Bridging neuroscience, control theory and software engineering — a multi-scale neuromodulated substrate for task-adaptive code understanding.
· the complete dossier ·
SYMPHONY will establish the first neuromimetic knowledge substrate for software systems: a representation in which modules, functions, tests, commit history and design decisions are encoded as nodes in a multi-scale network whose activation patterns are reconfigured, on demand, by task-specific neuromodulatory signals. A code representation that behaves less like a document and more like a nervous system — foregrounding what the engineer’s current task needs.
Five objectives across 36 months deliver a four-layer extraction pipeline (M12), a neuromodulatory reconfiguration mechanism (M18), a low-bandwidth task-control interface (M24), a 20 % F1 lift over LLM and knowledge-graph baselines (M30), and a user study with at least 60 stratified engineers (M33). The consortium pairs Real AI, Newcastle (Ramaswamy), CREATE-PRISMA / UNINA (Siciliano) and UP Robotics.
SYMPHONY targets TRL 1–4 and aligns with EIC Pathfinder for a persistent, task-adaptive, auditable substrate of code understanding.
Current approaches to machine code understanding divide into two families, each with a structural ceiling we expect to hit within this decade.
Claude Opus 4.5 crossed 80 % on SWE-bench Verified in December 2025. Independent re-evaluation at ICSE 2025 Companion dropped leading SWE-agent configurations from 12.47 % to 3.97 % once solution leakage and weak test cases were removed. ICLR 2026 SWE-Bench+ replicated the collapse on SWE-agent 1.0 with Claude 3.5 (57.6 % → 31.8 %). Production LLM context windows remain in the low hundreds of thousands of tokens while industrial codebases span millions — retrieval augmentation supplies local relevance, not system-level coherence.
SonarQube, PMD, IntelliJ inspections, architecture recovery and architecture knowledge graphs capture what is explicitly encoded — call graphs, dependency edges, type hierarchies, declared interfaces. They do not capture the design rationale or the contextual activation of architectural knowledge. Avgeriou et al. (2007) identified that tier as the hardest software-engineering knowledge to externalise, and it remains so.
The widening gap between software-system complexity (roughly doubling every four years since 1970) and individual human comprehension capacity (close to flat across the same period) is the addressable problem. SYMPHONY proposes to close it not by enlarging the engineer, but by reshaping the substrate they navigate.
SYMPHONY’s advance is not to improve either of the two incumbent families but to combine their information content under a different organising principle drawn from biology. Both cortical neuromodulation in the mammalian brain and descending corticospinal modulation in the vertebrate motor system are mechanisms by which a fixed underlying network produces qualitatively different, task-appropriate activation patterns in response to low-bandwidth descending signals.
Structural, behavioural, historical and rationale layers unified in a single graph-resident representation built for activation-based retrieval, not query-based retrieval.
A single substrate state surfaces different subnetworks under an externally specified task token, using the four-scale neuromodulatory primitives of Mei, Muller & Ramaswamy (2022) as the mathematical template.
Borrowed from Siciliano-school haptic shared control — a narrow scalar control surface, not a prompt window. Composable, auditable, bounded. This is the property that makes the substrate amenable to industrial governance.
The category SYMPHONY seeds is the auditable code-comprehension substrate — a class of artefact that does not exist today and is structurally distinct from both the dominant LLM-agent platforms (GitHub Copilot, Cursor, Claude Code, Replit Agent) and the legacy static-analysis toolchains. Its novelty sits at the intersection of three mature fields that have not yet been combined in this way.
Mei, Muller & Ramaswamy (Trends in Neurosciences, 2022) formalised a four-scale framework for integrating neuromodulation into deep networks — hyperparameter, plasticity, neuronal and dendritic. SYMPHONY transposes that framework from its native perceptual and motor domains into a symbolic and structural domain (source code) where signals are discrete, hierarchical, and linguistic. This is the load-bearing scientific bet.
