CognisCognis network
Engineered solutions

The walls other networks hit — engineered through

Where conventional stacks give up, Cognis Weave brings the frontier to bear: semantic communication, entanglement-assisted capacity, channel engineering, advanced coding, and a formal logical foundation. Every technique below is grounded in shipped, measured, or peer-reviewed work.

Where conventional stacks give up, Cognis Weave brings the frontier to bear. Three engineering levers do the heavy lifting. We change the question: for AI traffic you rarely need bit-exact bytes — you need meaning preserved, so semantic and task-oriented coding (DeepJSCC, information-bottleneck features) delivers the payload in a fraction of the channel uses that literal delivery would demand. We change the resources: pre-shared entanglement and quantum key material lift capacity and secrecy beyond what any classical-only link can reach, and repeaterless QKD records already prove the distances. We change the channel itself: reconfigurable intelligent surfaces, holographic and massive MIMO, and space-division multiplexing raise the ceiling of the medium before a single bit is sent. Underneath all three is a rigorous logical foundation — the Lambek calculus and the linear-logic/session-type correspondence — that turns 'these pieces compose correctly' into a machine-checked property. Every capability below is built on shipped, measured, or peer-reviewed work, and every one is a certified Cognis primitive.

Problem → solution

Every hard problem, handled

The obstacles that stop TCP, Tor, BitTorrent and LoRaWAN — and exactly how Cognis Weave gets through each one.

Shannon rate-distortionSolved

The move: Semantic / task-oriented communication (DeepJSCC, NTSCC+, Information Bottleneck): transmit a representation optimized for a declared downstream task and distortion measure, not for byte-exact reconstruction.

How it works: Change the fidelity criterion. Under a task-sufficient distortion measure the effective source is the IB minimal sufficient statistic, whose entropy is far below the raw byte entropy, so far fewer channel uses are legitimately required — fully inside Shannon's rate-distortion framework, not outside it. NTSCC+ (Wang et al., IEEE JSTSP 17(4):1022, 2023) is the first end-to-end system to beat VTM+5G-LDPC on PSNR.

In Cognis: Exposed as the task-distortion-transmit primitive gated by a fidelity-contract: a stream declares lossless / task-sufficient / perceptual-generative, and a receiver that actually needs bit-exactness refuses or downgrades the co-trained codec.

Shannon separation theoremOptimized

The move: DeepJSCC / DeepJSCC-f: a co-trained joint source-channel mapping with no error-free bitstream in the middle, giving graceful degradation (no cliff) and ~3 dB low-SNR PSNR gains (Kurka & Gunduz, IEEE JSAIT 2020).

How it works: Separation is optimal only asymptotically for stationary-ergodic point-to-point links with unbounded blocklength and delay. In the finite-blocklength, low-latency, non-ergodic regime it is provably not tight (and in multi-user networks provably suboptimal); DeepJSCC realizes the joint optimum the separated stack leaves on the table. It APPROACHES the JSCC optimum — it does not beat the fundamental joint limit, and its edge over separation shrinks to zero asymptotically.

In Cognis: The graceful-degrade-link transport policy: carry payloads over a joint mapping that degrades smoothly under adverse/variable channels within a tight latency budget, with entropy-guided-rate-allocation spending more channel uses on informative latents (NTSCC-style).

A noisy quantum channel's classical rate is capped by its Holevo (HSW) capacity.Beyond classical

The move: Entanglement-assisted classical capacity: pre-shared entanglement lifts the achievable classical rate above the unassisted ceiling — demonstrated at +16.3% over a lossy/noisy bosonic channel (Hao et al., PRL 126, 250501, 2021).

How it works: Add a new pre-shared resource. Entanglement-assisted capacity is a mathematically distinct, larger capacity (Bennett-Shor-Smolin-Thapliyal); the assisted/unassisted ratio grows without bound as the channel gets noisier. No theorem is violated — the 'break' is honest only because entanglement is counted as an extra resource, and the advantage is largest exactly where absolute rates are lowest.

In Cognis: The EntanglementAssistedChannel primitive: two agents burn a pre-established shared-context/correlation budget to communicate above their nominal bandwidth, with the accounting making the consumed resource explicit.

Channel capacities are additive; a zero-capacity channel is useless.Beyond classical

The move: Superadditivity / superactivation: two channels together carry strictly more than the sum, and two zero-quantum-capacity channels can jointly have positive capacity (Hastings 2009; Smith-Yard, Science 2008; Cubitt-Chen-Harrow).

