The Manifest sets out the identity, the work, the practice, and the operational principles of the Institute. It is a working document, versioned, philologically accurate by design.
The Institute reads the inverse problem as a precise operational act. In every instance of work, two functional roles are distinguished and brought into relation: a material to be inverted — the heterogeneous evidence on which the conversion operates — and a functional receiver — the structured model that accepts the inverted configuration and uses it. The two roles are not properties of the entities involved; they are states that materialise only at the moment of the inversion.
In the classical inverse-problem literature the distinction is implicit: data on one side, model on the other, an estimator in between. We make the distinction explicit because the Institute works on two methodological lines in which the material side is filled by very different apparatuses, while the receiver side remains the same category of structured physical or process model.
Measurements, structural priors, and accumulated knowledge of the system are converted into parameter values, closure conditions, and boundary structures. The receiver is the deterministic model — ODE compartmental, PDE distributed-parameter, state-space — augmented where needed by neural functional closures (PINNs, UDEs, graph neural networks).
A hybrid syntactic-vector RAG indexing a domain corpus, queried under LLM orchestration, is the material from which the inverted configuration is extracted. The receiver is the same category of mechanistic model as in Line B — a simulator that accepts the parametric profiles, initial conditions, and boundary structures the RAG-LLM has assembled.
The separation between material and receiver is not permanent. It materialises for a single instance of work and must be re-activated for any subsequent instance. The same RAG can serve a different receiver tomorrow; the same physical model can be calibrated against a different material next week; a quantity that is "material" in one inversion act can become "receiver-internal structure" in the next.
This is not a weakness of the formalism — it is its defining property. The inversion act is an episode, not a permanent system. The Institute's outputs are organised around discrete inversion episodes, each fully documented as such: which entity played which role, which corpus or dataset was the material, which model received the configuration, and what parametric structure resulted.
Without the community having explicitly announced it, the inverse problem has become the hidden mathematical structure of much of what we today call artificial intelligence. Generative models learn priors for classical Bayesianism. Neuro-symbolic architectures reformulate Tikhonov regularisation in the language of neural networks. The last two years add a development whose methodological consequences are still being absorbed: large language models, when orchestrating retrieval over heterogeneous documentary corpora, function as a new kind of material on which inversion can operate.
The Institute exists as a place of bidirectional translation between two traditions: the historical inverse-problem tradition, rich in theory but marginalised in the spaces where current innovation runs; and the contemporary currents of AI, rich in methods but poor in a unifying mathematical language. We aspire to be one of the few places where the two meet productively — in specific inversion episodes, on concrete substrates, with contemporary tools.
The Institute is small. It operates on two specific lines, each articulated in three sub-areas, where the founding nucleus has direct competence.
A hybrid syntactic-vector RAG indexes a high-value documentary corpus (EMA EPAR + Product Information; CNR-IRSA Verbania zooplankton archive; Lake Varese documentation). Under LLM orchestration, the system extracts and assembles the structured configuration the receiver requires.
A simulator that accepts the inverted configuration: parametric profiles, initial conditions, boundary structures, age-structured demographics. The receiver consumes what the material produces, and produces the simulated dynamics that the institutional stakeholders use.
The LLM orchestrates which evidence is retrieved when, weights its relevance, and produces the parametric structure. The conversion is mediated by the symbolic regularisers the corpus carries (citation graphs, regulatory schemas, structured domain ontologies). The receiver's cost function alone could not reach this information.
Line A reads the EMEA work (1995-2003) as the proto-form of this programme: every dossier of a pharmaceutical product was already, in formal terms, an applied inverse problem. The TRAG prototype now operates on the same corpus with hybrid RAG + LLM apparatus.
Sub-areas: methodological foundation TRAG; applied projects A·1 Daphnia Lake Maggiore (compartmental ODE receiver) and A·2 GLM-AED2 Lake Varese (lake biogeochemistry receiver); methodological collaboration with the Leverhulme Centre for Wildfires.
