What the papers say, and where they disagree with the plan

A triggered design needs one bond slow on a shelf and fast in a tank. If temperature is the only difference between those states, selectivity is S = exp[(Ea/R)(1/T_s − 1/T_w)], which rises without limit in Ea — so a stiff enough bond appears to always win.

It does not, because Ea is bounded from above by the other half of the duty cycle: recovery has to finish. Requiring a half-life of t_w at T_w gives Ea ≤ R·T_w·ln(A·t_w/ln2). Substitute that maximum back into S and the activation energy cancels completely:

No Ea. No ΔH, no ΔS, no ceiling temperature, no monomer structure. The maximum thermal selectivity of any thermally triggered chemistry is fixed by the two temperatures of the duty cycle and by how many attempts recovery is allowed — and by nothing else. Every measured system in the corpus below spans 333 °C of ceiling temperature and a fourfold range of enthalpy; none of it appears here.

The exponent is what makes it bite. At 25 °C against 60 °C it is 0.117, so a pre-exponential term around 10¹⁶ is crushed to about two orders. The bound barely responds to argument: three decades of A and a 24× range of soak time move it by under 0.3 orders.

So the specification’s two-factor trigger is not a design preference that makes the chemistry more elegant. It is forced, and the bound says by how much. Base-catalysed cleavage is first order in hydroxide, so one pH unit is exactly one order of rate, and 4.1 orders converts directly into a process floor: the recovery tank must reach pH 11.1 before any polymer can satisfy the duty cycle. A caustic bottle wash at pH 13 clears it with nearly two orders to spare; a mild alkaline rinse at pH 10 cannot.

The bound covers the thermal channel only, and its loophole is the honest one: a change of mechanism between shelf and tank is exactly what it does not constrain — which is what a second gate is. Assumptions are Arrhenius kinetics with a temperature-independent pre-exponential and one rate-limiting step; over a 35 K span neither is worth four orders.

The lab inverts the two-sided duty cycle in closed form: a barrier must be high enough to survive two years on a shelf and low enough to release in a caustic wash tank. At the defaults that window is 94.2–106.6 kJ/mol, just 12.4 kJ/mol wide. Nothing in that derivation knew whether such a barrier is reachable by any linkage anyone has made.

The vitrimer literature answers it. 11 dynamic covalent networks, each with a bond-exchange barrier a stress-relaxation experiment returned: 1 inside, 9 too labile to hold a shelf, 1 too stubborn to release. The clustering below the floor is the real signal — a network built to rearrange under moderate heat is a network that fails a shelf-life test. Those are the same property read twice.

So the constraint is not obviously false. A candidate has to be engineered into this window; no published chemistry gets there by selection. And the gate is a process variable rather than a material one — swap the caustic wash for plain water and the window closes entirely.

The document’s cost-parity play is to graft cleavable release points onto legacy polyethylene by reactive extrusion. The corpus says how much cleavable unit that route actually gets into a chain, and the two published routes do not overlap.

And that floor is the optimistic one. A reagent cannot enter a crystallite, so cleavage is confined to the amorphous phase — and a locked bond does not merely fail to react, it welds together the two fragments that would otherwise have separated there. The reactor shows the two variables collapse into one: f_eff = f · (1 − χ), and monomer yield is a function of that product alone. Half the bonds cleavable at 40% crystallinity recovers exactly like 30% cleavable and fully amorphous — 21.96% against 22.90% monomer mass.

Which is the uncomfortable part, and it is a second kill on the same target, independent of the first: the crystallinity a film needs for barrier and stiffness is subtracted directly from its recoverability. Not a trade-off a better catalyst engineers around — arithmetic on the reachable bond count. It holds for reagent-mediated cleavage below Tm on a process timescale; melt the lamellae and it lifts.

Cleavable units partition a chain into blocks of unbreakable backbone averaging 1/f repeat units, so composition alone floors the fragments before any kinetics. No reaction time gets below it. A route whose product swings between wax and shorter plastic across its own reported composition window is not yet a recovery route — either the target moves or the strategy does.