Selvaggio, Pacchierotti, Giordano & Siciliano showed across a decade of RA-L / ICRA / T-RO papers (2018–2025) that a low-bandwidth supervisory signal reshapes a high-degree-of-freedom autonomous controller into qualitatively distinct task behaviours — without rewriting the controller. SYMPHONY adopts that architectural primitive as the task-baton interface: a small set of scalar modulatory signals, intentionally narrow, so behaviour is composable, auditable and bounded.
Real AI’s Hominis programme builds foundation models for the real world — situated, auditable and compute-aware, trained at Leonardo / CINECA on EuroHPC allocation time. SYMPHONY uses that engineering base to assemble the substrate at industrial scale, but inverts the usual foundation-model deployment shape: instead of a fixed model serving prompts, the substrate is a slow, persistent representation whose activation is steered by the engineer’s task — a different system primitive.
At TRL 1–4 the question is whether the mechanism is sound, not whether an industrial artefact exists. Three independent lines of published evidence support it.
Simulation. Mei, Muller & Ramaswamy (2022) demonstrated, in simulation, that neuromodulatory units at the four scales produce the three behaviours SYMPHONY requires — faster adaptation, higher cumulative reward across task sequences, and resistance to catastrophic forgetting.
Hardware. The Siciliano-school programme (RA-L 2018; ICRA 2019; IROS 2019; RA-L 2020; T-RO 2022; ICUAS 2025; RAS 2025; JIRS 2023) and Caccavale & Finzi (TopiCS 2021, Autonomous Robots 2019) showed in physical hardware that low-bandwidth supervisory signals produce qualitatively distinct, context-appropriate behaviours from a single underlying autonomous controller — the architectural property SYMPHONY transposes.
Failure mode. The SWE-bench re-evaluations cited above, plus the context-window ceiling documented across LLM architecture-recovery work in 2024–2025, establish that the current dominant statistical approach is not on a trajectory to close the gap by incremental scaling alone. SYMPHONY is therefore not a scaling bet.
Each objective carries a quantitative threshold, a decision milestone, and a documented alternative path if the threshold is missed. RP1 closes at M12; RP2 closes at M36. The full objective table (with thresholds, milestones and partner leads) sits on the parent /synaptic/symphony page — section V.
O1 multi-layer extraction (M12) · O2 neuromodulatory reconfig (M18) · O3 low-bandwidth control (M24) · O4 benchmarked advantage (M30) · O5 equitable-access study (M33).
Real AI (NL) coordinates and leads O1, O4. Founded by Tarry Singh; builds Hominis on EuroHPC allocation time at Leonardo / CINECA. Newcastle (UK) — Sri Ramaswamy, chair in computational neuroscience, third author of Mei-Muller-Ramaswamy 2022; Blue Brain alumnus. Leads O2 and co-leads ethics for O5. CREATE-PRISMA / UNINA (IT) — Bruno Siciliano, director of PRISMA Lab, ERC Advanced Grant holder, Engelberger laureate; leads O3. UP Robotics (HR) — industrial-automation demonstrator codebase, supplies half the held-out task instances for O4 evaluation.
Downstream impact concentrates in four sectors where the cost of opaque, undocumented or unmaintainable software is already material and growing.
The European industrial base operates on code that routinely outlives the engineers who wrote it. UP Robotics confirms the pattern across its customer installations in Croatia and the wider region. A substrate that lets a maintenance engineer ask the engineering question directly, with a control surface procurement can audit, on real industrial code.
Telecommunications stacks, electricity-grid SCADA, public-administration legacy systems and rail signalling are dominated by software written across multiple decades and multiple contractors, with the implicit knowledge that bound it together largely retired. The cost of comprehension here is sovereign.
Code-comprehension and code-generation tooling is presently dominated by US-headquartered platforms operated under non-European governance. The 2026 EU policy push for sovereign foundation models and EU-jurisdiction inference creates a procurement-grade window for an auditable, European-jurisdiction code-AI stack.
Two decades of EU-funded computational artefacts — CERN analysis pipelines, ESA flight-software toolchains, EBI bioinformatics platforms — whose long-tail maintenance falls disproportionately on early-career researchers. A substrate that surfaces design rationale and historical context is precisely the sustainability tool the EU’s open-science programmes have been asking for.
