Some industries stay on old, wasteful technology even when a better option exists. The reason is that each firm's payoff depends on what its neighbors use: shared pipelines, supplier tooling, and customer habits all favor the incumbent. One line of work shows how to break such a lock-in with a temporary shock. In a population of learning firms, a regulator briefly raises how much the firms experiment, like turning up a temperature knob. Past a critical point the bad equilibrium disappears, and because of hysteresis (the system does not slide back when the knob returns to normal), the firms stay on the good technology after the shock ends (Catastrophe by Design in Population Games). That result assumes one regulator who sees every firm's state and turns one global knob. Real industries are split into information silos. Firms and sub-networks keep their operational and emissions data private, and the regulator sees only coarse summaries.
This note asks how to tip a lock-in that is coupled across silos when each controller sees only its own silo. We propose borrowing the structure of a journal editorial board. An editor-in-chief sets policy from short verdicts. Associate editors handle the detail inside their own domains and pass up decisions, not raw material. Applied to industry, each associate controller sees the full state of one silo and controls experimentation there. It reports upward only a single bit per period: whether its silo has passed its local tipping point. The top controller uses those bits to decide which silos to shock, and in what order, so that each silo that tips lowers the threshold its neighbors need to cross. The claim to test is that this staged, hierarchical schedule reaches a permanent transition at much lower total intervention cost than a uniform global shock. It should also succeed where independent per-silo control fails, and it should do this while disclosing almost nothing across silo boundaries.
Core Claim
When the network effects behind a lock-in cross silo boundaries, the critical temperature of each silo depends on how far its neighbors have adopted the new technology. That dependence creates a possible cascade. Tip the right silo first, and its neighbors need a smaller shock, or none at all.
A global knob cannot exploit this, because it must be set high enough for the most entrenched silo. Independent local controllers cannot exploit it either. They do not know when their neighbors have tipped, so they either overshoot or release too early and fall back into the locked equilibrium. A two-tier controller has a different information pattern: full local state below, one bit per silo above. That pattern is close to the minimum needed to schedule the cascade.
We predict two things. First, the two-tier controller's cost will approach that of an all-seeing centralized planner. Second, it will clearly beat both the uniform global shock and uncoordinated local control in networks with strong cross-silo coupling and heterogeneous entrenchment. In networks with weak coupling, or with equally entrenched silos, we predict the advantage vanishes. The approach has to survive that falsifiable boundary.
Model
There are $K$ silos. Each silo $k$ is a population of firms choosing between an incumbent and a clean technology, and $y_k$ is the fraction of silo $k$ using the clean one. The firms learn by Q-learning with softmax exploration at temperature $T_k$, following the dynamics in Catastrophe by Design. The clean technology's payoff in silo $k$ rises with a coupling-weighted adoption level $\sum_j W_{kj} y_j$, where the matrix $W$ encodes shared infrastructure. The motivating case is steel and cement plants whose switch depends on hydrogen and CO₂ pipelines no single plant controls (Feasible supply of steel and cement within a carbon budget).
Each silo's critical temperature therefore becomes a function $T_c^{(k)}(y_{-k})$ of the others' adoption. The controller's objective is to choose schedules $T_k(t)$ that minimize
$$C = \sum_{k=1}^{K} \int_0^{H} \big(T_k(t) - T_0\big)_+ \, dt,$$
where $T_0$ is the baseline temperature and $H$ is the horizon. The objective is subject to a constraint: once every $T_k$ is returned to $T_0$, all silos remain at the clean equilibrium. The integrand measures excess experimentation, which is a proxy for lost output during the shock. We also report the welfare loss of the firms directly, so that the result does not hinge on this proxy.
The information constraint is the crux. The top tier observes only $b_k(t) = \mathbb{1}[\,y_k(t) > y^*_k\,]$, one bit per silo per period. Each associate controller observes its own $y_k$ and payoffs and nothing from other silos.
Research Program
The work has three parts.
The first is analytical. For two and three coupled silos we will characterize how $T_c^{(k)}$ shifts with neighbor adoption. We will then derive conditions under which a sequential schedule strictly dominates the uniform global shock. We expect these conditions to involve asymmetry in entrenchment and the strength of the off-diagonal entries of $W$, and we expect the dominance region to be empty when $W$ is diagonal.
