The Situation

There is more plastic in the world every year. Approximately 460 million tonnes are produced annually; roughly 20 million tonnes enter the environment. Chemically recycled material accounted for about 0.2 percent of European plastics production in 2024.

The technical means to do better exist and have been demonstrated. Pyrolysis operates at commercial scale. Enzymatic depolymerization achieves over 95 percent conversion on waste streams nothing else can process. Catalytic routes can deconstruct polyolefins below 300°C. Electrically heated crackers cut process emissions by 90 percent or more. Solvent and delamination routes recover intact polymer layers from multilayer film.

Regulation exists. The EU Packaging and Packaging Waste Regulation began applying on 12 August 2026, with recycled content mandates and design-for-recyclability requirements landing in 2030. National incorporation bonuses exist. Extended producer responsibility schemes are expanding.

Capital has been available. Brand commitments have been made publicly and repeatedly.

And over the same period: Plastic Energy entered administration. Viridor closed three plants that were achieving 70–75 percent yield. Mura’s Böhlen project ended when a neighbouring cracker closed. Six US PET recycling operations shut, removing 615 million pounds of annual input capacity and 679 jobs. Nine of sixty-five European chemical recycling projects were cancelled. The Global Plastics Treaty reached its sixth negotiating session without a consolidated draft text.

The Question

Given the technical, regulatory, and financial means apparently available, why is plastic production rising while the capacity to recycle it contracts?

That is the whole of it. It is not a rhetorical question and we do not have the answer.

Fellows are asked to model the system well enough to produce a defensible account of where the failure actually sits — and to be able to say which parts of the system, if changed, would move the outcome, and which would not.

Scope is full.Wellhead to shoreline. Production, use, collection, sorting, reprocessing, export, leakage, and the regulatory, financial and market structures surrounding all of it. Where the boundary is drawn inside that scope is the team’s decision to make and defend.

What Success Looks Like

Not a solution. An explanation, modeled rather than asserted, that survives contact with the archive. A strong outcome has these properties:

• It accounts for the failures and the counterexamples — explanations that only explain failure are incomplete, because not everything has failed.
• It identifies where the binding constraint sits, and distinguishes that from where attention and intervention are currently directed.
• It is honest about what the model cannot represent.

A team that concludes its model cannot capture the variables that actually govern the outcome — and demonstrates rigorously why — has produced a first-class result. It is not a consolation outcome.

Method: The Model Stack

Four models, built in sequence. Each hands the next something it cannot proceed without. The stack is load-bearing: if a layer were removed, the finding would change.

The four models are four grains on one system. A requirement follows from that, and it is not optional: the layers must agree. If the state diagram says an asset survives and the parametric model says it does not, one of the two projections is lying. Finding out which is part of the work, not a failure of it.

Layer 1 — Block Diagram

What the system is made of, how it is wired, and where its boundary sits.

The boundary decision is the hard part and should be treated as a first-order analytical act, not a preliminary. A parts list records what things are; a wiring diagram records what things depend on. Mura’s Böhlen project died because a neighbouring cracker closed — on a parts list you delete the cracker and the diagram still looks correct. On a wiring diagram, Mura’s output port now connects to nothing.

A process block flow diagram usually carries two stream types: material and energy. This system carries four — material, money, information, and stress (leakage, host-community exposure, externality). Minimum interfaces to draw as connectors: bale specification, acceptance specification, certification and mass balance, residue offtake, and the debt covenant — the wire along which a yield shortfall travels to become an insolvency.

Layer 2 — Causal Loop Diagram, with Iceberg Synthesis

The feedback structure connecting the blocks.

Loops identified, polarity assigned, reinforcing distinguished from balancing. The Iceberg synthesis traces surface events in the archive — a closure, an administration filing, a stalled negotiation — down through patterns to the structures and mental models producing them. No loops are supplied in this brief. Identifying them is the work.

Layer 3 — State Diagram

What regime the system is in, and what moves it between regimes.