How it works: Quantum capacity is not a single-letter additive quantity — impossible for classical Shannon channels. This is the deepest genuinely-beyond-classical result here; large-scale experimental exploitation is still nascent (recent one-shot zero-error superadditivity from Kochen-Specker contextuality, arXiv:2601.08913, 2026).

In Cognis: The CapacityActivation combinator: compose two individually-lossy or individually-useless links/sources into a working one, encoding non-additivity as a first-class operator rather than assuming capacity adds.

PLOB boundSolved

The move: Twin-field / MDI-QKD: interfere pulses at an untrusted midpoint so key rate scales as sqrt(eta) — TF-QKD demonstrated over 1002 km fiber (Liu et al., PRL 130, 210801, 2023); MDI-QKD over 442 km (PRA 108, 022605, 2023).

How it works: Change the channel topology. PLOB is specifically the REPEATERLESS point-to-point bound; TF-QKD beats it by inserting a measure-in-the-middle node (effectively one relay), converting linear loss scaling to square-root without granting the node trust. It reframes the channel; it does not defeat the loss physics — and the 1002 km rate (~1e-11/pulse) needs SNSPDs at ~0.02 Hz dark counts.

In Cognis: The UntrustedMidpointRelay topology primitive plus InfoTheoreticKeyAgreement: insert a zero-trust rendezvous node to beat the direct rate-vs-distance wall, with trust guaranteed structurally (no-cloning / Bell violation) rather than conjecturally.

A wireless link is capped by C = B*log2(1+SNR); you cannot push a single stream past it.Optimized

The move: RIS / massive & cell-free MIMO / near-field beamfocusing: reshape H, raise SNR by 20+ dB (RIS field trial, FITEE 2024), and open LoS spatial DoF > 1 via spherical wavefronts (Bjornson et al., arXiv:2209.03082).

How it works: Attack the three quantities the bound is defined over, all channel properties, not the code. SNR moves the operating point up the log curve (~10*log10(N) dB coherent gain); MIMO capacity is a SUM of per-eigenmode Shannon terms, so more orthogonal modes add terms; near-field focusing gives even pure-LoS channels rank > 1, overturning 'LoS = one stream.' Each stream still obeys Shannon.

In Cognis: ReflectiveRelay (RIS analogue), CoherentCombine (beamforming), SpatialMultiplex (MIMO rank), NearFieldFocus (address by direction AND distance) and WaveDomainCompute (SIM: run the precoding math in the medium at light speed).

A single fiber core saturates near ~100 Tb/s (nonlinear Shannon limit from the power/ASE/Kerr tradeoff).Optimized

The move: Space-division multiplexing: 22.9 Pb/s in a 38-core/3-mode fiber (NICT, ECOC 2023); deployable 19-core standard-cladding at 1.02 Pb/s over 1808 km; first 2-core commercial submarine MCF (TPU, RFS 2025-2026).

How it works: Change the resource, not the bound: pack N cores and/or M modes so aggregate capacity scales ~N*M x per-channel capacity, each channel still obeying the same nonlinear Shannon limit. Genuine parallelism, not a loophole. OAM is a subspace of the spatial mode set (Zhao et al., Sci. Rep. srep27674, 2016) — a basis choice that lowers decode cost, not a new dimension.

In Cognis: SpatialFanout (N independent streams, aggregate = N x per-stream limit, explicitly labeled parallelism) with CrosstalkEqualizer for the coupled-core case (MIMO-DSP analogue) and an OrthogonalBasisSelector that treats basis choice as an efficiency knob, never a capacity gain.

At short blocklength you pay the finite-blocklength penalty R*(n,e) ~ C - sqrt(V/n)*Q^-1(e); the meta-converse is a hard upper bound.Optimized

The move: PAC codes (128,64) essentially meet the dispersion/RCU bound; learned codes (KO, DeepPolar+, ProductAE, TurboAE) beat the specific classical codes that defined short-block SOTA by ~0.7-1.8 dB.

How it works: Two honest categories. PAC/SPARCs/spinal APPROACH the FBL optimum or achieve capacity asymptotically; ML codes BEAT classical codes (RM/polar/turbo) at matched rate/length, closing slack to the FBL bound — not surpassing information theory. The one true reframing is feedback/rateless: it operates against a different, higher curve the no-feedback bounds assume away.