The document calls the cyclic-versus-linear mismatch one of its most important blind spots and leaves it there. The numbers are larger than the framing suggests: cyclic PGBL reaches 409 °C of separation between its ceiling and its degradation onset, against PBTL’s 340 °C — with a monomer whose ring strain is so poor it was long classed non-polymerizable. Identical thermodynamics to linear PGBL, 72 °C more onset, purely from removing chain ends.

ΔH°p in kJ/mol, ΔS°p in J/(mol·K). Trap = degradation onset − ceiling: the room a catalyst has before thermal degradation competes.

Signs are restored by physics, not read off the page: propagation is exothermic and entropically penalised, so both are negative. The check is that Tc = ΔH°p/ΔS°p must reproduce the printed ceiling — all measured entries agree to under 6 °C under the corrected pairing, and two miss by 55 and 105 °C under the naive one. That identity is a unit test, so a re-extraction cannot silently undo it.

The reactor now represents this directly. A ring has no chain ends, so it cannot be activated for depropagation until an opening event creates them — a separate barrier with its own rate, which is what the literature is really reporting when cyclic PGBL needs 300 °C against linear PGBL’s 220 °C. Set the opening rate to zero and monomer yield stays at exactly zero however long the reactor runs.

The specification tracks two states: storage and recovery. There is a third, and it is the severe one. A film has to be extruded 20–40 °C above its melting point — around 240 °C for a packaging-grade iCP, nearly 200 °C hotter than the recovery tank the chemistry was designed around.

Recovery caps the catalysed barrier at about 107 kJ/mol. Put that same barrier at melt temperature and the half-life is 5 milliseconds. The polymer would be gone before it reached the die. So withholding the catalyst until recovery is not a migration-risk nicety — it is the only reason a melt-processable iCP can exist at all.

Which sets up the useful number. The uncatalysed barrier must clear ~165 kJ/mol to survive a 60-second pass, and the catalysed one must stay under 107 to finish recovery. The difference is what the catalyst has to deliver.

58 kJ/mol is demanding but not unreasonable — upper-range organocatalysis, squarely enzymatic — and it is checkable against a candidate catalyst before a film is ever made. The gradient is what matters for design: a hotter line needs a higher-melting, more crystalline polymer, so barrier performance is paid for in catalytic burden. That is the same conflict finding 03 reaches from the other side — crystallinity costs you reachable bonds there, and catalytic burden here.

And a negative, recorded because it was the hypothesis that started this. Processing does not kill the hero polymer. PBTL clears with 19 kJ/mol of margin and a residence budget of about 82 minutes against a 30–120 second extruder — though only ~60 °C of melt headroom. What survives is a screen: the hottest tolerable melt sits ~18 °C below the TGA onset, so Td − Tm ≥ 48 °C, from two numbers every paper already reports. Four of the six thermally characterised systems here report no melting point at all.

The 18 °C offset is stable in Td — it moves 17.1 to 18.2 across a 160 °C range — but it is not universal: it tracks ln(t_res/loss), falling to ~8 °C for a 30-second residence and rising past 45 °C for a 0.1% loss budget. The assumed pre-exponential largely cancels, since it appears on both sides; across five decades the margin moves under 8 kJ/mol.

Crystallites are impermeable, so a penetrant must detour around them and a reagent cannot get in at all. The same morphology that makes a film a barrier is the morphology that makes it unrecoverable — which turns the two requirements into one exchange with a closed form.

Because a reagent reaches only f·(1 − χ) of the backbone, and f cannot exceed 1, monomer recovery caps crystallinity at χ ≤ 0.5 for any chemistry. Substituting that into Nielsen tortuosity gives the most barrier a recoverable film can have:

About 27× at a lamellar aspect ratio of 50, and 5–408× across the whole literature range of aspect ratio and chain immobilisation. Above B* there is no (χ, f) that satisfies both — for any polymer, any catalyst, any reaction time.