Direct measurements (time-series, stage histograms, abundance counts, radiotracer kinetics, fissile material balances) and accumulated knowledge of the system (structural priors, physical constraints, demographic regularities). Heterogeneous, partial, structurally constrained.
A deterministic model — ODE compartmental, PDE distributed-parameter, state-space — augmented where needed by neural functional closures (PINNs, UDEs, graph neural networks) that supply the components classical formulations had to assume or approximate crudely.
Recovery of vital rates, flow parameters, closure conditions, fissile material balances under correlated measurement noise. The symbolic constraint of the receiver (conservation laws, transport equations) guides the estimation; the neural closure absorbs the residual non-mechanistic component.
Line B continues directly the trajectory: 1973 IFAC, 1976 Brebbia, 1979-1983 expansion across nuclear safeguards / population biology / entomology / metrology, 1987 Memorie IIDr 45, 2008 Manca et al. (contribution analysis), 2026 Journal of Limnology.
Sub-areas: B·1 Cladocera Lake Maggiore; B·2 Material accountancy & pattern recognition; B·3 Radioecology Zn-65.
The two lines are not separate research programmes that happen to share an address. They are two parameterisations of the same inversion act. In both, a functional receiver — a physical or process model — is fed by a material whose conversion is the methodological problem. Line B is the consolidated form (the material is direct measurement, the conversion is statistical estimation guided by symbolic constraint). Line A is the frontier form (the material is a documentary corpus under LLM orchestration, the conversion is mediated retrieval guided by symbolic regularisers in the corpus itself).
The Institute's applied work on Lake Maggiore is the place where the two parameterisations converge on the same biological substrate. B·1 inverts time-series of Daphnia counts into demographic parameters with PINN-closed ODE models. A·1 inverts the 30+ years of Manca's heterogeneous documentation into the egg age-structures that those same demographic models require. The two episodes use different materials, the same receiver category, the same ecosystem.
The founding nucleus has practised the methodology of applied inverse problems for four decades across three successive divisions of the European Commission, in continuous collaboration with first-rank international institutions.
Formed in the 1980s, coordinated by the future Founding Director, with Bernard Meltzer as full-time Senior Advisor. Seven PhD students from Genova, Strathclyde, Edinburgh. Italian network included the Politecnico di Milano AI group (Somenzi, M. and D. Gini). Argentesi received an Honorary Fellowship from Strathclyde.
Analogical naive physics; Case-Based Reasoning for documentary systems; computational linguistics; robotic planning. Stack: Symbolics 3600 LISP Machine + JRC IBM mainframe.
The political-institutional decision to constitute the laboratory was Bonnaure's — a senior JRC director who moved resources from the declining nuclear programme to the nascent AI Laboratory at the end of the 1970s.
Parallel collaboration with the JRC entomology group on fruit-fly population dynamics (Athens 1982-83 chapters).
The founding nucleus moved into the new JRC Computational Nuclear Safeguards Division. Close collaboration with Jim Shipley's group at Los Alamos National Laboratory on Kalman filtering and pattern recognition for fissile materials control; and with the IAEA in Vienna on intelligence-based documentary analysis of inspection reports.
The statistical accountancy software for fissile material balance, based on the framework articulated in Argentesi, Casilli & Franklin, 1st ESARDA Annual Meeting, Brussels, 1979.
Euratom Luxembourg exercised directly the EU safeguards on behalf of IAEA; JRC Ispra was its R&D arm. Argentesi coordinated contracts with NUKEM Hanau, BNFL Risley, Christian Rovsing Copenhagen, and Reading (Curnow/Woods, source of SITMUF).
The founding nucleus was given charge of the European Technical Office for Medicinal Products, with the task of developing the regulatory software of the new European Medicines Agency (then EMEA, today EMA). The division was transferred to London in 1995 at the start-up of the Agency. The Founding Director served until 2003; Franco Rinaudo, now Co-Founding Director, was the first system engineer of the Agency.