In Cognis: RatelessChannel (take exactly the symbols the link needs, ACK to stop), LearnedNonlinearEncoder, SpatialCoupling, and a FiniteBudgetOptimalityGauge that scores every pipeline against its meta-converse/RCU bounds at the actual (n, error target) — exposing real slack, not a misleadingly small 'gap to capacity.'

Anonymity trilemmaEngineered around

The move: User-coordination rounds provably beat the ORIGINAL 2018 bound (Das et al., PoPETs 2020); PIR (SimplePIR ~10 GB/s, YPIR, Piano) and DPF systems (Express, Spectrum) reframe what is protected; BBS credentials prove a predicate instead of emitting an identifier.

How it works: Change the terms. Coordination adds a synchronized shared resource the 2018 model omitted — but the SAME authors immediately prove a STRONGER impossibility that re-binds all known ACNs. PIR moves privacy from the network channel to computation (retrieval privacy, a different quantity), and beats the linear-server bound only by adding preprocessing that Beimel-Ishai-Malkin explicitly excludes.

In Cognis: coordination-round (the one lever that beats the base bound, with the stronger-impossibility caveat attached), private-fetch (PIR), dpf-write, oblivious-store, cover-channel (DP budget made explicit, no false 'infinite' claim), and anon-credential (BBS predicate proof).

Snapping two components together carries no interoperability guarantee; deadlocks and protocol mismatches surface only at runtime.Solved

The move: Session types as linear-logic propositions (Caires-Pfenning, CONCUR 2010; Wadler CP/GV, ICFP 2012): a channel IS a proposition, a process IS a proof, composition IS the cut rule, and cut-elimination gives deadlock-freedom as a corollary (machine-checked, LinearActris, POPL 2024).

How it works: Change the channel of guarantee from runtime testing to static proof. Substructural/linear logic (Lambek 1958 -> linear logic) forbids weakening and contraction, so a channel is used exactly once, matching real messages; cut-elimination makes progress a theorem. The guarantee is bought by RESTRICTION — it lives inside the typed, linearity-respecting, acyclic fragment; cyclic topologies need Priority CP and do not evade general concurrency undecidability.

In Cognis: LinearChannel, Cut, CoherenceCheck (n-ary multiparty), FunctorialAdapter (protocol refinement as a functor preserving F(g.f)=F(g).F(f)), StringDiagram and OpenSystemCospan — the categorical spine that makes every other Frontier primitive compose with a proof, not a test.

No optical RAM and no cheap all-optical logic, so you cannot build an all-optical router/computer (delay-bandwidth-product wall; Miller's cascadability/fan-out criteria, Nat. Photonics 4, 3, 2010).Engineered around

The move: Reframe the architecture: optical circuit switching (Google Apollo, tens of thousands of 136x136 3D-MEMS OCS, ~108 W vs ~3 kW) and analog photonic tensor cores running ResNet/BERT near FP32 (Lightmatter, Nature 2025, 65.5 TOPS at ~78 W).

How it works: Change the question and the channel role, not the physics. Circuit switching never buffers or inspects packets, so it needs no optical RAM; analog MVM uses optics for the fixed parallel linear algebra it is good at and keeps memory/nonlinearity/control in electronics; co-packaged optics uses light as interconnect, exactly Miller's prescription. The delay-bandwidth-product and cascadability limits remain fully intact — dense optical RAM still does not exist in 2026.

In Cognis: CircuitReserve (dedicate a contention-free lane instead of buffering), AnalogLinearOffload (send the big fixed linear transform to a fast approximate channel, keep state/nonlinearity/control in the reliable one), WavelengthParallel, and a BoundedDelayLine that admits its hard delay-bandwidth budget instead of pretending to be RAM.

Key security rests on computational hardness and is only conjecturally safe against a future (quantum) adversary.Beyond classical

The move: Information-theoretic key agreement: QKD (TF/MDI/DI-QKD) whose security rests on physical law — no-cloning, monogamy of entanglement, loophole-free Bell violation (Nadlinger et al., Nature 607, 682, 2022 for DIQKD).

How it works: Change the security basis with a new physical resource. Eavesdropping is detectable as physical disturbance; DIQKD certifies integrity from a Bell violation treating devices as black boxes. Notes: information-theoretic security is real but sits behind vulnerable trusted nodes, side channels, and documented satellite implementation attacks; DIQKD runs at trickle key rates over short ranges.

