Can Europe capture the value of its own battery waste, and what can European refiners then serve of the recycled-lithium minimum that binds in August 2031?

What this model cannot tell you

• The model balances tonnes, recovery rates and metal value. It does not compute a rate of return, a payback or a net present value, and it should not be read as saying whether a plant is worth building. The corpus's returns material exists and is adjudicated, and it is denominated in euros while every price in the corpus is in dollars with no exchange rate published anywhere in the evidence, so joining them would be inventing a number.
• Copper and manganese are in the powder and are recovered, and the corpus publishes no price for either. Every value share this model reports is therefore a share of lithium, cobalt and nickel only, and it understates what a real refinery sells. It also understates the difference between the routes, because manganese is where smelting loses most after lithium.
• Lithium here is an index, never a tonnage. The powder is assayed at 2 to 6 per cent lithium compounds and the payable is quoted as a share of the lithium carbonate price, and nothing in the corpus converts between the two. The coverage comparison works because the conversion cancels on both sides of a ratio; ask this model for tonnes of lithium carbonate and it will not answer.
• The level of the coverage rail is conditional and the ordering is not. The 14 per cent of 2030 EU lithium demand that recycling could cover is published (C-42) and the feedstock, collection, yield and recovery assumptions behind it are not; the corpus publishes no EU lithium demand tonnage anywhere, and this model publishes no conversion from contained lithium compounds to carbonate equivalent, so there is no route in this evidence to an absolute grounding and the specification does not manufacture one. What it does instead is declare a reference basis in full, compute it forward from five named published values rather than back-solving it from the model's own output, and measure how little the corpus constrains it: the published yield, assay and recovery bands alone leave the reference index anywhere from 3.024 to 12.348 kt, a range 4.083333 times wide. Every coverage number and both quota-service ratios inherit that. Read quota_service_ratio as a position against the most generous configuration the corpus's own published values allow, not as an absolute share of the statutory minimum. The model reports the same configuration's ratio at both ends of that band and prints verdict_robust_to_normalisation, which is nought at the defaults because the central case reads strained at one end and holds at the other.
• What the normalisation does not touch. It divides both sides of every comparison, so it is absent from every route gate, every margin, every payable, every crossover price, every utilisation and every value share, and it cancels exactly out of eu_origin_share_of_recycled_li, which is also the ratio between the two quota-service rails. The finding that European refiners hold a little over a third of the recycled lithium in Europe's own waste, and every statement this model makes about which configuration is better or worse than which, stands whatever the normalisation is. Only levels against thresholds are conditional.
• There is no out-of-sample test of the coverage rail in this package, and the refusal is a finding rather than an omission. A genuine test would need a second published statement of recovered lithium against a feedstock tonnage on the same basis the model uses. The nearest thing the corpus holds is C-70, which reports 17,100 tonnes of regenerated lithium salts from 120,000 tonnes of used batteries at one Chinese operator. That is a salts basis, this model is on a contained-lithium-compounds basis, and the corpus publishes no conversion between them - the same absence that stops the coverage rail being grounded in the first place. Comparing them would produce a number that looks like a validation and is a unit error, so the comparison is refused. V-REFERENCE-IDENTITY, which used to be described as a calibration proof, is an algebraic identity that cannot fail and is now labelled one.
• Capacity is measured on one basis and refining on another. The corpus's capacity series mixes an installed pre-treatment figure with an announced refinement figure and compares both directly against tonnes of arisings, and the model follows that comparison because it is the corpus's own. A tonne of pre-treatment capacity and a tonne of refining capacity are not the same tonne, and where the model says utilisation it means the corpus's comparison, not a plant engineer's.
• The European refining share is a capacity ratio standing in for a flow. It is built from the operating facility record, plant by plant, with the classification printed in datasets/eu_facilities.csv so anyone can disagree with it. It is not a measured export share, because the corpus holds no black-mass export tonnage and the report lists that absence among the things it could not establish. The two ends of its range are site-level observations, not a distribution.
• The route mix rests on one number whose geography is unknown. Hydrometallurgy's 65.6 per cent processing share is the only route split the corpus publishes and the source does not say whether it is global or European. Its evidence range is that single point, so any other allocation is the reader's own.
• Utilisation as this model computes it runs above what operators report. At the 2024 anchor the model returns about 35.6 per cent aggregate utilisation while operators report under 10 per cent at many plants. The gap is arisings that never reach a plant, which the corpus does not measure, and the model surfaces it as a diagnostic rather than smoothing it away. The low end of collection_share is exactly the value that reconciles them.
