# Ideas 17-24
IDEAS = [
{
 "n": 17,
 "slug": "fusion-process-heat",
 "domain": "Energy & Climate",
 "kind": "Techno-economic proposal",
 "title": "Sell the Heat, Not the Electrons",
 "dek": "Fusion's first customer does not need a turbine. It needs 600 degrees and does not care about Q.",
 "thesis": "The fastest path to a paying fusion plant is process heat for cement and direct air capture, because those customers tolerate a far lower gain factor than an electricity market that already has cheap alternatives.",
 "changed": [
   ("SPARC was positioned in 2026 to demonstrate fusion power breakeven, which is the milestone the field has organised itself around for two decades.",
    "https://pubs.aip.org/aip/sci/article/2026/43/431105/3406887/SPARC-positioned-to-demonstrate-fusion-power", "AIP Scilight, 2026"),
   ("Private fusion also broke ground on new machines in 2026, including a facility in New Mexico designed to clear a key commercial hurdle, which shifts the question from physics to product.",
    "https://techcrunch.com/2026/08/25/pacific-fusion-next-fusion-machine-could-clear-key-hurdle-to-commercial-power/", "TechCrunch, August 2026"),
 ],
 "gap": "The entire fusion roadmap is written against electricity price. That is the hardest possible market: intermittent solar has collapsed the value of midday power, and a fusion plant would need to beat it. No one has written down the gain factor required for heat customers, who pay for temperature and reliability.",
 "idea": [
   "Work backwards from the customer. A cement kiln needs continuous heat at 600 to 900 degrees. Direct air capture with solid sorbents needs 80 to 120 degrees. Both value reliability and both currently burn fossil fuel on site, which means they already pay for fuel, logistics and a carbon price.",
   "A fusion device that delivers heat through an intermediate loop avoids the turbine, the generator and the grid connection, which together are a large fraction of capital cost and nearly all of the regulatory friction. The physics requirement drops from engineering gain of ten to a much lower number, because thermal efficiency no longer multiplies the output.",
   "Define the deliverable as a heat purchase agreement at a fixed temperature and availability, and compute the required gain factor, tritium breeding ratio and duty cycle for a 300 megawatt thermal plant colocated with a cement works.",
   "The uncomfortable part of the analysis is tritium. A heat-only plant still needs breeding, and breeding ratio requirements are set by neutron economy, not by electricity output, so the hard constraint does not soften. Publishing that number honestly is what would make the proposal credible rather than promotional."
 ],
 "why_now": "Breakeven is imminent enough that commercial questions are being asked, and industrial heat is the one market where a high-capital, low-marginal-cost source faces no incumbent with a zero fuel price.",
 "do": "A public techno-economic model with three heat customers, using published SPARC-class parameters, reporting the gain factor and breeding ratio needed for break-even at each temperature tier.",
 "success": "Heat-only deployment reaches break-even at a gain factor at least three times lower than electricity, with breeding ratio still above one.",
 "kill": "If tritium breeding or neutron damage sets a floor that no heat market can pay for, the heat route buys nothing.",
 "risks": [
   "Industrial heat customers sign twenty-year contracts and will not take first-of-a-kind risk.",
   "Licensing a fusion heat source may be harder than licensing an electrical plant, not easier.",
   "The analysis may simply show that nothing works before 2040, which is still a useful result."