So HCON 2 is not false, but it is not the binary the document implies either. It is a continuous exchange, and the exchange rate is computable: giving up monomer for a DP ≤ 5 oligomer lifts the barrier ceiling about 3.9×. That is the actionable form. The 90% virgin-equivalent target is what the barrier requirement is being paid for with.

And it explains something about the literature. PBTL is a polythioester — every backbone linkage is a cleavable thioester, so f = 1 by construction and its recovery is limited by crystallinity alone. A polymer whose selling point is intrinsic crystallinity with a melting point to 213 °C would need χ ≤ 0.5 to give monomer from the solid state. The published depolymerisations avoid this by running in solution, which dissolves the lamellae and removes χ from the equation. A legitimate escape, and not a free one: the cost lands exactly where the chemical-recycling lifecycle literature says it does. One polythioester in this corpus needed 1 g of polymer per 100 mL of solvent for full recyclability. The solvent is not an implementation detail — it is the price of the crystallinity the film needs.

No absolute transmission rate appears anywhere in this module, and a test enforces that. The corpus reports no OTR for any iCP; producing one here would be manufacturing the exact number the audit says is missing. Everything above is a ratio to the same chemistry fully amorphous, with α and β as stated inputs rather than fitted values.

The third decisive experiment — recover monomer from contaminated film, repolymerise, repeat — sounds like something to simulate. It is not. It is a recursion with a fixed point, and solving it says more than running it.

Each cycle recovers a fraction y of the mass and purges the rest; new impurity i arrives from contamination and from the degradation products the cycle generates; purification removes a fraction p. Impurity partitions with mass, so c_n+1 = (c_n·y + i)(1 − p), whose fixed point is:

Finite for any p > 0. So “zero multi-generational accumulation” is the wrong target — accumulation always happens and always converges. The question is whether it converges below the limit, and that is answerable before anyone runs a single cycle.

Set p = 0 and it collapses to c* = i/(1 − y) — the reciprocal of the purge. Yield loss is not only a cost; it is the loop’s only impurity outlet, and the specification’s ≥99.9% per-cycle recovery target closes it. That is a 1000× concentration factor on anything purification does not remove. The standard result that a recycle loop needs a purge, arriving where the circular framing does not expect it — because “circular” reads as “lose nothing”.

Which reframes the yield target. Once purification is real, the recovery yield barely moves the answer — holding the steady state at one cycle’s contamination needs about 50% removal per pass whether the yield is 90% or 99.9%. The lever is the separation train, which is exactly where the chemical-recycling lifecycle literature says the burden sits.

And the five-cycle test is either conclusive or actively misleading. Convergence is geometric, so with purification working the loop settles in two to five cycles and “at least five” is well chosen. Without it, five cycles reads 5.0 against an eventual 1000 — 0.5% of the problem, and reports a pass. Two loops reading within 14% of each other after five cycles can be heading 54× apart. The experiment has to report the purification efficiency alongside the impurity number, or it has concluded nothing.

c is a single lumped impurity in units of one cycle’s fresh input. That is deliberate: migration limits are per substance, so a real answer needs one recursion per migrant with its own i and p. The lumped form gives the shape — the fixed point, the purge inversion, the convergence rate — which is what generalises. Assumes impurity partitions with mass on purge, constant p across cycles, and no further reaction; the last is optimistic for anything that degrades into something worse.

The scoping document’s strongest recommendation for its second opportunity area is to bind it to the first: search monomer space jointly over polymerisability, barrier, shelf stability and triggered depolymerisation rather than one property at a time. The seven results above make that evaluable — every constraint computes exactly from properties, with no fitted model anywhere in the chain.

What it is not is a search over monomer space. That needs a map from structure to property — from a drawn molecule to ΔH°p, to a melting point, to a crystallinity — and no such map exists here. Building one would undo the discipline of everything above it. So this is a feasibility envelope over property space: is this set of properties admissible, and which constraint binds first.