In Cognis: InfoTheoreticKeyAgreement plus DeviceIndependentCertification: verify an untrusted component/sub-agent's honesty from its correlation statistics alone (Bell-type bound), without inspecting its internals — a structural rather than conjectural trust primitive.

Semantic & Task-Oriented Communication

Conventional wisdom says lossless is impossible below source entropy. True — for bit-exact delivery. Frontier does not fight that theorem; it reframes the target. Semantic-lossless means the meaning survives with residual task/semantic distortion below a declared threshold, and for AI-to-AI traffic that is the criterion that actually matters.

The mechanism is Weaver's three levels operationalized (Gunduz et al., IEEE JSAC 41(1):5-41, 2023): Level A reproduces symbols, Level B conveys meaning, Level C causes an effect. Task-oriented coding trains the encoder to keep only what a downstream task needs, formalized as a Variational Information Bottleneck rate-distortion tradeoff (Shao, Mao & Zhang, IEEE JSAC 2022). Under a task-sufficient distortion measure the effective source is the minimal sufficient statistic — far lower entropy than the raw bytes — so you legitimately transmit far fewer channel uses. This stays inside Shannon rate-distortion; it does not break the floor, it moves to a lower-entropy source.

On the channel side, DeepJSCC (Bourtsoulatze, Kurka & Gunduz, IEEE TCCN 2019) keeps a continuous joint source-channel mapping instead of a fragile error-free bitstream, giving no cliff effect — graceful PSNR-vs-SNR degradation — and, with feedback, ~3 dB low-SNR gains (Kurka & Gunduz, IEEE JSAIT 2020). NTSCC+ (Wang et al., IEEE JSTSP 17(4):1022, 2023) is the first end-to-end system to beat VTM (VVC intra) + 5G-LDPC on plain PSNR. Generative semantic comm (mm-GESCO, arXiv:2408.05455) reaches ~200x compression by synthesizing a perceptually faithful — not faithful — reconstruction.

Verdict. No theorem broken. DeepJSCC approaches the JSCC optimum that separation leaves on the table in the finite-blocklength / low-latency / non-ergodic regime; asymptotically its edge over separation vanishes. Task-oriented comm reframes the question. The hype to discount: 'beats Shannon,' 'infinite compression,' and counting a generative decoder's hallucinated detail as delivered information. Hard limits: gains need a matched co-trained codec, are brittle to distribution/channel shift, and are lossy w.r.t. anything outside the trained task — so the fidelity-contract primitive must refuse to substitute a semantic codec where bit-exactness is genuinely required.

Certified primitives: task-distortion-transmit, graceful-degrade-link, entropy-guided-rate-allocation, generative-reconstruct, fidelity-contract.

Quantum Networking

This is where the frontier most clearly exceeds a classical limit — but only by paying a new physical resource, and only where the detail is honored.

Entanglement-assisted capacity is a mathematically distinct, larger capacity than the unassisted Holevo capacity (Bennett-Shor-Smolin-Thapliyal), and the assisted/unassisted ratio grows without bound as the channel gets noisier. It has been demonstrated: +16.3% over a lossy, noisy bosonic channel at equal transmit power (Hao et al., PRL 126, 250501, 2021). The 'break' is honest precisely because entanglement is counted as a pre-shared resource, and the advantage peaks where absolute rates are lowest. Superadditivity / superactivation goes deeper: two zero-capacity channels can jointly carry information (Smith-Yard, Science 2008; Hastings 2009 for Holevo non-additivity) — flatly impossible for classical Shannon channels. High-dimensional superdense coding beats the C=2 qubit ceiling (2.09 with ququarts, Science Advances 2018).

On key distribution, TF-QKD beats the PLOB repeaterless bound (Pirandola et al., Nat. Commun. 8, 15043, 2017) by moving the measurement to an untrusted midpoint, turning linear loss scaling into square-root — demonstrated over 1002 km fiber (Liu et al., PRL 130, 210801, 2023). Security changes basis from computational hardness to physical law (no-cloning, Bell violation), with DIQKD certifying integrity from a loophole-free Bell test alone (Nadlinger et al., Nature 607, 682, 2022).