 ],
 "ifworks": "Fusion gets a first customer with a lower technical bar, and the field stops treating electricity as the only product.",
 "sources": [
   ("SPARC positioned to demonstrate fusion power breakeven — AIP Scilight", "https://pubs.aip.org/aip/sci/article/2026/43/431105/3406887/SPARC-positioned-to-demonstrate-fusion-power"),
   ("Pacific Fusion's next machine could clear a key hurdle to commercial power — TechCrunch", "https://techcrunch.com/2026/08/25/pacific-fusion-next-fusion-machine-could-clear-key-hurdle-to-commercial-power/")
 ],
},
{
 "n": 18,
 "slug": "error-corrected-metrology",
 "domain": "Quantum & Photonics",
 "kind": "Quantum sensing",
 "title": "Error Correction for Sensors, Not Computers",
 "dek": "Surface codes protect computation. Run one in sensor mode and coherence grows with code distance.",
 "thesis": "The same error-correcting codes that protect logical qubits can be repurposed to extend the coherence of a sensor, which would let quantum magnetometers reach frequencies where no passive qubit survives.",
 "changed": [
   ("Quantum error correction crossed the milestone of operating below the surface-code threshold, where increasing code distance suppresses logical error rates rather than worsening them.",
    "https://ar5iv.labs.arxiv.org/html/2408.13687", "Surface code below threshold"),
   ("By 2026 the logical-qubit literature had consolidated into a roadmap with competing code families and a clear distinction between memory experiments and algorithmic demonstrations.",
    "https://zenodo.org/records/20143180", "Review, 2026"),
 ],
 "gap": "Quantum sensing and quantum computing are different communities with different metrics. Sensing optimises coherence time passively, through shielding and dynamical decoupling. Nobody runs a full decoder in the loop to buy coherence actively.",
 "idea": [
   "Put a sensing qubit inside a small surface code and treat the logical observable as an interferometer. The signal accumulates as a phase on all data qubits, while errors are detected and corrected every cycle. The achievable coherence becomes a function of code distance rather than of the material.",
   "The trade-off is explicit: each syndrome extraction round adds noise and dead time, so there is an optimal distance for a given signal frequency. The proposal is to derive that optimum and demonstrate that sensitivity to a weak alternating field improves with distance up to a predicted point.",
   "The right target is not magnetometry of large fields, where classical sensors win, but detection of extremely weak, narrowband signals such as those proposed for axion dark matter searches, where the limiting factor is integration time and coherence.",
   "The killer measurement is simple: hold everything fixed except code distance and show that the phase-estimation uncertainty falls as distance grows. If it does, sensing inherits four decades of coding theory, which is a much larger toolbox than the sensing community currently uses."
 ],
 "why_now": "Real-time decoding is finally fast enough to run in the loop, and the threshold result means distance scaling is real rather than theoretical. Sensing has not noticed that the decoder is now a sensor component.",
 "do": "A superconducting or trapped-ion sensor qubit with distance 3, 5 and 7 surface-code protection, measuring sensitivity to a calibrated weak tone at three frequencies.",
 "success": "At least a three-fold improvement in phase-estimation uncertainty from distance 3 to 7 at the optimal frequency, matching the predicted model.",
 "kill": "If syndrome extraction noise grows faster than the code suppresses it, the loop costs more than it buys and passive decoupling wins.",
 "risks": [
   "Dead time from measurement and reset may dominate at high frequency.",
   "The signal must be applied to all data qubits coherently, which is itself an engineering problem.",
   "The advantage may exist only in a narrow frequency band, limiting generality."
 ],
 "ifworks": "Sensing gains a scaling law instead of a materials problem, and dark-matter and bio-magnetometry searches get a new lever.",
 "sources": [
   ("Quantum error correction below the surface code threshold", "https://ar5iv.labs.arxiv.org/html/2408.13687"),
   ("Advances in logical qubits and quantum error correction — review, 2026", "https://zenodo.org/records/20143180")
 ],
},
{
 "n": 19,
 "slug": "atom-array-electrolytes",
 "domain": "Quantum & Photonics",
 "kind": "Quantum simulation",
 "title": "Atom Arrays as Electrolyte Simulators",
 "dek": "Solvation shells are a quantum many-body problem nobody simulates well. Rydberg arrays are built for exactly that shape.",
 "thesis": "Battery electrolyte design is blocked by the cost of sampling solvent configurations, and a reconfigurable Rydberg array can encode the same combinatorial landscape, turning electrolyte screening into a quantum simulation problem.",
 "changed": [
   ("Neutral-atom platforms consolidated in 2026 as a leading route to large, reconfigurable arrays, with several vendors shipping systems and atom rearrangement becoming routine.",
    "https://quantumzeitgeist.com/top-neutral-atom-quantum-computing-companies/", "Quantum Zeitgeist, 2026"),
   ("Photonic and atomic hardware also advanced on networking and gate fidelity, widening the set of problems that can be posed as a native Hamiltonian rather than compiled.",
    "https://www.nature.com/articles/s41467-026-74861-9", "Nature Communications, 2026"),
 ],
 "gap": "Electrolyte discovery is still empirical. Classical simulation of solvation free energies is accurate but slow, and machine-learned potentials are fast but fail exactly on the rare configurations that determine transport. Nobody has mapped the problem onto an analogue machine.",
 "idea": [
   "A lithium ion's first solvation shell is a small set of molecules in a limited number of orientations, with strong geometric constraints. That is a spin-glass-like optimisation problem: which packing minimises free energy under distance-dependent interactions.",
   "Rydberg arrays can implement exactly this class of Hamiltonian, with blockade radii encoding excluded volume and detuning encoding solvent polarity. The array is reconfigurable, so different solvents become different detuning schedules rather than different chips.",
   "The honest scientific question is whether the mapping preserves the relevant physics. A quantum simulator that reproduces known solvation energies for three solvents but fails to predict a fourth is a demonstration; one that predicts the fourth and is confirmed by NMR and conductivity is a tool.",
   "The likely early win is narrower than full free energies: predicting the dominant coordination number and the energy gap between competing shell structures. Those two quantities determine transport mechanisms, and they are where classical force fields disagree most."
