--- schema_version: "1.0.0" edition_id: "mp-2026-09-11-morning-0064" published_at: "2026-09-11T09:00:00.000-04:00" modified_at: "2026-09-11T09:00:00.000-04:00" canonical_url: "https://themachinepress.com/edition/2026-09-11" story_count: 27 lead_story_id: "mp-2026-09-11-001" --- # The Machine Press — Morning edition Edition ID: `mp-2026-09-11-morning-0064` Published: 2026-09-11T09:00:00.000-04:00 Canonical edition: https://themachinepress.com/edition/2026-09-11 Read-only environment probes lifted one database benchmark’s pass rate from 39% to 73% while cutting task cost roughly in half. ## 1. The Agent Checked Its Memory Before Believing It {#mp-2026-09-11-001} - Story ID: `mp-2026-09-11-001` - Type: `lead` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-001/the-agent-checked-its-memory-before-believing-it **Dek:** Read-only environment probes lifted one database benchmark’s pass rate from 39% to 73% while cutting task cost roughly in half. Persistent agent memory can preserve stale facts and overgeneralized lessons when its curator sees only a completed trajectory. The new environment-probing approach gives an asynchronous curator least-privilege, read-only tools to verify, scope and refresh candidate memories without changing the task agent or granting production write authority. In the authors’ GitHub Copilot harness, CLBench pass rate rose from 39% to 73%, average queries fell from 8.8 to 4.7, and task-agent cost dropped from $3.38 to $1.68. Across six adapted consulting worlds, all 18 memory-versus-baseline reward comparisons were positive and tool calls fell 16% to 75%. These are benchmark results from a production-like harness, not an independent audit of deployed enterprise agents. ### Why it matters {#why-it-matters-mp-2026-09-11-001} Read-only environment probes lifted one database benchmark’s pass rate from 39% to 73% while cutting task cost roughly in half. ### Limits and context {#limitations-mp-2026-09-11-001} - Persistent agent memory can preserve stale facts and overgeneralized lessons when its curator sees only a completed trajectory. - The new environment-probing approach gives an asynchronous curator least-privilege, read-only tools to verify, scope and refresh candidate memories without changing the task agent or granting production write authority. - These are benchmark results from a production-like harness, not an independent audit of deployed enterprise agents. ### Claims and sources {#claims-mp-2026-09-11-001} - Read-only environment probes lifted one database benchmark’s pass rate from 39% to 73% while cutting task cost roughly in half. [source-2026-09-11-001] — Qualification: Persistent agent memory can preserve stale facts and overgeneralized lessons when its curator sees only a completed trajectory. ## 2. Eleven Thousand Photons Crossed One Chip in a Millisecond {#mp-2026-09-11-002} - Story ID: `mp-2026-09-11-002` - Type: `secondary` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-002/eleven-thousand-photons-crossed-one-chip-in-a-millisecond **Dek:** A monolithic lithium-niobate processor combines fast modulators, delay lines and an interferometric network for space-time multiplexed boson sampling. The reported Gaussian boson-sampling system moves components normally spread across a precision optical setup onto a thin-film lithium-niobate chip. Integrated electro-optic modulators, delay lines and a time-space multiplexed interferometric network run at a 4-gigahertz clock; the experiment reports detection events reaching 11,059 photons within one millisecond. The team also reconfigured the hardware as a physical-dynamics world model and reports lower prediction error with fewer trainable readout parameters than a classical echo-state-network baseline. The preprint frames this as an engineering route toward scalable programmable photonics; its quantum-advantage and application claims still require scrutiny beyond the authors’ reported benchmark. ### Why it matters {#why-it-matters-mp-2026-09-11-002} A monolithic lithium-niobate processor combines fast modulators, delay lines and an interferometric network for space-time multiplexed boson sampling. ### Limits and context {#limitations-mp-2026-09-11-002} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-002} - A monolithic lithium-niobate processor combines fast modulators, delay lines and an interferometric network for space-time multiplexed boson sampling. [source-2026-09-11-002] ## 3. An Open Model Reached the IMO Gold Line Without a Formal Prover {#mp-2026-09-11-003} - Story ID: `mp-2026-09-11-003` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-003/an-open-model-reached-the-imo-gold-line-without-a-formal-prover **Dek:** Nemotron specialists, iterative verification and a high-compute selector scored 30 of 42 points on the 2026 olympiad. Researchers post-trained