The same EMA corpus is the substrate on which the TRAG prototype of Line A operates — a direct methodological continuity from the regulatory informatics of the 1990s to the LLM-orchestrated RAG architectures of 2026.
After 2003, members of the founding nucleus operated individually: system engineering at the European Commission, Brussels (F. Rinaudo, now retired); Head of Documentary Analysis Division at the IAEA, Vienna (L. Costantini); pharmaceutical-informatic consulting (F. Argentesi). The 2026 Institute consolidates and reopens that practice within the contemporary AI methodology.
Four concentric layers, deliberately kept simple. The two Directors set scientific direction jointly; Founding Fellows contribute on a sustained basis; Associates collaborate on specific outputs; the In Memoriam category recognises foundational figures who can no longer participate in person. The full roster is on the People page.
The Institute does not appear in a vacuum. Six anchor points trace the formal inverse-problem lineage its founders have practised.
Argentesi, Di Cola & Verheyden at the 3rd IFAC Symposium on Identification and System Parameter Estimation. First formal article of the lineage.
Argentesi, Di Cola & Guerri at the first international conference on applied numerical modelling. PDE, Tikhonov regularisation, adjoint system, gradient optimisation.
Brussels 1979 (NUMSAS), Parma 1981 (Kalman entering ecology), Athens 1982-83 (Cavalloro), Statistica Applicata 1982 (Canali). One framework, four substrates.
Argentesi, de Bernardi, Di Cola & Manca — comparative analysis of estimators for cladoceran birth rates. Foundational baseline of B·1.
The largest historical exercise in computer-assisted documentary inverse inference in Europe. Substrate of the contemporary TRAG prototype.
Two convergent methodological languages: PINNs/UDEs/GNNs for Line B and LLM-orchestrated RAGs for Line A. The inverse-problem structure is unchanged; the regularisation channels have expanded.
An informal collective of senior independent researchers, internet-native, with no physical premises. No legal personality, no employees, no funding obligations, no degree-granting authority. Light enough to move quickly; identifiable enough to be cited; distributed enough that no institutional cycle determines its direction.
Not a university and does not simulate one. Not a startup, has no commercial mandate. Not a consortium and does not lead first-line Horizon Europe programmes. Not an advocacy organisation. Not a member-recruiting body.
Peer-reviewed papers in venues where each theme is best read. Technical Notes self-published with persistent DOIs via Zenodo. Open-source software, datasets, and reproducibility material through public repositories. Methodological commentary.
Maximum philological accuracy in the representation of facts, dates, roles, and institutional affiliations — both its own and those of the figures it acknowledges. The Institute reports what is verifiable, no more, no less.
To move from manifest to functioning Institute, three quarters of operational steps in the first twelve months.
Formal co-signature of the Manifest by the Co-Founding Director. Confirmation of Founding Fellow status. Setup of institutional infrastructure (hosting, email, GitHub, Zenodo). Publication of v1.0 of the website.
Argentesi 2026 (post-acceptance at Journal of Limnology) as Technical Note 2026/1. Technical Note 2026/2 on TRAG as inverse-problem architecture for EMA. Dialogue opening with CNR-IRSA Verbania. Methodological dialogue with Leverhulme.
First peer-reviewed papers under joint Institute affiliation. Progressive recovery and re-publication of archival outputs. First open seminar. Review of the Manifest at one year and publication of v1.0.
The 1980s JRC AI Laboratory of Ispra, where the founding nucleus was formed, has not been systematically registered in the mainstream historiography of artificial intelligence. The Institute, while not constituted as an historical-archival body, considers that the laboratory's role — as a continental node of the European first-generation AI network, peer of Edinburgh, Milano, Pisa, and connected directly to the founding Edinburgh tradition through Bernard Meltzer — deserves documentary preservation. The Institute will progressively make available on its website such testimony and documentation as the founding members can contribute.
Cambridge UK, June 2026
Flavio Argentesi, Founding Director
Co-signature by Franco Rinaudo, Co-Founding Director — pending formal acceptance