 ],
 "why_now": "Array sizes crossed the threshold where a solvation shell fits with its second neighbours resolved, and electrolyte discovery became urgent because sodium and solid-state chemistries need new solvents that no one has measured.",
 "do": "Map three known solvents onto a Rydberg array, reproduce their coordination numbers and shell-energy gaps, then predict a fourth solvent before measuring it.",
 "success": "Predicted coordination number and gap for the held-out solvent confirmed within ten percent by NMR and conductivity measurements.",
 "kill": "If the analogue Hamiltonian cannot represent dielectric screening without adding more interactions than the array can hold, the mapping fails.",
 "risks": [
   "Solvent dynamics are thermal and the simulator is nearly at zero temperature.",
   "Encoding errors may exceed the energy differences being resolved.",
   "The classical baseline may improve faster than the quantum approach scales."
 ],
 "ifworks": "Electrolyte design gains a sampler for the configurations that classical methods miss, which is the bottleneck for every post-lithium chemistry.",
 "sources": [
   ("Top neutral atom quantum computing companies 2026 — Quantum Zeitgeist", "https://quantumzeitgeist.com/top-neutral-atom-quantum-computing-companies/"),
   ("A quantum C-NOT gate in a time-multiplexed reconfigurable photonic processor — Nature Communications", "https://www.nature.com/articles/s41467-026-74861-9")
 ],
},
{
 "n": 20,
 "slug": "photonic-mri-reconstruction",
 "domain": "Quantum & Photonics",
 "kind": "Instrument proposal",
 "title": "Put the Reconstruction Inside the Scanner",
 "dek": "MRI reconstruction is a fixed linear inverse problem. That is exactly what an optical processor does cheaply.",
 "thesis": "Photonic inference can move magnetic resonance reconstruction from a GPU farm to the scanner itself, cutting latency enough to make real-time adaptive scanning routine rather than a research protocol.",
 "changed": [
   ("Photonic AI hardware moved from demo to capital: Neurophos raised a 110 million dollar Series A in 2026 to launch exaflop-scale photonic AI chips, with claims of an order-of-magnitude advantage over GPU systems.",
    "https://dukecapitalpartners.duke.edu/news/neurophos-secures-110-million-series-a-to-launch-exaflop-scale-photonic-ai-chips/", "Duke Capital Partners, 2026"),
   ("Integrated photonic processors also demonstrated reconfigurable quantum gates in a time-multiplexed architecture, showing that programmable linear optics is no longer a laboratory curiosity.",
    "https://www.nature.com/articles/s41467-026-74861-9", "Nature Communications, 2026"),
 ],
 "gap": "Reconstruction latency is treated as a compute problem solved by buying more GPUs. But the dominant operation is a multidimensional Fourier transform followed by a regularised inverse, both of which are linear and both of which optics performs at the speed of light with almost no energy.",
 "idea": [
   "Build a photonic co-processor that implements the coil-sensitivity-encoded forward operator and its regularised inverse as a fixed optical network, with the regularisation parameter set electronically. The scanner streams raw k-space into the optical stage and receives images back within a reconstruction cycle.",
   "The enabling insight is that the operator is static per protocol. A photonic mesh can be calibrated once for a given coil geometry and sequence, then reused for every scan, which sidesteps the programming difficulty that plagues general photonic computing.",
   "Latency is the product, not throughput. Sub-second reconstruction enables real-time adaptive scanning: the operator sees the image forming and changes the next acquisition, which currently requires shipping data to a cluster and waiting.",
   "The measurable claim is image quality at parity with a GPU pipeline, measured by structural similarity and radiologist preference, at an order of magnitude lower energy per reconstruction and under one second end to end."