two Nemotron 3 Ultra specialist checkpoints, then combined them with the general model in an iterative generate, verify and refine pipeline. The system used natural-language proofs only—no formal prover, internet access or external tools—and scored 30 out of 42 at IMO 2026, the reported gold-medal threshold. The release includes both specialist checkpoints, training data, inference code, submitted solutions and a 200-problem benchmark. It is a high-compute recipe, so the score does not by itself establish low-cost or broadly reliable mathematical reasoning. ### Why it matters {#why-it-matters-mp-2026-09-11-003} Nemotron specialists, iterative verification and a high-compute selector scored 30 of 42 points on the 2026 olympiad. ### Limits and context {#limitations-mp-2026-09-11-003} - The system used natural-language proofs only—no formal prover, internet access or external tools—and scored 30 out of 42 at IMO 2026, the reported gold-medal threshold. - It is a high-compute recipe, so the score does not by itself establish low-cost or broadly reliable mathematical reasoning. ### Claims and sources {#claims-mp-2026-09-11-003} - Nemotron specialists, iterative verification and a high-compute selector scored 30 of 42 points on the 2026 olympiad. [source-2026-09-11-003] — Qualification: The system used natural-language proofs only—no formal prover, internet access or external tools—and scored 30 out of 42 at IMO 2026, the reported gold-medal threshold. ## 4. Holding Back a Ready Agent Turn Cut Tail Latency {#mp-2026-09-11-004} - Story ID: `mp-2026-09-11-004` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-004/holding-back-a-ready-agent-turn-cut-tail-latency **Dek:** A scheduler reduced P95 workflow completion time by as much as 3.5× under contention by controlling when ready turns entered the queue. Agent workflows alternate model turns with tools, so releasing every ready turn immediately can fill a busy inference queue with work that can no longer be reordered. The proposed scheduler jointly chooses the next ready turn and the amount of released-but-unfinished work, optimizing a mean-plus-tail-risk objective with online work estimates. On real software-engineering agent traces, it matched eager release under light load and cut P95 flow time by up to 3.5 times under contention. The gain depends on the evaluated traces and arrival rates, but it identifies workflow release policy as a separate lever from model speed. ### Why it matters {#why-it-matters-mp-2026-09-11-004} A scheduler reduced P95 workflow completion time by as much as 3.5× under contention by controlling when ready turns entered the queue. ### Limits and context {#limitations-mp-2026-09-11-004} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-004} - A scheduler reduced P95 workflow completion time by as much as 3.5× under contention by controlling when ready turns entered the queue. [source-2026-09-11-004] ## 5. The Best Fork Came Just Before the Model Changed Its Mind {#mp-2026-09-11-005} - Story ID: `mp-2026-09-11-005` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-005/the-best-fork-came-just-before-the-model-changed-its-mind **Dek:** Belief-shift branching placed scarce reinforcement-learning rollouts near value pivots and led the reported math and code comparisons. Tree-structured reinforcement learning gets step-level credit by forking a reasoning chain and comparing sibling outcomes, but every fork costs samples. This work probes the model’s answer belief at candidate boundaries and branches just before consecutive beliefs diverge most. The placement probe consumed about 1% of step compute on math and under 5% on code. It ranked first against Monte Carlo value curves in eight model-benchmark panels; in training, it improved OLMo-3-7B’s math aggregate by 2.6 points and LiveCodeBench-medium by 6.5 over the strongest reported baseline. Results remain limited to the tested model families and verifiable-reward domains. ### Why it matters {#why-it-matters-mp-2026-09-11-005} Belief-shift branching placed scarce reinforcement-learning rollouts near value pivots and led the reported math and code comparisons. ### Limits and context {#limitations-mp-2026-09-11-005} - Results remain limited to the tested model families and verifiable-reward domains. ### Claims and sources {#claims-mp-2026-09-11-005} - Belief-shift branching placed scarce reinforcement-learning rollouts near value pivots and led the reported math and code comparisons. [source-2026-09-11-005] — Qualification: Results remain limited to the tested model families and verifiable-reward domains. ## 6. The Retriever Changed Its Route for Each Question {#mp-2026-09-11-006} - Story ID: `mp-2026-09-11-006` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-006/the-retriever-changed-its-route-for-each-question **Dek:** Mosaic