Run it over the corpus and the answer is an absence. Every constraint can return “insufficient” rather than a plausible default, and that is mostly what happens:

No source in the corpus reports a crystallinity, a cleavable fraction, a purification efficiency or a recovery yield — for any of the nine measured systems. A screen that filled those in would return nine confident verdicts and every one would be fiction. Returning insufficient is the result, and it is the same shape as everything else in this audit: the binding constraint on deciding whether an intrinsically circular polymer can be packaging is not the physics and not the chemistry — it is that the literature does not report the properties the decision needs.

One thing IS settled before a chemistry is chosen. The process-side constraints do not depend on the polymer at all: the recovery tank must reach pH 11.1, the non-thermal gate must supply 4.1 orders, and the usable barrier window is 94–107 kJ/mol. That part of the design can be specified now, and it is the part a screening campaign should be built around.

The proposed hybrid: cut a plastic backbone electrochemically, feed the fragments to electricity-producing bacteria, harvest the current. Before building a simulator for that, one page of arithmetic decides whether the loop can produce electrons at all. Everything per repeat unit of backbone — the only basis on which the two halves are comparable.

First: cutting is cheap, so the loop cannot fail on cutting. A −CH₂− unit carries 6 electrons to CO₂. A two-electron cleavage every 29 units costs 1.15% of the electrons in the fragment — and 0.17% at the DP-200 end. The expert review predicted this ledger would come out negative because deconstruction spends electrons. It does not, and that retires the fix that prediction implies: making the cleavage catalytic rather than stoichiometric would optimise a term worth one percent.

Second: it fails on bioavailability, which is a different problem. Electricity-producing bacteria eat acetate, lactate, formate — small, oxygenated molecules. The melt-grafted polyethylene route floors its fragments at DP 29–200. A DP-29 alkane is not food for anything on the anode, so the ~174 electrons it contains are unreachable at any coulombic efficiency. The loop is dead there — because the electrons are inaccessible, not because they were spent.

Third, and this is the part that changes the answer: size is necessary and not sufficient. Cutting a polyolefin shorter produces a shorter alkane, and ethane is not fermentable either. Bioavailability needs chain length AND oxygen in the fragment, and cutting only delivers the first. Oxidising an alkane end to a carboxylate is a further 6-electron step.

So the two published cleavable architectures do not merely differ by 100× in composition — they differ in whether their fragments are edible at all. Ring-opening copolymerisation gives hydroxy-acids and lactones, already oxygenated: break-even at CE 0.167, just above the measured floor of 0.13. Radical insertion into polyethylene gives alkanes: charging the functionalisation quadruples the requirement to CE 0.667, which is the top eleven percent of everything ever measured on a bioanode.

A positive electron ledger is necessary, not sufficient. This says nothing about inhibition of the culture by fragments, catalyst carryover poisoning the anode — for which finding 07 is the right tool, since the catalyst is an impurity in the fermentation feed and the loop concentrates it — the missing fermentation step between cutting and the anode, or the applied potential the cleavage needs. It is an electron count, not an energy balance, and deliberately the cheaper of the two.

Finding 09 says bioavailability needs two things at once: a short enough fragment and oxygen in it. Two conditions is two axes, and drawing them turns a sentence into a picture — the region a depolymerisation reactor emits and the region a bioanode can metabolise do not overlap, and they miss on both axes at once.

The controls are the point rather than decoration. Raise the cleavable fraction, drop the crystallinity, and the emitted fragment travels left — that is cutting, and finding 03 already gives its floor at 1/(f·(1−χ)). On a hydrocarbon backbone the marker never reaches the box however far it goes, because the remaining distance is vertical, and no amount of cutting adds oxygen to a molecule.

Switch the backbone to oxygenated and the marker drops onto the box’s row immediately, at any composition — then it only has to travel left. That is the same separation finding 09 reached from the electron side, and it is why the two published architectures were never really competing: one of them is playing on a different axis.

What the picture names is the missing stage. Nothing in the proposed hybrid moves a fragment downward. Cutting moves it left; oxidation moves it down, and the only unit operation that does that at scale, cheaply, on a mixed stream, is a fermenter. So the honest architecture is not two tiers but three — deconstruction, fermentation, bioanode — and the middle one is absent from the specification, from the platform brief, and from this codebase.