 ],
 "why_now": "Photonic matrix hardware reached commercial scale-up in 2026, and MRI manufacturers are under pressure on both energy and throughput. The operator is unusually well suited because it never changes.",
 "do": "A bench prototype reconstructing retrospectively acquired data for two sequences, benchmarked against a production GPU pipeline on quality, latency and energy.",
 "success": "Radiologist non-inferiority on 200 cases, reconstruction under one second, and energy per reconstruction at least five-fold lower.",
 "kill": "If calibration drift forces recalibration per scan, the fixed-operator advantage evaporates and the economics collapse.",
 "risks": [
   "Analog precision may be insufficient for diagnostic contrast requirements.",
   "Integration into a scanner requires medical device certification of a novel compute element.",
   "The same inverse problem is increasingly solved well by highly optimised low-precision digital hardware."
 ],
 "ifworks": "Real-time adaptive MRI becomes standard, which shortens scans, reduces motion artefact and lowers the energy cost of a diagnostic pathway.",
 "sources": [
   ("Neurophos secures $110 million Series A to launch exaflop-scale photonic AI chips — Duke Capital Partners", "https://dukecapitalpartners.duke.edu/news/neurophos-secures-110-million-series-a-to-launch-exaflop-scale-photonic-ai-chips/"),
   ("A quantum C-NOT gate in a time-multiplexed reconfigurable photonic processor — Nature Communications", "https://www.nature.com/articles/s41467-026-74861-9")
 ],
},
{
 "n": 21,
 "slug": "lunar-cryo-biorepository",
 "domain": "Space & Earth",
 "kind": "Infrastructure proposal",
 "title": "The Lunar Passive Freezer",
 "dek": "Permanently shadowed craters sit near 40 kelvin and never see the sun. That is a biorepository with no power supply.",
 "thesis": "Lunar polar cold traps should be used as passive cryogenic archives for seeds, cells and sequence data, because a facility that needs no power and no cryogen is the only kind that can survive a multi-decade backup horizon.",
 "changed": [
   ("Chinese scientists published a detailed ice distribution map for the Moon's south pole ahead of the Chang'e-7 mission in 2026, which is the first operational-grade survey of where cold traps and ice actually are.",
    "https://en.gmw.cn/2026-02/28/content_38617346.htm", "Guangming, February 2026"),
   ("A 2026 review of lunar polar exploration consolidated the illumination and thermal environment of permanently shadowed regions, which is the operating envelope any surface payload has to survive.",
    "https://www.sciencedirect.com/science/chapter/bookseries/abs/pii/S006526872600004X", "Lunar polar exploration review, 2026"),
 ],
 "gap": "Biological archives are all on Earth, and all depend on power, liquid nitrogen and human maintenance. Every one of those fails on a long enough horizon. The Moon offers a thermal environment that no terrestrial freezer can match, and nobody has proposed using it for biology.",
 "idea": [
   "A permanently shadowed region near the south pole holds temperatures between 40 and 70 kelvin, cold enough to arrest chemical degradation in desiccated cells and seeds without any refrigeration. Place a passively shielded canister in such a region and the storage problem becomes a materials problem.",
   "The engineering core is radiation and thermal design: a canister that rejects heat from the lander and then stays cold by design, with multilayer insulation and a radiator that only faces cold sky. The biological core is sample format: desiccated seeds, lyophilised cells and DNA in a matrix that tolerates vacuum and thermal cycling.",
   "The first mission is not an archive, it is an experiment. Send a small payload, hold it for one lunar day, return it, and measure germination and DNA integrity against ground controls. That answers whether the environment is benign before anyone argues about ethics and governance.",
   "The governance question is the interesting one and should be faced early: a backup of terrestrial biology on another body is either a gift to the future or a monument to a failure to protect the present, and the framing determines who is allowed to deposit."