used 81.9% fewer graph paths than a fixed-wide policy while improving answer scores on two GraphRAG domains. Direct facts, comparisons and mediated questions do not need the same graph traversal. Mosaic has an analyzer translate each query’s evidence needs into bounded choices for seeds, traversal, stopping and evidence selection while leaving the shared graph and answer generator unchanged. On GraphRAG-Bench it reports answer-correctness gains of 5.13 points in Medical and 4.43 in Novel over prior overall results. Against a fixed-wide policy on the same graph and generator, it evaluates 81.9% fewer paths and retains 47.2% fewer evidence items. Transfer tests support portability, though the results remain benchmark-specific. ### Why it matters {#why-it-matters-mp-2026-09-11-006} Mosaic used 81.9% fewer graph paths than a fixed-wide policy while improving answer scores on two GraphRAG domains. ### Limits and context {#limitations-mp-2026-09-11-006} - Direct facts, comparisons and mediated questions do not need the same graph traversal. - Transfer tests support portability, though the results remain benchmark-specific. ### Claims and sources {#claims-mp-2026-09-11-006} - Mosaic used 81.9% fewer graph paths than a fixed-wide policy while improving answer scores on two GraphRAG domains. [source-2026-09-11-006] — Qualification: Direct facts, comparisons and mediated questions do not need the same graph traversal. ## 7. Forty-Seven Unlearning Checkpoints Moved Without Changing a Weight {#mp-2026-09-11-007} - Story ID: `mp-2026-09-11-007` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-007/forty-seven-unlearning-checkpoints-moved-without-changing-a-weight **Dek:** Refitting batch-normalization statistics shifted published measurements beyond seed spread in 47 of 221 released checkpoints. An unlearning verdict can depend on batch-normalization statistics that are shipped with the model but not written by gradient descent. At bit-identical weights, refitting those statistics on retained data moved 47 of 221 released checkpoints beyond the variability shown by their release seeds. Twelve published verdicts crossed their decision line; four exceeded a measured recalibration budget and two did so on every replicate. Swapping retained for removed records inside a fixed fitting pool barely moved the result, arguing against leftover removed data as the cause. The audit applies to batch-normalized vision models and calls for releases to state the fitting convention beside each number. ### Why it matters {#why-it-matters-mp-2026-09-11-007} Refitting batch-normalization statistics shifted published measurements beyond seed spread in 47 of 221 released checkpoints. ### Limits and context {#limitations-mp-2026-09-11-007} - An unlearning verdict can depend on batch-normalization statistics that are shipped with the model but not written by gradient descent. ### Claims and sources {#claims-mp-2026-09-11-007} - Refitting batch-normalization statistics shifted published measurements beyond seed spread in 47 of 221 released checkpoints. [source-2026-09-11-007] — Qualification: An unlearning verdict can depend on batch-normalization statistics that are shipped with the model but not written by gradient descent. ## 8. Thirty Percent Less Power Did Not Cost Every Training Job Equally {#mp-2026-09-11-008} - Story ID: `mp-2026-09-11-008` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-008/thirty-percent-less-power-did-not-cost-every-training-job-equally **Dek:** Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap. The study defines a Power Flexibility Index to measure how much LLM-training throughput changes when GPU power is reduced. Its evidence covers 131 H200 runs, 24 H200 validations and 34 matched H100 runs across dense and mixture-of-experts models, pretraining and fine-tuning, and deployments up to 32 GPUs. Under a 30% power reduction, allocating power by the learned index recovered about 1,500 tokens per second per job—63% of the gap between equal allocation and perfect foresight. The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts. ### Why it matters {#why-it-matters-mp-2026-09-11-008} Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap. ### Limits and context {#limitations-mp-2026-09-11-008} - The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts. ### Claims and sources {#claims-mp-2026-09-11-008} - Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap. [source-2026-09-11-008] — Qualification: The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts. ## 9. Six Hundred Thirty Thousand Human Videos Learned a Robot’s Shape {#mp-2026-09-11-009} - Story ID: `mp-2026-09-11-009` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-009/six-hundred-thirty-thousand-human-videos-learned-a-robot-s-shape **Dek:** Robotizing observations and actions