The blank space between the marker and the box is deliberate and is the most important thing on the chart. How bioavailable a partially-oxidised C8 oligomer is to a mixed exoelectrogenic community is unmeasured. Shading it as a gradient would invent exactly the number the ledger refuses to assume, so the box has hard edges and the space between is left as what it is: unknown, and the first thing worth measuring.

Finding 08 ends on an absence: four properties no source reports for any polymer, and not one published iCP fully screenable. The obvious hope is that first-principles simulation supplies the missing inputs — and the last few years have made that hope much more reasonable than it used to be.

Open Molecules 2025 holds ~140 million DFT calculations over ~83 million unique systems, and crucially they are finite and reactive rather than infinite crystals — the right shape for ring-opening thermodynamics, where you need both sides of the reaction. Equivariant potentials (MACE, NequIP) turn those into force fields that run molecular dynamics at near-DFT accuracy, and tooling like mlip v2 and chemtrain-deploy carries them to million-atom systems.

So the question is not whether the machinery exists. It is which of the eight inputs the joint screen needs it can actually produce, and the answer is bounded:

The four properties no source reports are exactly the four simulation cannot produce either. Crystallinity is set by nucleation and cooling history — a rare event many orders beyond a nanosecond, and a processing history that is not in the Hamiltonian. Cleavable fraction is what a reactor incorporated. Purification efficiency belongs to a separations column. Recovery yield is a plant mass balance. None of them is a property of a molecule, so no potential, dataset or amount of compute closes them.

What it does close is the one axis the joint-feasibility module said it lacked — the structure-to-property map for ceiling temperature. That is the second opportunity area’s strongest real claim, and it is genuinely available now.

Two things the framing usually gets wrong, both worth carrying. Accuracy is not the binding axis: a 23-model benchmark finds large potentials buy 3–5 meV/atom over lightweight ones while losing orders of magnitude in throughput, with the worst cases barely faster than the DFT they replace — and for a screening campaign the value of a surrogate is how many candidates it lets you reject. And these models are trained on systems of ≤350 atoms, benchmarked on 192, and deployed on 10⁴–10⁶ — roughly 3.5 orders of extrapolation, which is precisely where they are least validated.

Finding 11 leaves exactly one input that first principles can supply: the ceiling temperature, since T_c = ΔH°p/ΔS°p. The obvious next move is to run a pretrained potential over the nine measured monomers and check it against their published ceilings. Before spending that, one piece of error propagation decides whether the check could succeed — and it turns on something the benchmarks do not publish.

ΔH°p is not an energy. It is a difference of two total energies — the strained ring, and the same atoms as an opened chain. Learned potentials are scored in meV per atom on absolute energies, and a difference of two large numbers inherits their errors:

The trouble is that ring strain in these monomers runs 5.4 to 21.1 kJ/mol — very small energies. At the 3–5 meV/atom the 23-model benchmark reports, on a 20-atom monomer with independent errors, σ(ΔH) is 13.6 kJ/mol. For γ-butyrolactone, the canonical intrinsically-circular monomer at −5.4 kJ/mol, the error bar is 253% of the quantity. The sign is not reliable, and T_c inherits the same relative error — hundreds of kelvin on a number whose whole interest is where it sits relative to room temperature.

The honest counterargument is that those errors are not independent. A potential evaluating a ring and its opened chain sees largely the same atoms in largely the same environments, so the errors should correlate, and correlated error cancels in a difference. At ρ = 0.99 the same case falls to 1.4 kJ/mol, which is perfectly usable.

So the question is not how accurate the potential is. It is how correlated its errors are across a ring-opening, and inverting for the correlation that would hold ΔH°p to 10% of its own value gives ρ ≥ 0.976–0.998, with the tightest demand on exactly the monomers the field cares about most — because they are the ones with the least ring strain.