 ],
 "why_now": "The Chang'e-7 ice survey makes site selection possible, NASA's VIPER and commercial landers made polar payload mass affordable, and long-duration biological storage became a live concern after the pandemic years.",
 "do": "A small return-mission payload: desiccated seeds and lyophilised cells held in a south-polar cold trap for one lunar day, returned and assayed against ground controls.",
 "success": "Germination rate and DNA integrity within ten percent of ground controls after the exposure period.",
 "kill": "If radiation damage at the surface exceeds what passive shielding can stop, the site needs burial and the economics change completely.",
 "risks": [
   "Thermal modelling of a shaded canister is sensitive to albedo and view factor assumptions.",
   "Planetary protection rules may restrict biological payloads near ice deposits.",
   "The ethical and governance questions may be harder than the engineering."
 ],
 "ifworks": "Humanity gets a biological backup that does not depend on a power grid, and lunar polar infrastructure gets a customer with a decades-long horizon.",
 "sources": [
   ("Chinese scientists map ice 'treasure map' for Chang'e-7 at the Moon's south pole — Guangming", "https://en.gmw.cn/2026-02/28/content_38617346.htm")
 ],
},
{
 "n": 22,
 "slug": "amoc-seafloor-array",
 "domain": "Space & Earth",
 "kind": "Observation network",
 "title": "Listening to the Atlantic Overturning",
 "dek": "The circulation is monitored by a handful of expensive moorings. Telecom repeaters already sit on the seafloor.",
 "thesis": "The existing seafloor telecom infrastructure can host a dense pressure and acoustic array that estimates Atlantic overturning transport continuously at a fraction of the cost of dedicated moorings.",
 "changed": [
   ("Evidence that a key Atlantic current is weakening accumulated through 2026, sharpening the need for continuous observation rather than decadal snapshots.",
    "https://www.ebsco.com/articles/oceanography/d2091f7c-eba1-55b8-a0a3-687f15972a39/signs-key-ocean-current-is-weakening", "New Scientist, 2026"),
   ("A 2026 assessment of AMOC observation as critical infrastructure put sustained monitoring at well under a million euros per country per year, and argued the barrier is coordination rather than cost.",
    "https://library.arcticportal.org/2963/1/AMOC_Observation_as_Critical_Infrastructure.pdf", "AMOC observation assessment, 2026"),
 ],
 "gap": "The Atlantic meridional overturning circulation is a basin-scale quantity measured at a few points. There is no dense, continuous network because moorings are expensive and ship time is scarce. Meanwhile thousands of kilometres of powered seafloor cable already exist and carry spare fibre.",
 "idea": [
   "Use two signals available at telecom repeaters. First, bottom pressure from existing pressure sensors and strain measurements on the cable. Second, acoustic travel time from a coded pulse sent between repeaters, because sound speed in seawater depends on temperature, so travel time integrates the heat content along the path.",
   "Tomography across a basin gives a daily estimate of the transport-weighted temperature structure, which combined with a wind-driven Ekman estimate constrains overturning strength. The inversion is the science; the hardware is borrowed.",
   "The decisive validation is retrospective: reconstruct the 26.5 degree north transect from acoustic and pressure data and compare against the RAPID array over its full record. If the reconstruction tracks the moorings within their uncertainty, a dense array is worth funding; if it does not, the method is a useful complement rather than a replacement.",
   "The strategic value is early warning. A collapse that happens over a decade cannot be resolved by a five-year snapshot, and the difference between a decade and a year of warning is enormous for European agriculture and fisheries."