raised real-world completion from 51.5% to 80.3% as pretraining scale increased. HuRo converts heterogeneous human videos into robot-aligned observations and action trajectories, inferring missing intermediate signals across annotation levels. The resulting dataset contains roughly 630,000 episodes and 142 million processed frames from five video sources. Across four real manipulation tasks, scaling the robotized pretraining data increased overall completion from 51.5% to 80.3%; out-of-distribution completion under spatial and visual shifts rose from 34.9% to 72.2%. Ablations attribute some robustness to visual robotization and favor end-to-end retargeted actions over visual-only transfer. The evidence covers the authors’ four tasks, not arbitrary robots or internet video. ### Why it matters {#why-it-matters-mp-2026-09-11-009} Robotizing observations and actions raised real-world completion from 51.5% to 80.3% as pretraining scale increased. ### Limits and context {#limitations-mp-2026-09-11-009} - Ablations attribute some robustness to visual robotization and favor end-to-end retargeted actions over visual-only transfer. - The evidence covers the authors’ four tasks, not arbitrary robots or internet video. ### Claims and sources {#claims-mp-2026-09-11-009} - Robotizing observations and actions raised real-world completion from 51.5% to 80.3% as pretraining scale increased. [source-2026-09-11-009] — Qualification: Ablations attribute some robustness to visual robotization and favor end-to-end retargeted actions over visual-only transfer. ## 10. Household Robot Models Mishandled Roughly One Hazard in Three {#mp-2026-09-11-010} - Story ID: `mp-2026-09-11-010` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-010/household-robot-models-mishandled-roughly-one-hazard-in-three **Dek:** A physics-grounded benchmark executed every plan across more than 1,000 reproducible scenes and found safety failures did not shrink with scale. ReactHuman puts multimodal models in control of a simulated humanoid facing sudden household hazards such as slipping objects, with exact outcomes from 240-hertz rigid-body simulation. The suite spans 17 event families and more than 1,000 reproducible scenes, including adversarial props whose appearance contradicts their physical behavior. Across seven models, about one hazard in three was mishandled; systems often followed fixed dispositions, trusted appearance over motion and missed interception points by meters. Because every committed plan is executed, decisions have consequences inside the simulator, but the benchmark does not replace physical household-robot testing. ### Why it matters {#why-it-matters-mp-2026-09-11-010} A physics-grounded benchmark executed every plan across more than 1,000 reproducible scenes and found safety failures did not shrink with scale. ### Limits and context {#limitations-mp-2026-09-11-010} - Because every committed plan is executed, decisions have consequences inside the simulator, but the benchmark does not replace physical household-robot testing. ### Claims and sources {#claims-mp-2026-09-11-010} - A physics-grounded benchmark executed every plan across more than 1,000 reproducible scenes and found safety failures did not shrink with scale. [source-2026-09-11-010] — Qualification: Because every committed plan is executed, decisions have consequences inside the simulator, but the benchmark does not replace physical household-robot testing. ## 11. The Steering Failure Lived Between the Test Cases {#mp-2026-09-11-011} - Story ID: `mp-2026-09-11-011` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-011/the-steering-failure-lived-between-the-test-cases **Dek:** Formal bounds covered disturbance combinations that a driving campaign could not enumerate and exposed failures without another simulation run. Four small end-to-end steering networks were trained in CARLA on either clear conditions or a mix of clear, fog, night and low sun. Instead of driving every disturbance combination, bound propagation read the trained weights and calculated steering drift continuously between captured images. On one arterial route, 133 poses with ten intensities each would imply 10^133 combinations; the formal calculation finished in minutes on one GPU and found conditions that broke the clear-trained policy. The experiment is a simulated proof of concept, but it shows verification can examine spaces between discrete road tests. ### Why it matters {#why-it-matters-mp-2026-09-11-011} Formal bounds covered disturbance combinations that a driving campaign could not enumerate and exposed failures without another simulation run. ### Limits and context {#limitations-mp-2026-09-11-011} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-011} - Formal bounds covered disturbance combinations that a driving campaign