The second is algorithmic. We will build the two-tier controller in a simple form. Associate controllers run a local ramp policy that raises $T_k$ until their bit flips and then releases slowly. This ramp also addresses the open problem, flagged in Catastrophe by Design, that abrupt shocks are politically impractical. The top tier runs a bandit or greedy scheduler over which silo to activate next, using only the bit vector. The top tier could alternatively be learned with reinforcement learning on bit-vector observations.
The third is stress testing. We will add an adversarial incumbent silo whose controller lowers its own temperature to resist the transition, which models the entrenched stakeholders the literature notes formal models ignore. We will also add noisy or delayed bits, which model the upward reporting latency that hierarchies suffer (An Analysis of the Tsingtao Beer Merger).
The editorial analogy does real work here rather than serving as decoration. Editorial front matter of large journals describes exactly this split: an editor-in-chief sets policy while associate editors adjudicate within domains and pass up decisions (IEEE Transactions on Systems, Man, and Cybernetics publication information). The embodied-carbon ledger of A Blockchain-Based Approach For Embodied Carbon Management shows the same pattern in industry. Suppliers keep raw mix data off-chain, and only endorsed hashes move up the tiers. That paper supplies the governance pattern and its confidentiality motivation. It does not model any lock-in dynamics, which is the part we add.
Two further papers motivate the work. The Impact of the Degree of RMOR on Corporate ESG Performance finds that asset-heavy firms blunt ordinary incentives, which motivates using nonlinear shocks at all. The cusp-catastrophe case study (A Case Study of Discontinuous Innovation Based on Cusp Catastrophe Model) offers only a qualitative fold picture. We make the fold quantitative and multi-silo, with a measurable early-warning bit.
Baselines
The naive reference point is doing nothing: $T$ is held at $T_0$. This establishes that the network is genuinely locked and gives the welfare gap a transition would close. Next to it sits the permanent linear subsidy, a constant boost to the clean payoff kept until adoption passes a threshold and then kept on. We include it because it is what policymakers actually use, and its cumulative cost over the horizon is the economic deadweight that the hysteresis argument claims to avoid.
The strongest existing method is the uniform global shock of Catastrophe by Design. It raises a single $T$ for all silos past the largest critical temperature and then drops it, and we run it with the same ramp shape and an equal grid search over peak and duration. That paper gives a formal mechanism but no reported numerical cost for this setting; no reported result found on Lacuna for multi-silo transition cost. We therefore treat our own tuned implementation as the reference to beat, and we report its cost on every network. Beating it means reaching a permanent transition in all silos at lower cost $C$ and lower firm welfare loss, across at least the high-coupling, heterogeneous-entrenchment regime, with confidence intervals over random seeds and random $W$.
Independent local control is the second strong competitor. Each silo runs the same ramp-and-release policy as our associate controllers, but with no top tier and no bits. This is the ablation that isolates the key ingredient, the upward bit plus scheduling. If it matches the hierarchy, the hierarchy adds nothing.
A second ablation keeps the top tier but makes its schedule random rather than bit-driven. This separates "staging helps" from "informed staging helps."
The oracle centralized planner sees all $y_k$ and payoffs and optimizes the schedule by dynamic programming or exhaustive search on small $K$. It gives the upper bound and lets us state how much of the achievable saving the one-bit hierarchy captures.
All methods get the same networks, firm-learning parameters, horizon, and tuning budget, measured in simulated episodes.
Why not just…?
Why not just turn one global knob, as the original bifurcation paper does? Because the knob must exceed the critical temperature of the most entrenched silo for long enough that the slowest silo tips. That spends excess experimentation in every silo that would have tipped cheaply, or by cascade alone. Whether this waste is large is an empirical question. If the analysis shows the dominance region is small for realistic coupling, the global knob is good enough and we will say so. The cost comparison across coupling strength is designed to show exactly where the crossover lies.
