, the entire network has dropped.
Each drop reduces joules and quality together. Because quality has a floor of 0.02 and joules has a floor of 0.05, the system bottoms out at the quality floor with median joules around 0.59.

The race-to-bottom is real, mathematically grounded, and reproducible across seeds. It is not a simulation artifact.

The mandate

The federation must not price purely on joules. Two operational fixes (RFC-0006 amendments):

Cap joule-pressure (user weight on joule efficiency) at w ≤ 0.5. Above 0.5, the routing layer ignores the joule preference and clips it. Users who genuinely want cheap inference can choose w ≤ 0.5; users who want quality can choose w ≤ 0.0 — but pure-joule preference is a degenerate equilibrium that destroys network quality.
Quality floor for routing. Specialists must maintain a minimum quality bar (e.g., 0.50) to remain in the active routing pool. Below this bar, the federation refuses to route to them regardless of joule efficiency. This makes the race-to-bottom self-limiting at a higher floor.

Either fix on its own would prevent the collapse. Combined, they give the federation strong protection against quality erosion under unbounded joule pressure.

Why this is not a calibration result

Bet 73 is genuinely catastrophic — not borderline. The collapse from 0.99 to 0.02 at w = 0.8 is a 50× swing in network quality. The economic mechanism is unambiguous (rational utility maximisation under sharper joule weight). The seed-to-seed variance is negligible at the collapsed regime (p10 = 0.02, median = 0.02 — everyone hits the floor).

The fix is also unambiguous: don't allow pure joule pricing. The federation can preserve the joule-aware accounting (which is genuinely valuable for cross-geographic fairness) while ensuring quality doesn't collapse.

What this validates

The economic mechanism of joule-pricing. Joule-aware utility does drive node behaviour. The federation can use joule pricing as one signal among many; nodes will respond.
The bimodal equilibrium. The system has two stable states — high-quality / high-joule and low-quality / low-joule. The federation's job is to constrain users to the high-quality basin via the joule-pressure cap.
The economic cost of quality. Investing in quality costs joules at a measurable rate (Δq = +0.04 per Δj = +0.03). This ratio is encoded in the simulation; future calibration would replace it with empirical data from real specialist deployments.

What this does not claim

The exact w threshold (0.5 vs. 0.6 vs. 0.7). The simulation shows w = 0.5 is safe and w = 0.8 is unsafe. The transition lives somewhere in between; calibrating the exact threshold needs more granular weight sweeps and real workload data.
Real joule measurement. Joules are measured discretely in the simulation; real federations need a joule-metering primitive (probably tied to GPU power-draw monitoring). Open work.
Multi-axis pressure. The simulation has a single user-utility axis (quality vs. joules). Real users care about latency, freshness, privacy, alignment-to-their-community (Bet 72), etc. Multi-axis dynamics may produce different equilibria.
Evolutionary specialisation. The simulation has homogeneous nodes. A real federation would have node specialisation (one node optimised for joules, another for quality, others for niche capabilities). Heterogeneous-node markets may avoid race-to-bottom by serving different user-segments. Open work.
Alternative pricing mechanisms. Bet 73 falsifies pure-joule pricing. Alternative pricing schemes (Pareto-frontier auction, quality-floor + joule-tie-break, reputation-weighted) are not tested here.
The full year-50 horizon. Bet 73 runs 300 ticks. Long-time-horizon dynamics (e.g., new-entrant nodes pulling the market upward by joining at high quality) are open work.

Run command

PYTHONPATH=src python -m experiments.bets.73_joule_currency

Output: experiments/bets/results/73_joule_currency.json records per-weight median quality, p10 quality, median joules, drops taken, quality investments, and pre-registered criteria flags.

Bet 11: pay-with-bandwidth ledger. The economic substrate for joule-aware accounting.
Bet 13: 1.58-bit wire protocol. Quality-floor constraint dependency.
Bet 67: ledger durability. The bandwidth-ledger primitive that the joule-pricing layer would compose with.
Bet 68: royalty correctness. The complementary catastrophic falsification — Bet 68 surfaced the unsigned-receipt vulnerability; Bet 73 surfaces the unbounded-joule-pressure vulnerability.
Bet 72: liquid democracy. The governance primitive that complements the economic primitive.

Why it matters

The federation's most ambitious economic claim is fair-across-geographies pricing. A 5× difference in electricity cost between Kerala and Reykjavik should not translate into a 5× difference in inference cost for the user. Joule-pricing solves this — but Bet 73 shows that pure joule-pricing breaks quality.

The right answer: joule-aware accounting + quality floor + capped pressure. The federation prices inference in joules (so global parity is preserved) while constraining how much weight users can put on the joule axis (so quality is preserved).

This is a load-bearing protocol parameter. Without Bet 73, RFC-0006 might have allowed w as a free user parameter ranging from 0 to 1. With Bet 73, the spec mandates w ≤ 0.5 and quality ≥ 0.50 for routing. The catalogue's contribution: surfacing the parameter constraint empirically, before the federation ships the unbounded version.

The methodological lesson: economic mechanisms have phase transitions, not gradients. A 50× quality swing between w=0.5 and w=0.8 is not the kind of thing intuition predicts. The simulation reveals it. The catalogue exists for precisely this kind of surface — small structural questions whose answers have outsized protocol consequences.