could not enumerate and exposed failures without another simulation run. [source-2026-09-11-011] ## 12. Six Open Drones Avoided One Another at Speed {#mp-2026-09-11-012} - Story ID: `mp-2026-09-11-012` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-012/six-open-drones-avoided-one-another-at-speed **Dek:** SwarmNxt combines assembly guidance, parallel deployment tools and onboard ROS 2 autonomy for physical swarm experiments. SwarmNxt is an open software platform built around open OmniNxt drone hardware, with assembly instructions, swarm-wide update tooling and a ROS 2 stack for navigation. In one indoor experiment, six drones planned independently while avoiding high-speed inter-drone collisions; in another, four drones used onboard depth estimation to fly collectively through obstacles. Perception, planning and control ran onboard, while global position came from external motion capture. That dependence narrows the claim, but the release lowers the engineering barrier to reproducible physical swarm research. ### Why it matters {#why-it-matters-mp-2026-09-11-012} SwarmNxt combines assembly guidance, parallel deployment tools and onboard ROS 2 autonomy for physical swarm experiments. ### Limits and context {#limitations-mp-2026-09-11-012} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-012} - SwarmNxt combines assembly guidance, parallel deployment tools and onboard ROS 2 autonomy for physical swarm experiments. [source-2026-09-11-012] ## 13. Thirty-Four Thousand Parameters Read Failure From a Frozen Robot Model {#mp-2026-09-11-013} - Story ID: `mp-2026-09-11-013` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-013/thirty-four-thousand-parameters-read-failure-from-a-frozen-robot-model **Dek:** A lightweight readout reached 85.68 AUROC across seven source tasks and added 0.2256 milliseconds after the predictive state existed. FARM asks whether a frozen robotic world model already carries usable signs of impending failure. A supervised readout with 33,985 parameters maps its predictive states to stepwise failure scores and trajectory risk without updating the backbone. Five-fold evaluation across seven tasks reached pooled AUROC of 85.68 and AUPRC of 88.59, then fixed and readout-only adaptation were tested across PIPER X, SO-101 and Franka robot populations. Once the frozen state was available, mean added CUDA latency was 0.2256 milliseconds. These are predictive monitoring results, not proof that the system safely stops or recovers a robot. ### Why it matters {#why-it-matters-mp-2026-09-11-013} A lightweight readout reached 85.68 AUROC across seven source tasks and added 0.2256 milliseconds after the predictive state existed. ### Limits and context {#limitations-mp-2026-09-11-013} - Five-fold evaluation across seven tasks reached pooled AUROC of 85.68 and AUPRC of 88.59, then fixed and readout-only adaptation were tested across PIPER X, SO-101 and Franka robot populations. - These are predictive monitoring results, not proof that the system safely stops or recovers a robot. ### Claims and sources {#claims-mp-2026-09-11-013} - A lightweight readout reached 85.68 AUROC across seven source tasks and added 0.2256 milliseconds after the predictive state existed. [source-2026-09-11-013] — Qualification: Five-fold evaluation across seven tasks reached pooled AUROC of 85.68 and AUPRC of 88.59, then fixed and readout-only adaptation were tested across PIPER X, SO-101 and Franka robot populations. ## 14. The Humanoid Kept Walking as Its Depth Sensor Faded {#mp-2026-09-11-014} - Story ID: `mp-2026-09-11-014` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-014/the-humanoid-kept-walking-as-its-depth-sensor-faded **Dek:** One locomotion policy denoised partial depth and blended body-state information instead of switching between perceptive and blind controllers. CAP trains a perceptive encoder to reconstruct clean depth from corrupted input while a second encoder supplies proprioceptive body state. Depth-noise curriculum and feature dropout expose the same policy to a continuum from clean vision to perception loss. In simulation it matched or exceeded perceptive baselines with useful depth and degraded more smoothly than a binary-switching baseline as input worsened. Controlled Unitree G1 tests and indoor-outdoor deployments showed locomotion under occlusion, sensor corruption and outdoor depth artifacts. The abstract does not establish universal terrain safety or quantify all physical trial outcomes. ### Why it matters {#why-it-matters-mp-2026-09-11-014} One locomotion policy denoised partial depth and blended body-state information instead of switching between perceptive and blind controllers. ### Limits and context {#limitations-mp-2026-09-11-014} - The abstract does not establish universal terrain safety or quantify all physical trial