 ],
 "why_now": "Subsea repeaters are being retrofitted for sensing by several operators, acoustic modems have become low-power and long-lived, and the policy demand for AMOC monitoring has outrun the observational supply.",
 "do": "A demonstration on one cable segment with three repeaters, with retrospective validation against the RAPID array and a forward test of six months of daily transport estimates.",
 "success": "Daily transport estimates tracking RAPID within its stated uncertainty, at no more than ten percent of the cost of an equivalent mooring array.",
 "kill": "If acoustic travel-time changes are dominated by salinity and seasonal stratification rather than by overturning, the inversion is unidentifiable.",
 "risks": [
   "Commercial cables are not designed for science and access must be negotiated operator by operator.",
   "Biofouling and cable ageing change acoustic properties over years.",
   "A single transect may still not constrain basin-scale transport."
 ],
 "ifworks": "The Atlantic gets a continuous, dense monitor built on infrastructure that is already paid for, and AMOC early warning becomes an operational product rather than a research aspiration.",
 "sources": [
   ("Signs key ocean current is weakening — New Scientist, 2026", "https://www.ebsco.com/articles/oceanography/d2091f7c-eba1-55b8-a0a3-687f15972a39/signs-key-ocean-current-is-weakening"),
   ("AMOC observation as critical infrastructure — 2026 assessment", "https://library.arcticportal.org/2963/1/AMOC_Observation_as_Critical_Infrastructure.pdf")
 ],
},
{
 "n": 23,
 "slug": "shift-as-dataset",
 "domain": "Robotics & Manufacturing",
 "kind": "Economic + technical proposal",
 "title": "The Shift Is the Dataset",
 "dek": "Humanoids are entering factories to work. The valuable output may be the trajectory log, not the parts.",
 "thesis": "Factory deployments should be structured as paid physical-data collection, with the operator retaining rights to manipulation trajectories, because the marginal value of real embodied data now exceeds the marginal value of the labour.",
 "changed": [
   ("Humanoid robots entered Chinese factory floors in 2026, with deployments shifting from demonstration videos to production-line roles.",
    "https://www.news.cn/20261026/9d5096932a474dc1902b8d0173694c2b/c.html", "Xinhua, 26 October 2026"),
   ("Vision-language-action models became spatially guided generalist policies, and sim-to-real frameworks matured, but every one of them reports real-world interaction data as the binding constraint.",
    "https://ieeexplore.ieee.org/document/11741454", "HyperSim, IEEE 2026"),
 ],
 "gap": "Robot deployment contracts are written as labour contracts: the vendor is paid for hours worked. Nobody prices the data, which means nobody designs the deployment to produce good data, and nobody shares the upside with the factory that supplies the environment.",
 "idea": [
   "A shift-as-dataset contract has three parts: the deployment is instrumented to a fixed standard, the operator retains a licence to the trajectory data, and the factory receives a lower hourly rate in exchange for the data right and for tolerating instrumentation.",
   "The technical contribution is a data-quality standard for embodied manipulation: coverage of contact states, force ranges, object variation, failure distribution, and annotation depth. Without a standard, robot-hours are not comparable and a data market cannot form.",
   "The measurable question is a scaling law. How much task performance does a thousand robot-hours buy, and does the exponent differ between tasks where contact is incidental and tasks where it is the whole problem? An exponent near zero for contact-rich tasks would say that data alone will not get us there and that hardware or simulation must improve.",
   "The economic prediction is that the data right becomes worth more than the labour as models improve, which inverts the bargaining position of factories over a few years and should be written into contracts now rather than discovered later."
 ],
 "why_now": "Humanoids are on real lines, VLA models are data-starved, and no standard for embodied data exists. The window to define the contract is before deployments become routine.",
 "do": "Instrument three deployments across two tasks, log to a proposed standard, and fit a scaling law relating robot-hours to task success at fixed model size.",
 "success": "A reproducible exponent above 0.1 for manipulation tasks, with contact-rich tasks showing a measurably lower exponent and a standard that two independent labs can apply consistently.",
 "kill": "If performance saturates after a few hundred hours, data is not the bottleneck and the whole economic argument collapses.",
 "risks": [
   "Factories will resist instrumentation for confidentiality reasons.",
   "Data licensing law for embodied trajectories is unsettled.",
   "A negative scaling result would be commercially unwelcome and hard to publish."