No MLIP paper reports that correlation. Per-atom RMSE bounds it in neither direction — a model can have excellent absolute accuracy with uncorrelated errors, or mediocre accuracy with errors that cancel almost exactly. That is the useful result: it does not say learned potentials cannot do this, it says the published numbers cannot tell you whether they can, and it names the one measurement that would.

So instead of a result I could not honestly produce here — no ASE, no model weights, and nine monomers whose 3D structures would have to be built first — what is committed is the measurement itself, written down before anything runs: dataset, procedure, pass criteria, baseline, and what a failure would mean. The scoping document’s own critique of its second opportunity area is that “95% prediction accuracy” is undefined without a tolerance, a dataset, a prospective test and a baseline. This supplies all four, dated, so the criteria cannot be chosen after seeing the answer.

Finding 12 wrote the calibration down before running anything. This is the run: MACE-OFF23 on four lactones of unambiguous structure, by the incremental oligomer method, thirty conformers per species relaxed with the potential. Script and raw output are committed.

It failed. Against the one anchor with a citation — γ-butyrolactone at -5.4 kJ/mol — the prediction is -48.11 kJ/mol after correcting the largest artifact I could identify. The preregistration’s primary criterion, ρ, was never measurable at all: it needs a DFT reference this machine does not have.

But it failed somewhere I did not predict, and that is the useful part. Four separate setup effects each independently exceed the 5.4 kJ/mol being measured — and the potential’s accuracy, which finding 12 argued would be the binding constraint, is not the largest of them.

Two convergence checks support that reading rather than assuming it. On conformer count the answer moved −38.8 kJ/mol between 10 and 30 conformers and then not at all between 30 and 60 — so the minimum is converged and residual sampling is not the cause. On end groups, capping both hydroxyls shifted the answer +17.1 kJ/mol and roughly halved the conformer spread, confirming gas-phase intramolecular hydrogen bonding is real and sub-dominant.

So the correction to my own analysis: the energy model is not the binding constraint on computing a ceiling temperature. The thermodynamic protocol is. Exhaustive conformer search, end-group handling, extrapolation to the long-chain limit, thermal corrections — a DFT calculation set up as naively as this one would be wrong by a similar amount. Which reframes what a learned potential is actually for here: not being more accurate than DFT, but making a protocol careful enough to be correct cheap enough to run at all.

This is what preregistration is for. Without criteria fixed in advance it would have been easy to report that the model got the sign right and the four-membered ring most strained, and call that a partial success. The criteria were fixed, they were not met, and the run is recorded as a failure with its error budget itemised. Note also the licence: MACE-OFF23 is released under the Academic Software License, which does not permit commercial use.

Finding 13 itemised why the first run failed. Every item was then addressed, and PySCF was installed so the preregistration’s primary criterion — ρ, the correlation of the potential’s errors across a ring-opening — could finally be measured instead of assumed.

The fixes worked. Methyl-capping the chain ends removed the hydrogen-bond artifact and cut the increment drift from 19.9 to 3.0 kJ/mol. Replacing successive differences with a linear fit of E(n) on n returned R² = 1.00000000 over n = 2…5, so the long-chain limit is reached rather than hoped for. Adding a proper RRHO thermal correction shifted the answer +6.39 kJ/mol, positive as it must be — an opened chain has torsional modes a rigid ring does not. Error on the anchor fell 63%.

And ρ turned out to vindicate the hypothesis. Measured on lactone methanolysis against B3LYP/6-31G(d) — a real ring-opening, small enough for DFT on every species — the potential reproduces DFT with a mean absolute error of 1.65 kJ/mol. For three of four systems ρ_eff ≥ 0.998, comfortably past what the propagation demanded. The errors really do cancel across a ring-opening.

It fails on exactly one system, and it is the worst possible one: γ-butyrolactone, the least-strained ring in the set and the monomer this whole field is built around. The discrepancy is 5.5 kJ/mol — larger than γ-BL’s entire ΔH°p. The preregistered threshold is not met, on a specific and interpretable failure rather than a diffuse one, which makes it a fine-tuning target rather than a reason to abandon the method.