Verdict. Entanglement-as-resource and superadditivity are the durable, genuinely-beyond-classical primitives. But TF-QKD 'beating PLOB' reframes the channel (an untrusted relay), does not defeat loss physics, and the 1002 km rate (~1e-11/pulse) needs SNSPDs at ~0.02 Hz dark counts — very exotic hardware. True memory-based repeaters have not convincingly beaten the repeaterless bound end-to-end over long deployed fiber; metropolitan heralded entanglement (Delft, Harvard, Nature 2024) is a milestone, not a network. 'The quantum internet is here' is hype — the field sits at QIA-roadmap stages 2-3.

Certified primitives: EntanglementAssistedChannel, CapacityActivation, UntrustedMidpointRelay, HeraldedEntanglementLink, InfoTheoreticKeyAgreement, DeviceIndependentCertification.

Channel Engineering

None of this bends C = B*log2(1+SNR) — it engineers the channel the bound is stated relative to, attacking the three quantities the bound is defined over.

(1) SNR. Reconfigurable intelligent surfaces (RIS) tune a passive metasurface so reflected paths add coherently: field trials on a commercial 5G network report 21.5 dB channel-power and 23.8 dB SNR gains (FITEE 2024), and 26 dB received-power gain through a 30 cm concrete wall (arXiv:2308.03263). Coherent arrays give ~10*log10(N) dB. You move up the log curve; you do not bend it.

(2) Spatial DoF. MIMO capacity is a sum of per-eigenmode Shannon terms, so adding orthogonal modes adds terms. Holographic/continuous apertures supply more modes per square meter (IEEE JSAC 2024), and — the genuine intuition-breaker — near-field beamfocusing gives even pure-LoS channels rank > 1 by exploiting spherical wavefronts, so an extremely large aperture multiplexes users by distance as well as angle (Bjornson et al., arXiv:2209.03082). Each stream still obeys Shannon.

(3) H itself. RIS and stacked intelligent metasurfaces (SIM) reconfigure propagation; SIM even executes precoding/DoA math in the wave domain at light speed (arXiv:2411.19687; IEEE TWC 2025). ISAC is the odd one out: it raises effective spectral efficiency by extracting a second function (sensing) from spectrum that would otherwise be blanked (Keysight/MediaTek pre-6G demo, Nov 2025).

Verdict. Zero theorem violations. RIS SNR gains, massive-MIMO array gain, and cell-free macro-diversity (~5x 95%-likely per-user SE, Ngo et al., IEEE TWC 2017) are real and deployed/field-validated — approaches to the bound from a better operating point. Near-field LoS rank > 1 is a real reframing, not a violation. Watch two overreaches: 'infinite modes from a continuous aperture' (DoF is capped by aperture-in-wavelengths and bandwidth) and 'RIS in your phone now' (excluded from 3GPP Rel-18/19 normative work; a 6G study item ~2025-2027).

Certified primitives: ReflectiveRelay, CoherentCombine, SpatialMultiplex, NearFieldFocus, WaveDomainCompute, DualFunctionChannel, MacroDiversity.

Spatial Multiplexing

A single-mode fiber core saturates near ~100 Tb/s — the nonlinear Shannon limit set by the power / ASE-noise / Kerr-nonlinearity tradeoff (single-mode C-band records like NICT's 402 Tb/s across O-U bands, June 2024, show how far one spatial channel goes before you must add channels). Space-division multiplexing does not raise that limit; it multiplies channels: pack N cores and/or M modes so aggregate capacity scales ~N*M x per-channel capacity, each channel still obeying the same bound.

The records are real and clearly labeled as parallelism: 22.9 Pb/s in a 38-core / 3-mode fiber with S+C+L WDM (NICT, ECOC 2023) — but lab-grade, on non-standard fat-cladding fiber with offline DSP. What actually deploys is the low-count, standard-125um, weakly-coupled branch: 1.02 Pb/s over 1808 km in 19-core standard-cladding fiber (NICT/Sumitomo, May 2025), the first commercial 2-core submarine MCF (TPU cable, RFS 2025-2026), and a deployed 7-core submarine link with a multi-core amplifier (410.5 Tb/s over 140 km, Nat. Commun. Eng. 2026). Mode-division multiplexing works but is DSP-limited: MIMO tap count scales with differential mode delay.

The OAM verdict matters. Orbital-angular-momentum multiplexing is not a new physical dimension — it is the azimuthal subspace of the full spatial mode set, and a complete-basis MDM or plain LOS-MIMO matches or beats it (Zhao et al., Sci. Rep. srep27674, 2016). In free space, turbulence wrecks mode orthogonality. In ring-core fiber, OAM has a legitimate but modest merit — lower inter-mode crosstalk shrinks the required MIMO to small fixed blocks (4x4) — a decoding-cost win, never a capacity win.