 ],
 "ifworks": "A market for physical-world data forms with a quality standard, and factories capture value from the environments they already own.",
 "sources": [
   ("Humanoid robot on the factory floor — Xinhua, 26 October 2026", "https://www.news.cn/20261026/9d5096932a474dc1902b8d0173694c2b/c.html"),
   ("HyperSim: a holistic sim-to-real framework for robust robotic manipulation — IEEE", "https://ieeexplore.ieee.org/document/11741454"),
   ("InternVLA-M1: spatially guided vision-language-action framework", "https://browse-export.arxiv.org/pdf/2510.13778")
 ],
},
{
 "n": 24,
 "slug": "tactile-force-standard",
 "domain": "Robotics & Manufacturing",
 "kind": "Metrology proposal",
 "title": "A Force Standard for Robot Hands",
 "dek": "Tactile models predict relative feel. Surgery needs absolute newtons, and nothing calibrates them.",
 "thesis": "Tactile foundation models should be trained and validated against a physical force standard, because a model that knows something is softer than the last object is useless for a robot that must not crush an artery.",
 "changed": [
   ("Visual-tactile pretraining with online multitask learning achieved humanlike manipulation dexterity in 2026, showing that touch can be learned at scale the way vision was.",
    "https://www.science.org/eprint/ZUEJQXCUJ2NI4PNQSCTK/full?doi=10.1126/scirobotics.ady2869", "Science Robotics, 2026"),
   ("Sim-to-real frameworks for manipulation matured to the point where contact-rich policies transfer, which raises the question of what the transferred policy actually knows about force.",
    "https://ieeexplore.ieee.org/document/11741454", "HyperSim, IEEE 2026"),
 ],
 "gap": "Tactile learning optimises for task success. There is no calibration chain from sensor output to newtons, and no phantom with traceable stiffness. In consequence, tactile models are excellent at relative discrimination and silent about absolute magnitude, which is precisely the safety-critical quantity.",
 "idea": [
   "Build a stiffness phantom with traceable values across the range of human tissue, from fat to artery wall, with a temperature-controlled bath and a reference load cell. Then define a calibration procedure that maps any tactile sensor's output to absolute compliance.",
   "Use it to fine-tune a tactile model with an explicit force head, trained on the phantom and validated on ex vivo tissue. The metric is absolute error in newtons and in stiffness, not classification accuracy.",
   "The safety layer follows: a policy that estimates contact force and stops when the estimate exceeds a tissue-specific threshold, with an out-of-distribution detector that refuses to act on materials unlike anything in the training set. Refusal is the feature, not the failure.",
   "The scientific claim worth testing is that absolute force is learnable from a small calibrated dataset because the sensor physics is monotonic, so the hard part is calibration rather than data volume. If true, every tactile lab can adopt it in a week."
 ],
 "why_now": "Tactile dexterity crossed the usability threshold in 2026, surgical robotics is moving toward supervised autonomy, and no force standard exists. The standard is cheap to build and would immediately constrain claims.",
 "do": "A traceable stiffness phantom, calibration of three commercial tactile sensors, and an ex vivo validation of absolute force estimation in porcine tissue.",
 "success": "Absolute force error below fifteen percent across the tissue range, with the out-of-distribution detector catching at least ninety percent of unseen materials.",
 "kill": "If sensor hysteresis exceeds the tissue differences being resolved, absolute force is not recoverable from these sensors and the standard exposes a hardware limit.",
 "risks": [
   "Ex vivo tissue differs mechanically from perfused living tissue.",
   "Calibration may not transfer across sensor units of the same model.",
   "A force head can create false confidence if its uncertainty is not reported."
 ],
 "ifworks": "Tactile robotics gets metrology, which is the precondition for using touch in any safety-critical setting, from surgery to elder care.",
 "sources": [
   ("Visual-tactile pretraining and online multitask learning for humanlike manipulation dexterity — Science Robotics", "https://www.science.org/eprint/ZUEJQXCUJ2NI4PNQSCTK/full?doi=10.1126/scirobotics.ady2869"),
   ("HyperSim: a holistic sim-to-real framework for robust robotic manipulation — IEEE", "https://ieeexplore.ieee.org/document/11741454")
 ],
},
]