• The two published cost sources for hydrometallurgy disagree by nearly a factor of three, 1,555 to 5,662 dollars a tonne against about 1,300. The model uses the first because it is the only source that costs all four routes on one basis, carries the second as a live check on the margin it returns, and prints the ratio rather than choosing.
• Smelting is barred by a legal gate that rests on a mid-point. Its published lithium recovery band is 0 to 30 per cent, and even its top is below the 50 per cent the statute requires from the end of 2027, so the gate would close at any point in the band. But the recovery target itself is regulation that has already slipped once elsewhere, and the reader can switch it off. When they do, the model's most profitable route returns and the lithium goes to slag.
• The chemistry shift is approached and not modelled. Genuine LFP powder carries no cobalt and no nickel, which is below every composition the corpus assays, so the thin-powder scenario reaches the bottom of the published bands and stops there. It understates the effect it is built to show, and the corpus holds no chemistry mix for European feedstock to do better with.
• Second-life storage is not in the model at all. Every pack diverted into stationary storage arrives at a recycler a battery lifetime later, and the corpus rates the flagship ramp only 30 per cent likely on its own assessment. The diversion is real and unquantified, and it makes the arisings path here slightly generous.
• The projection's own band and the corpus's published band disagree at the horizon, and the horizon cap holds in the year rather than in the value. A single compound rate calibrated on the 2030 band cannot also reproduce the 2040 band, because the corpus's two bands imply different growth regimes and C-06 names them: production scrap peaks around 2030 and end-of-life batteries take over after 2035. The model carries one driver, so its 2035 path sits at position 1.005695 in the published 2035 band and its 2040 path at 1.010215 in the published 2040 band, both just outside the top, and its own 2040 band runs 986.6 to 4,915.3 kt against a published 1,500 to 2,100. The positions are computed and printed rather than clipped, because clipping would import a regime change the corpus publishes no rate for. Nothing in the verdict depends on any year after 2031.
• The backcast is two points and one of them is the point that fixes the driver. Two points cannot distinguish compound growth from several other shapes that would fit as well, and the model claims only that a constant rate is consistent with every European arisings figure the corpus publishes, within a tolerance it states.
• Nothing in this model tells you whether the law will hold. It computes what European refiners could serve of an obligation as it is currently written, on dates as they are currently set. The corpus records the EU postponing one battery obligation by two years already, and the reader can soften the recovery target and the quota level and watch the world change. Whether the legislature will is not a quantity this or any model carries.
• The largest thing this model holds fixed that the world does not is the origin rule. Nothing in the regulation as harvested requires the recycled content in a 2031 battery to have been recycled in Europe, so the model computes both rails and reports the gap between them. It does not predict which rail the Commission's implementing acts will create. That gap, quota_origin_gap_ratio, is the single number this specification would most like to be able to close and cannot.
• What would need new research: a measured European collection rate for EV batteries, so the supply rail stops resting on a legal duty; a black-mass export tonnage series, so the refining share stops being a capacity ratio; a European black-mass grading standard, so composition stops being a range spanning a factor of four on cobalt; a published copper and manganese price on the same basis as the other metals, so the value shares stop understating; and the Commission's recycled-content calculation methodology, due by delegated act from 18 August 2026, which decides whether the origin gap is real.
• The emissions rail is computed and then stops, and that is a disclosure rather than an oversight. scrap_share_2030 moves the scrap and end-of-life split and, through it, ghg_saving_blended; ghg_saving_eol moves the same blend; nothing downstream consumes either. The corpus publishes no yield, grade or value difference between production scrap and end-of-life packs, so the split cannot reach the material balance; it prices no carbon in this subject, so the saving cannot reach a margin; and C-50's maximum carbon thresholds, dated February 2028 for EV batteries and February 2029 for industrial, are carried in the corpus with their dates and without their numbers, so the saving cannot reach a compliance gate either. Connecting the branch would mean inventing a carbon price or a threshold. The calculation is small, correct and sourced to C-75, and it is left terminal and named as terminal.
• Three parameters are carried for reference and enter no equation: bm_cu_content, bm_mn_content and hydro_breakeven_kt. None is reader-settable, because a lever that moves nothing is worse than an absence. Copper and manganese cannot be valued, since the corpus publishes no price for either, and they cannot be given a mass-basis consequence either, since C-49's whole-process efficiency target is defined over the whole battery and this model tracks only the black mass. C-89's break-even scale cannot be compared against anything here because C-89 does not say whether its 7,000 tonnes is battery input or black mass. Four validation vectors pin the copper and manganese inertness at both ends of their published bands, so a builder who wires either to an output will fail the vectors rather than ship a control that appears to do something.