Why not just let each silo manage its own transition? This is the most serious threat. If local adoption alone tells a controller enough about whether its neighbors have tipped, which happens when coupling is weak or tipping is fast, independent control may match the hierarchy. The evidence that it can fail comes from coordination studies. Uncoordinated LLM supply-chain agents deadlock under demand shocks, and when they negotiate they drift into an unmonitored informal hierarchy (Agentic LLMs in the Supply Chain). Designated leadership cuts task time by 9.76% for teams of three GPT-3.5-turbo agents and keeps communication cost linear up to nine agents (Embodied LLM Agents Learn to Cooperate in Organized Teams). Neither result is about lock-in, so the local-control ablation is the experiment that decides this, and we report its failure modes separately: premature release followed by reversion, and overshoot.
Why not just pool the data and plan centrally? Firms in multi-enterprise networks refuse to cede proprietary data (Blockchain-based Enterprise Community Data Access Control Methods). Training schemes that pool observations, such as the centralized-training multi-agent method in Multi-Agent Deep Reinforcement Learning for Multi-Echelon Inventory Management, assume the sharing that silos forbid. We keep the central planner as an oracle precisely to measure the price of the information constraint. If one bit per silo recovers most of the oracle's savings, pooling is unnecessary.
Why not just optimize the profit–carbon trade-off directly with multi-objective reinforcement learning? Decomposition-based multi-objective reinforcement learning finds dense trade-off curves in multi-echelon networks. Reinforcement Learning for Multi-Objective Multi-Echelon Supply Chain Optimisation reports up to 75% higher hypervolume than NSGA-II, and a front about eleven times denser than modified PPO, on its Complex network. But it optimizes a fully observed operating point. It does not address the question of escaping a stable bad equilibrium with an intervention that can later be withdrawn. We use it as a complement rather than a rival. It supplies the operating policy inside the supply-chain experiment below, while the intervention schedule is ours.
Proposed Experiments
The first experiment is the stylized multi-silo population game described above. It uses $K$ from 2 to 16, random coupling matrices $W$ with controlled off-diagonal strength, and entrenchment heterogeneity varied by payoff asymmetry. We report four quantities: cumulative cost $C$, firm welfare loss, probability that all silos remain clean after the intervention is withdrawn, and time to transition. Results are plotted against coupling strength and heterogeneity, and they decide the core claim. The hierarchy must beat the tuned global shock and the independent-control ablation in the high-coupling regime. It must also capture a stated fraction of the oracle's savings, which we will fix before running, at 70%. The advantage must disappear as $W$ approaches diagonal. If it persists at zero coupling, the model has a bug. If it never appears, the claim is wrong.
The second experiment embeds the mechanism in the Messiah multi-echelon simulator (Reinforcement Learning for Multi-Objective Multi-Echelon Supply Chain Optimisation). Manufacturers choose between an incumbent and a low-carbon process. The low-carbon option's cost falls with adoption by upstream suppliers who share infrastructure. Each echelon is a silo, and only the bits cross echelon boundaries. We report the same cost and permanence metrics. We also report where the post-transition operating point lands on the profit–carbon front, compared with the incumbent's front.
The third experiment is the stress test. It adds an adversarial incumbent silo, bit delays of one to ten periods, and bit-flip noise. It measures how quickly the hierarchy's advantage erodes, and whether a smooth ramp still triggers a permanent transition where an abrupt shock does.
A positive result would show that an editorial-style split of authority is a near-sufficient information structure for engineering industrial tipping points under confidentiality. A negative result in the local-control comparison would be equally informative: it would show that silos do not obstruct transitions as much as the governance literature assumes.
References
- Catastrophe by Design in Population Games: Destabilizing WastefulLocked-In Technologies
- Feasible supply of steel and cement within a carbon budget is likely to fall short of expected global demand
- An Analysis of the Tsingtao Beer Merger
- IEEE Transactions on Systems, Man, and Cybernetics publication information
- A Blockchain-Based Approach For Embodied Carbon Management Along The Construction Supply Chain
- The Impact of the Degree of RMOR on Corporate ESG Performance
- A Case Study of Discontinuous Innovation Based on Cusp Catastrophe Model : Implications for Predictive Risk Management
- Agentic LLMs in the Supply Chain: Towards Autonomous Multi-Agent Consensus-Seeking
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Blockchain-based Enterprise Community Data Access Control Methods and Applications
- Multi-Agent Deep Reinforcement Learning for Multi-Echelon Inventory Management
- Reinforcement learning for multi-objective multi-echelon supply chain optimisation