Verdict. MCF real and deploying; MDM real but DSP-limited; OAM a useful basis choice oversold as a new dimension. SDM approaches/multiplies the bound; it never beats it, and its promoters do not claim otherwise.

Certified primitives: SpatialFanout, WeaklyCoupledPartition, CrosstalkEqualizer, OrthogonalBasisSelector.

Frontier Codes

At short blocklength the enemy is not capacity C but the finite-blocklength penalty: R*(n,e) ~ C - sqrt(V/n)*Q^-1(e) (Polyanskiy-Poor-Verdu), bracketed by the meta-converse (upper) and RCU/DT achievability (lower) bounds down to n~100. At n=128 that gap is a real 1-2 dB, and it — not C — is what short-block work fights. Three honest categories:

(a) Approach the bound, proven asymptotically. SPARCs with AMP/VAMP + spatial coupling and spinal codes are proven to achieve capacity; polar codes (Arikan 2009) are the original capacity-achieving construction, standardized in 5G NR control channels. Asymptotic results, not short-block wins.

(b) Approach the finite-blocklength optimum. The cleanest real result: PAC codes (Arikan 2019) at (128,64) under list/sequential decoding essentially meet the dispersion/RCU bound — near-optimal, not superior to any bound (arXiv:2011.03177).

(c) Beat classical codes, not any limit. ML-designed codes — KO (ICML 2021), DeepPolar+ (arXiv:2506.10166, 2025), ProductAE (~0.7-1.8 dB gains), TurboAE (NeurIPS 2019) — genuinely and reproducibly beat the specific hand-designed codes (RM, polar, turbo) at matched rate/length. Real engineering advance; closing slack to the FBL bound, not surpassing information theory.

The one legitimate 'looks like beating the bound' is feedback / rateless: spinal codes and Deepcode change the resource model the no-feedback FBL bounds assume away, so they operate against a different, higher curve — a bent question, not a broken theorem.

Verdict. Shannon capacity is a hard ceiling and the meta-converse a proven upper bound; nothing here exceeds either. The FiniteBudgetOptimalityGauge is the anti-hype instrument: always score against the meta-converse/RCU bound at the actual (n, error target), and check whether a 'beats the limit' claim's baseline is a classical code (fair) or an information-theoretic bound (then it is only approaching it).

Certified primitives: RatelessChannel, LearnedNonlinearEncoder, SpatialCoupling, SnrMatchedRedundancy, CurriculumDecoderTraining, FiniteBudgetOptimalityGauge.

The Lambek / Linear-Logic Foundation

This is the part of Frontier that is not about beating a communication bound at all — it is what lets every other certified primitive compose safely. In ordinary engineering, snapping two components together carries no interoperability guarantee; deadlocks and protocol mismatches surface at runtime. Frontier changes the channel of guarantee from testing to static proof, via exact correspondences (not analogies):

  • Curry-Howard-Lambek: propositions = types = objects of a cartesian closed category; proofs = programs = morphisms. Lambek proved this is an equivalence of categories (functorial, arXiv:1612.02816).
  • Propositions as sessions (Caires-Pfenning, CONCUR 2010; Wadler CP/GV, ICFP 2012): a channel is a linear-logic proposition, a process is a proof, and composing two processes over a channel is the CUT rule.
  • Substructural resources: Lambek's 1958 calculus drops weakening and contraction, so a channel is used exactly once — matching real messages that cannot be silently copied or dropped. Residuation (tensor left-adjoint to two implications) is exactly the adjunction tensor ⊣ internal-hom.
  • Cut-elimination = deadlock-freedom: Gentzen's theorem, ported to linear session types, makes progress a corollary of typing. Machine-checked for deadlock AND leak freedom in LinearActris (POPL 2024).
  • Coherence generalizes duality to n parties (Carbone-Montesi-Schurmann-Yoshida, CONCUR 2015), so multiparty protocols become coherence proofs synthesizing their own mediator.

Protocol refinements become functors (F(id)=id, F(g.f)=F(g).F(f) are theorems), string diagrams give sound-and-complete graphical specs (Mac Lane coherence; Joyal-Street), and structured cospans build large protocols from glued open subsystems (Baez-Courser; Fong-Spivak).