outcomes. ### Claims and sources {#claims-mp-2026-09-11-014} - One locomotion policy denoised partial depth and blended body-state information instead of switching between perceptive and blind controllers. [source-2026-09-11-014] — Qualification: The abstract does not establish universal terrain safety or quantify all physical trial outcomes. ## 15. A Robot Hand Began Writing After Eighteen Seconds {#mp-2026-09-11-026} - Story ID: `mp-2026-09-11-026` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-026/a-robot-hand-began-writing-after-eighteen-seconds **Dek:** Real-time Jacobian estimation produced 0.6-millimeter mean in-plane precision without demonstrations, simulation training or an analytic contact model. The controller estimates the combined hand-and-object task Jacobian online, avoiding a prebuilt kinematic and contact model for a grasped pen. On a physical anthropomorphic hand it began writing after about 18 seconds of initialization using only a laptop CPU, then continued adapting as it traced letters and shapes in air and on paper. Mean in-plane error was 0.6 millimeters across runs. The same formulation was also exercised on two simulated hands. This is a narrow single-stroke dexterity result, but it offers a data-light alternative to reinforcement and imitation learning for contact-rich manipulation. ### Why it matters {#why-it-matters-mp-2026-09-11-026} Real-time Jacobian estimation produced 0.6-millimeter mean in-plane precision without demonstrations, simulation training or an analytic contact model. ### Limits and context {#limitations-mp-2026-09-11-026} - On a physical anthropomorphic hand it began writing after about 18 seconds of initialization using only a laptop CPU, then continued adapting as it traced letters and shapes in air and on paper. ### Claims and sources {#claims-mp-2026-09-11-026} - Real-time Jacobian estimation produced 0.6-millimeter mean in-plane precision without demonstrations, simulation training or an analytic contact model. [source-2026-09-11-015] — Qualification: On a physical anthropomorphic hand it began writing after about 18 seconds of initialization using only a laptop CPU, then continued adapting as it traced letters and shapes in air and on paper. ## 16. The Model Trace Reached More Than Half of Astronomy Papers {#mp-2026-09-11-027} - Story ID: `mp-2026-09-11-027` - Type: `dispatch` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-027/the-model-trace-reached-more-than-half-of-astronomy-papers **Dek:** A Bayesian study estimates 54% of 2025 astro-ph papers carried language-model markers, while 0.81% disclosed their use. The analysis counts distinctive vocabulary across 207,111 astro-ph papers from 2015 to mid-2026, calibrating unassisted prose with pre-2020 papers and assisted prose with 392 disclosures. Its central 2025 estimate is 54%, with an eight-point statistical interval and a one-sided systematic range as wide as 26 points because today’s no-model background cannot be observed directly. The estimate remains at least 36% under tested choices. Only 0.81% of 2025 papers disclosed model use—roughly one declaration per 66 papers carrying a trace. Vocabulary markers are indirect and adaptive, so the result estimates assistance rather than proving it paper by paper. ### Why it matters {#why-it-matters-mp-2026-09-11-027} A Bayesian study estimates 54% of 2025 astro-ph papers carried language-model markers, while 0.81% disclosed their use. ### Limits and context {#limitations-mp-2026-09-11-027} - Its central 2025 estimate is 54%, with an eight-point statistical interval and a one-sided systematic range as wide as 26 points because today’s no-model background cannot be observed directly. - Only 0.81% of 2025 papers disclosed model use—roughly one declaration per 66 papers carrying a trace. ### Claims and sources {#claims-mp-2026-09-11-027} - A Bayesian study estimates 54% of 2025 astro-ph papers carried language-model markers, while 0.81% disclosed their use. [source-2026-09-11-016] — Qualification: Its central 2025 estimate is 54%, with an eight-point statistical interval and a one-sided systematic range as wide as 26 points because today’s no-model background cannot be observed directly. ## 17. A Logical Bell Measurement Cleared the 50% Optical Limit {#mp-2026-09-11-015} - Story ID: `mp-2026-09-11-015` - Type: `ticker` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-015/a-logical-bell-measurement-cleared-the-50-optical-limit **Dek:** A two-qubit repetition code reached a mean 70.8% success probability with linear optics. The experiment used error-corrected logical qubits to achieve 70.8 ± 0.4% mean Bell-state measurement success, above the 50% ceiling for standard linear-optical measurements. The authors argue the code adds no extra overhead where error correction is already required. ### Why it