The decisive comparison is what locates the remaining 22 kJ/mol. On the same methanolysis reaction MACE gives -43.09 and DFT gives -37.6 — they agree with each other to 5.5 kJ/mol. Both are roughly 30 kJ/mol more exothermic than the published bulk polymerisation enthalpy of -5.4 kJ/mol.

Two independent quantum methods agreeing with each other and both differing from experiment by the same amount is not a model-accuracy problem. It is a phase problem. An isolated gas-phase chain is a poor model of a bulk polymer, and chain packing and restricted conformational freedom are precisely what make γ-BL’s polymerisation barely favourable in the first place. Gas-phase oligomers do not become a bulk polymer by getting longer.

So, corrected twice by measurement: the learned potential is accurate enough for this, except on weakly-strained rings. The binding constraint is the thermodynamic model of what a polymer is — and that has nothing to do with machine learning.

One known defect, stated rather than buried: the Hessians returned 3–6 imaginary modes per species. They were dropped from the RRHO sum, which is standard practice but means the zero-point energy is underestimated and the +6.39 kJ/mol thermal shift is a lower bound. And the DFT reference is B3LYP/6-31G(d), not the source’s ωB97M-V/def2-TZVPD — adequate as a consistent reference for measuring whether errors cancel, not adequate as reference data for the true value.

Finding 14 located the residual 22.4 kJ/mol as a phase error rather than a model error, which is a diagnosis and not a fix. This is the fix, and it was written down as a falsifiable prediction before it was run.

γ-butyrolactone is unusually polar — ε ≈ 39, polar enough to be sold as a solvent in its own right. The ring holds its C=O and C–O dipoles in alignment; the opened chain does not. So a dielectric should stabilise the monomer more than the polymer, making ring-opening less exothermic. Sign alone is a weak test, so the prediction was sharpened first: the shift should be largest for γ-BL and should fall with ring size. A uniform shift would have killed it.

It held, including the part that could have failed. The shift at ε = 39 runs +16.6, +17.0, +14.4, +13.0 kJ/mol for rings 4 through 7 — γ-BL highest, then monotonically down. Every curve rises, so the direction is right everywhere, and the ordering is right where it was risky.

Then the obvious shortcut turned out to be wrong by 44%. Those four numbers come from methanolysis — a model reaction, small enough for DFT on every species. Transferring its shift onto the real polymerisation is the natural move. Running the dielectric directly on the polymerisation instead — ring against methyl-capped n = 2 and n = 3 oligomers, 45 atoms at the largest — gives +11.83 kJ/mol, not +17.0. The model reaction overstates the correction by 5.16 kJ/mol.

The reason is structural. Methanolysis splits the ring into a small ester and a free alcohol — two separately solvated polar fragments. Polymerisation adds the unit to a chain, so the product side is one longer molecule with a smaller change in exposed dipole. A model reaction transfers its sign and its ordering reliably; it does not transfer its magnitude, and nothing short of running the real one reveals that.

And “82% closed” still flatters it. The residual −10.52 kJ/mol is nearly twice the size of the quantity being predicted (−5.4). The prediction went from 12× too exothermic to 3× too exothermic — real progress, and not yet a usable number for this monomer. Had the transferred shift been used instead, this section would have claimed 91%.

The limitation is not generic. ±10 kJ/mol is 10–20% on a normally strained monomer and would be perfectly usable. It is fatal only for weakly strained rings, whose ΔH°p is itself only a few kJ/mol — and those are exactly the monomers a T_c-targeting search cares about, because a small ΔH°p is what puts T_c near room temperature in the first place.

That is the same wall the ρ measurement hit from an unrelated direction: γ-BL was also the one lactone where the potential’s errors failed to cancel. Two independent measurements — error correlation against DFT, and absolute agreement after phase correction — single out the same system. The hard case is a property of the problem, not an artefact of one run.