Verdict. These reframe rather than defeat — they are exact correspondences, so no theorem is broken and no impossibility beaten. The genuine win is that deadlock-freedom is a proved corollary, not a slogan. Three caveats kept explicit: the guarantee is bought by restriction (pure CP/DILL forbids cyclic topologies; cycles need Priority CP at extra cost) and does not evade the halting problem or general concurrency undecidability; functoriality holds only when the functor laws are actually discharged; and coherence-checking has real static cost. 'Category theory makes it automatically compositional' is the hype to avoid — the rigor is earned, not free.

Certified primitives: LinearChannel, Cut, CoherenceCheck, FunctorialAdapter, StringDiagram, OpenSystemCospan.

The Privacy Frontier

The anonymity trilemma (Das-Meiser-Mohammadi-Kate, 2018) is a proven impossibility and still stands: strong sender/recipient unlinkability against a global passive adversary cannot have both low bandwidth overhead and low latency. Every honest move here changes the terms the trilemma is stated over, not the inequality.

Change the question — private retrieval. PIR moves privacy from the network channel to computation: hide what you fetch rather than who you are. SimplePIR/DoublePIR hit ~10 GB/s per core (Henzinger et al., USENIX Sec 2023), YPIR removes the client hint (Menon-Wu, USENIX Sec 2024), and Piano achieves sublinear online time from PRFs alone (Zhou-Park, IEEE S&P 2024). This is a legitimately different problem (retrieval-privacy vs communication-anonymity), not a trilemma defeat.

Change the resource — preprocessing. The linear-server bound (Beimel-Ishai-Malkin, CRYPTO 2000) and ORAM's Ω(log N) bound (Larsen-Nielsen) hold only for stateless clients / unmodified DBs / no offline phase. DEPIR, Piano, and YPIR add a one-time preprocessing phase, converting per-query cost from linear to sublinear — beating the number, not the theorem, by leaving the theorem's model.

Change the question — coordination. User coordination provably beats the original 2018 bound (Das et al., PoPETs 2020) — but the same authors then prove a stronger impossibility that re-binds all known ACNs. A model change, not an escape.

Change the channel of identity. BBS anonymous credentials (IETF CFRG draft; ETSI TR 119 476) prove a predicate over attributes with full cross-use unlinkability, so no correlatable identifier is ever emitted. DPF-based systems (Express, USENIX Sec 2021; Spectrum, NSDI 2022) give cryptographic sender-unlinkability with sublinear write cost.

Verdict. Still a hard theorem. Real bends-by-reframing (PIR, coordination, credentials), one approach-the-bound that hits a physical ceiling (SimplePIR ~10 GB/s = memory bandwidth, not a crypto limit). Hype to flag: 'unconditional/absolute' anonymity around deployed mixnets (Nym/Loopix) overstates guarantees that are parameter- and traffic-dependent and degrade against active/long-running adversaries; DP systems (Vuvuzela) leak a bounded amount of metadata that accumulates against the budget over rounds — 'metadata-private' is true only within that budget, not forever. The cover-channel primitive therefore exposes the anonymity/latency/bandwidth budget for the caller to spend knowingly.

Certified primitives: private-fetch, oblivious-store, cover-channel, dpf-write, anon-credential, coordination-round.

Validation

Performance targets

The benchmarks Cognis Weave is engineered to hit — and how each capability is measured.