The consortium has agreed in principle on a Real AI × UP Robotics joint venture as the primary exploitation vehicle for the SYMPHONY substrate. Terms are set out in the Consortium Agreement at grant preparation and refined into a binding term sheet by M40. Newcastle and CREATE retain academic ownership of their respective contributions — the neuromodulation framework and the haptic shared-control formalism — and grant the JV a non-exclusive, royalty-bearing licence for industrial use.
(i) The four-layer graph schema and the substrate’s neuromodulated activation mechanism — open under Apache-2.0 and CC-BY 4.0, with a single defensive patent application planned by Real AI to prevent enclosure by third parties.
(ii) The task-baton control formalism — Newcastle and CREATE retain academic IP; JV gets a non-exclusive, royalty-bearing licence for industrial deployment.
(iii) Industrial integration code and SCADA-specific adapters — proprietary to the JV, sold or licensed to industrial customers under standard commercial terms.
(iv) Benchmark datasets and the O5 user-study protocol — open under CC-BY 4.0 to support replication and meta-research.
EIC Transition follow-on. Primary follow-on instrument is EIC Transition, applied for in Year 4 (target submission window M37–M40, immediately after project end). The Transition application will use SYMPHONY’s O4 quantitative advantage and O5 user-study evidence as its core scientific case, with a TRL-5 industrial pilot at UP Robotics as its technological case.
Regulation, certification, standardisation. Three regimes bear on downstream deployment and are tracked from project start: the AI Act’s high-risk classification (a code-comprehension substrate in regulated industries likely falls within Annex III); the Cyber Resilience Act’s software-of-unknown-pedigree provisions; and emerging code-attribution obligations. Working notes are contributed to CEN-CENELEC JTC 21 (AI) by M30 and ISO/IEC JTC 1/SC 42 by Year 4.
SYMPHONY is a deep-tech research bet whose value compounds on three layers: the science, the European procurement window for sovereign code-AI, and the commercial JV that the consortium has committed to from grant signature.
Independent re-evaluation has publicly cracked the SWE-bench narrative. The incumbent statistical approach to code understanding is documented as not on a trajectory to close the comprehension gap by scaling alone. The market is short an architecturally distinct alternative, and the EU is short a sovereign one.
The mathematical primary source for the mechanism (Ramaswamy, Newcastle) and the architectural primary source for the control surface (Siciliano, CREATE-PRISMA) are co-investigators. The industrial coalface is supplied by UP Robotics. Real AI integrates and coordinates. The pedigree depth is the moat — it cannot be assembled by a competing team in less than five years.
A JV term sheet executed by M40; an EIC Transition grant secured by Year 5; a defensive system-architecture patent filed; at least one industrial pilot beyond UP Robotics' existing customer base contracted; the JV cash-positive on bilateral industrial contracts under both the project EIC follow-on and the post-project commercial path.
The EIC Investor Programme and Real AI’s existing European deep-tech investor network are the two routes for parallel private capital. Targeted briefings are part of the project dissemination plan from M18.
Whether multi-scale neuromodulation — demonstrated in continuous perceptual and motor domains characterised by embodied feedback — transfers to a symbolic and structural domain (source code) where the signals are discrete, hierarchical, and linguistic. This is not a question of engineering polish; it is a question of whether the biological principle generalises. The five objectives are constructed so their decision milestones surface a clear answer within 36 months.
A substrate trained on a proprietary codebase could externalise design rationale in ways that defeat normal IP protections. Mitigation: technically, the substrate is bound to its training codebase and cannot be queried about unexposed projects; legally, training corpora are bounded by Consortium Agreement IP clauses.
A substrate that surfaces answers too easily can erode the apprenticeship value of unfamiliar code. The O5 user study detects this directly: a pre-registered secondary endpoint measures 30-day retention with a non-inferiority margin such that a positive speed effect cannot come at the cost of substantive deskilling.
DNSH commitment in §1.3 already binds SYMPHONY to per-paper tCO₂e reporting (Green Algorithms methodology) and an absolute compute-budget cap declared in the DMP. Make this a published metric in M30 / M36 deliverables; an over-budget month triggers a documented re-plan rather than overrun.