Two properties should be represented where present: irreversibility (transitions that cannot be undone at the price at which they occurred) and absorbing states (from which no exit exists). Weeks 1–3 build dated event logs for eight ventures — six that stalled or ended, two still running. Each is a trace the state diagram must reproduce.

Layer 4 — Parametric Model

Values attached to the relationships surfaced by the preceding layers, propagated through to outcome.

• Scenarios must be narrated, not merely numbered.
• Correlation structure must be explicit — treating correlated variables as independent understates the tail and will be marked as an error.
• Every parameter traceable to a source or a flagged estimate.

Sequence

Twelve weeks, 1 September to 22 November 2026. Teams of four. Every capstone deliverable falls in the week after the session that teaches what it requires — nothing is due before the concept behind it has been taught.

Running alongside: five sixty-minute Forge lab clinics in weeks 2, 4, 6, 8 and 10 — archive tooling and provenance, agent architecture, keeping the evidence base live, multi-dimensional encoding, and provenance audit.

The Research Assistant

The system under study is moving while it is being modeled. Crude prices move, mandates are drafted and delayed, ventures announce and fail. A model built against a static snapshot is stale before it is presented.

Teams build an agent system that maintains the evidence base rather than merely querying it. Its composition is their decision. The binding requirement: the assistant accelerates forensic work without sacrificing transparency. Every retrieved claim must remain traceable to its source, and the team must be able to explain what the system did and why.

Deliverables

• The question set — initial (Week 3) and final (Week 11), with resolution status
• Eight venture chronologies — dated event logs, source type and confidence per row
• The grain justification (Week 5) and the boundary justification (Week 6)
• The Model Stack — block diagram, causal loop diagram with Iceberg synthesis, state diagram, parametric model
• A statement that the layers agree, or an account of where they disagree and which projection is at fault
• The account — maximum ten pages
• The multi-agent research assistant, with provenance preserved
• The gap register, opened Week 4 and updated throughout
• A one-page statement of what the team would need to know next, and how it would find out

Evaluation Criteria

• Quality of the question set. Did the team see what does not add up, or did it summarize?
• Boundary justification. Is the cut defensible, and does the enclosed system contain the failure modes that have actually occurred?
• Grain justification. Is the aggregation defended, and does the team know which materially different situations its model describes identically?
• Retrodictive validity. Does the state diagram reproduce the eight chronologies?
• Agreement between layers. Do the four models tell the same story?
• Treatment of correlation. Were variables modeled as independent that are not?
• Explanatory coverage. Does the account explain the counterexamples as well as the failures?
• Intellectual honesty. Did the team state what it does not know?
• Recognition of model limits. Did the team identify what its model structurally cannot represent?
• Communicability. Could someone outside this field use it?

Explicitly not evaluated: optimism, ambition of the proposed intervention, or novelty of technology invoked.

Design Constraints

• No single composite score. Value of any kind — environmental, social, economic — is never collapsed into one number, in any deliverable, at any stage.
• No advocacy. Forensic analysis. This binds in every direction: a team that sets out to indict an industry has made the same error as a team that sets out to defend one.
• No molecular engineering. The challenge is at system level.
• No national averages where regional data exists. Bale prices, tariffs, collection rates and policy regimes diverge sharply below the national level.
• No undisclosed assumptions. An assumption not in the register does not exist.
• Verify everything. Figures in the archive are compiled from public reporting and must be independently confirmed before a team relies on them.

Competency Mapping

Heavily exercised: Systems Thinking · Model-Based Systems Engineering · Investigative Research · Quantitative Data Analysis · Policy Analysis · Circular Economy · Supply Chain Analysis · Multi-Agent AI Design · Climate Finance

Meaningfully exercised: Environmental Justice · ESG Analysis

Available depending on team direction:Community Engagement — if a team’s boundary encloses host communities or informal-sector collection as active blocks rather than externalities, this becomes live.