  1. Semantic-lossless for AI traffic: for a fixed downstream task (e.g. ImageNet top-1 or a RAG QA benchmark), a co-trained task-distortion-transmit codec will match the task metric of a bit-exact-then-infer baseline while using at least 3x fewer channel uses at a target SNR, AND will fail a bit-exactness audit (nonzero byte error) — proving it is semantic-lossless, not lossless. Falsified if it cannot beat 3x, or if it silently passes bit-exact audits (meaning it wasn't actually operating below the entropy floor).
  2. No-cliff over a real fading channel: on a hardware-in-the-loop Rayleigh/AWGN testbed, graceful-degrade-link (DeepJSCC-f) will show monotonic, continuous PSNR-vs-SNR with no >3 dB PSNR drop across any 1 dB SNR step from +10 to -5 dB, while a matched VTM+LDPC separation stack exhibits a cliff (>10 dB PSNR drop) below its design SNR. Falsified if DeepJSCC also cliffs, or if the separation baseline degrades just as gracefully.
  3. Entanglement-assisted advantage grows with noise: on a bosonic-channel testbed at equal transmit power, the measured entanglement-assisted / unassisted classical-rate ratio will strictly increase as channel noise increases (low-brightness regime), reproducing the >1.16x point and extrapolating higher. Falsified if the ratio is flat or shrinks with added noise, which would mean the demonstrated gain was not the BSST resource advantage.
  4. TF-QKD square-root scaling: a midpoint-relay QKD link will show secret-key-rate scaling as sqrt(eta) (slope ~1/2 on a log-rate vs log-transmittance plot) and exceed the PLOB linear bound at long distance, while a genuine point-to-point link on the same fiber shows slope ~1 and stays under PLOB. Falsified if the relay link also scales linearly, proving no channel-topology benefit.
  5. Near-field LoS rank > 1: with an extremely large aperture array and two receivers at the same angle but different ranges in the radiative near field, a beamfocusing precoder will deliver two independent streams with per-stream SINR high enough for >1.5x the single-stream capacity, whereas a far-field angle-only precoder collapses to rank 1 (streams uncancellable). Falsified if same-angle users cannot be separated by range.
  6. PAC codes meet the FBL bound: an independent implementation of a (128,64) PAC code under sequential/list decoding on AWGN will land within 0.25 dB of the RCU/dispersion finite-blocklength bound at BLER 1e-3, and no code at that (n,rate,error) will beat the meta-converse. Falsified if PAC misses the RCU bound by >0.5 dB (then it isn't near-optimal), or if any code claims to cross the meta-converse (then a measurement/definition error exists).
  7. Learned codes beat classical, not bounds: at matched (n, rate), a trained LearnedNonlinearEncoder (KO/DeepPolar+/ProductAE) will beat the corresponding RM/polar/turbo code by >=0.5 dB at BLER 1e-4 on AWGN, while the FiniteBudgetOptimalityGauge shows it remains strictly below the meta-converse. Falsified if it fails to beat the classical baseline, or if it appears to exceed the meta-converse (indicating a rigged baseline or a metric that hides discarded information).
  8. Deadlock-freedom is a typing corollary: a corpus of Cognis Weave compositions typed with LinearChannel/Cut will have zero runtime deadlocks under exhaustive/randomized scheduling for every configuration that typechecks in the acyclic linear fragment, and every deadlock we can construct will require a cyclic topology that fails to typecheck (or needs Priority CP annotations). Falsified if a typechecking acyclic configuration deadlocks at runtime, which would break the cut-elimination adequacy claim.
  9. PIR at memory bandwidth, not beyond: SimplePIR/YPIR throughput per core will track the machine's measured memory bandwidth within ~15% and will NOT exceed it, confirming the ~10 GB/s figure is a physical (not cryptographic) ceiling; preprocessing PIR (Piano) will show sublinear per-query server work only after a near-linear one-time preprocessing pass. Falsified if throughput exceeds memory bandwidth (measurement error) or if 'sublinear' query cost appears with no preprocessing phase (violating Beimel-Ishai-Malkin, so a modeling error).
  10. Coordination beats the base trilemma but hits the stronger wall: an ACN with synchronized coordination-rounds will achieve a lower bandwidth x latency product at fixed anonymity than the 2018 trilemma permits for uncoordinated protocols, yet will still fail to beat the PoPETs-2020 strengthened bound. Falsified if coordination cannot beat the 2018 bound (then the reframing is empty), or if it beats the stronger 2020 bound (then a proven impossibility is wrong — extraordinary, demanding a re-derivation).
  11. Optical RAM stays impossible: any slow-light / delay-line buffer we build will obey the delay-bandwidth-product ceiling (stored-bit count bounded, one-bit floor ~ one wavelength) and will fail to provide random access to a variable-length packet, while an optical-circuit-switch fabric (CircuitReserve) delivers ns switch-traversal with zero buffering. Falsified if a demonstrated optical buffer holds a variable-length packet with true random access beyond the DBP limit — which would overturn Tucker et al. and is not expected.
  12. Analog-linear offload energy budget: an AnalogLinearOffload photonic MVM will beat a current GPU/TPU on energy-per-MAC ONLY when DAC/ADC, laser, and electronic-control overhead are included in the accounting, at a specified batch size and precision; the advantage will shrink or vanish at low batch or high precision. Falsified if the claimed advantage disappears entirely once full-system overhead is counted (confirming the 'TOPS/W without overhead' hype), or, positively, confirmed if it survives an end-to-end wall-plug measurement.