What a mean-field dielectric cannot represent, stated rather than buried: ddCOSMO is a continuum, with no packing, no entanglement, no crystallinity and no explicit neighbour. A real melt has all four. ε was set to the monomer’s — right for the early reaction, wrong for a converted melt. Closing the last 10 kJ/mol needs periodic DFT on a packed chain, or COSMO-RS with a melt parameterisation. That is a different class of calculation again, and it is the only thing now standing between this and a working T_c predictor.

Finding 15 closed 82% of the error with a dielectric continuum and named exactly what a continuum cannot represent: packing, entanglement, crystallinity, an explicit neighbour. This computes the two phases directly rather than correcting between them — a polymer chain that bonds through the periodic boundary against a cell of liquid monomer, both held at their measured densities, both sampled at temperature.

Two things make this cleaner than every earlier attempt. A chain closed through the boundary has no ends, so the per-repeat energy is the long-chain limit — no methyl caps, no fit over n, no R² to report, nothing extrapolated. And a repeat unit and a monomer are both C₄H₆O₂: ring-opening is an isomerisation, so the two sides have identical atom counts and identical degrees of freedom, and every kinetic and vibrational term cancels exactly in the difference. The RRHO treatment — with the 3–6 imaginary modes that were the known defect of the gas-phase runs — is not needed and cannot contaminate the answer.

The result overshoots, in the opposite direction to everything before it. ΔH°p comes out at -1.54 ± 0.29 kJ/mol against a published -5.4 — the right sign now, and 3.9 kJ/mol short. The continuum was 10.5 kJ/mol too exothermic; this is 3.9 too endothermic, so the two bracket the measured value from either side.

Getting there took a second run, and the reason is worth stating. The first attempt returned +4.03 kJ/mol, holding the polymer at 299 K against a liquid at 278 K. ASE reads a thermostat’s friction in inverse ASE time units — about 10.18 fs each — so the value passed asked for a coupling ten times weaker than it looked. A scan then found a second cause: every coupling biases hot, falling monotonically with strength, because rescaling the relaxed cell to the bulk density leaves it strained and a weak bath cannot drain that fast enough. Rerunning at 0.05 fs⁻¹ cut the mismatch between phases from 21 K to 4.4 K.

An estimate made before that rerun put its value at about 3 kJ/mol. It was worth 5.57 — the estimate assumed only the temperatures would change, when a properly thermostatted liquid also samples a different structure. Error fell from 9.43 to 3.9 kJ/mol, a 59% reduction.

Two failures bracketing a value is worth more than one near miss. It says the true answer for a weakly strained monomer lies inside a 14 kJ/mol window set entirely by how the condensed phase is modelled — and that the residual is a phase-model error rather than a potential error. That is the conclusion finding 14 reached by argument; here it is measured, from both sides.

The validation is where this run withdraws an earlier claim. A cohesive energy tests whether the potential binds a condensed phase at all. Comparing the liquid’s thermal average against a 0 K minimised molecule gives 5.34 kJ/mol, which is wrong by construction — a thermal average against a potential minimum omits the molecule’s own vibrational energy. Correcting that reference on the first run gave 49.8 kJ/mol, inside the 48–52 Trouton band, and this page previously reported the potential as validated to about 5%.

That was flattered by the cold liquid. With the liquid properly at 299.7 K the same calculation gives 42.52 kJ/mol, about 12% below the band. The honest reading is that MACE under-binds this liquid by roughly a tenth — precisely what a finite cutoff with no explicit long-range dispersion term predicts. The bracket survives, because both phases carry the same class of error and it largely cancels in their difference; the “validated to 5%” claim does not.

Known defects, in order of how much they matter. A perfectly in-register single-chain crystal is not a semicrystalline polymer, and the real material is perhaps half amorphous — that is now the largest identified defect, and it points the same direction as the residual. The two phases still sit 4.4 K apart, worth about 0.7 kJ/mol, down from 3 but not zero. One liquid configuration, packed by rejection sampling and squeezed in stages rather than equilibrated from a melt. And the finite cutoff under-binds the liquid by about a tenth, quantified above rather than assumed away.