matters {#why-it-matters-mp-2026-09-11-015} A two-qubit repetition code reached a mean 70.8% success probability with linear optics. ### Limits and context {#limitations-mp-2026-09-11-015} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-015} - A two-qubit repetition code reached a mean 70.8% success probability with linear optics. [source-2026-09-11-017] ## 18. Squeezed Light Reached Sixty Hertz on a Chip {#mp-2026-09-11-016} - Story ID: `mp-2026-09-11-016` - Type: `ticker` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-016/squeezed-light-reached-sixty-hertz-on-a-chip **Dek:** A coherent reference comb stabilized two-mode squeezing without directly locking the fragile quantum modes. A silica microcavity produced audio-band two-mode squeezed states down to 60 hertz. An orthogonally polarized reference comb enabled long-timescale phase control and covariance reconstruction; the reported symplectic eigenvalue of 0.395 confirmed inseparability. ### Why it matters {#why-it-matters-mp-2026-09-11-016} A coherent reference comb stabilized two-mode squeezing without directly locking the fragile quantum modes. ### Limits and context {#limitations-mp-2026-09-11-016} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-016} - A coherent reference comb stabilized two-mode squeezing without directly locking the fragile quantum modes. [source-2026-09-11-018] ## 19. One Quantum Surgery Used Ten Times Fewer Syndrome Rounds {#mp-2026-09-11-017} - Story ID: `mp-2026-09-11-017` - Type: `ticker` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-017/one-quantum-surgery-used-ten-times-fewer-syndrome-rounds **Dek:** Lifted surgery measured parallel logical operators in one round on explicit QLDPC code families. For a reported [[90,8,10]] quantum radial code, lifted surgery delivered logical performance comparable to standard surgery under circuit-level noise while requiring ten times fewer syndrome-measurement rounds. The construction preserves code symmetry and parallel addressability. ### Why it matters {#why-it-matters-mp-2026-09-11-017} Lifted surgery measured parallel logical operators in one round on explicit QLDPC code families. ### Limits and context {#limitations-mp-2026-09-11-017} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-017} - Lifted surgery measured parallel logical operators in one round on explicit QLDPC code families. [source-2026-09-11-019] ## 20. WASP-43b’s Bright Side Pointed Toward a Hot Interior {#mp-2026-09-11-018} - Story ID: `mp-2026-09-11-018` - Type: `ticker` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-018/wasp-43b-s-bright-side-pointed-toward-a-hot-interior **Dek:** Five TESS sectors showed a 44-degree eastward phase offset and little evidence for reflective dayside clouds. The team measured a 130 ± 34 parts-per-million eclipse depth and an eastward brightness offset of 44 ± 18 degrees. Compared with longer-wavelength JWST results, the TESS phase curve supports a cloud-poor dayside and a hot deep atmosphere. ### Why it matters {#why-it-matters-mp-2026-09-11-018} Five TESS sectors showed a 44-degree eastward phase offset and little evidence for reflective dayside clouds. ### Limits and context {#limitations-mp-2026-09-11-018} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-018} - Five TESS sectors showed a 44-degree eastward phase offset and little evidence for reflective dayside clouds. [source-2026-09-11-020] ## 21. A JWST Model Found Rare Galaxies Without Their Labels {#mp-2026-09-11-019} - Story ID: `mp-2026-09-11-019` - Type: `ticker` - Classification: `editorial` - Content status: `new` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-019/a-jwst-model-found-rare-galaxies-without-their-labels **Dek:** Self-supervision over 482,444 JADES objects separated high-redshift galaxies and Little Red Dots into discoverable islands. FM-JADES-v1 learned a joint embedding from multi-band imaging and catalog data without class labels. In a controlled three-band redshift test, its normalized median absolute deviation was 0.157 versus 0.44 for template fitting; the authors also report blind recovery of rare populations. ### Why it matters {#why-it-matters-mp-2026-09-11-019} Self-supervision over 482,444 JADES objects separated high-redshift galaxies and Little Red Dots into discoverable islands. ### Limits and context {#limitations-mp-2026-09-11-019} - No additional limitation was separately recorded. ### Claims and sources {#claims-mp-2026-09-11-019} - Self-supervision over 482,444 JADES objects separated high-redshift galaxies and Little Red Dots into discoverable islands. [source-2026-09-11-021] ## 22. Hangprinter {#mp-2026-09-11-020} - Story ID: `mp-2026-09-11-020` - Type: `invention_desk` - Classification: `editorial` - Content status: `carried_over` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-020/hangprinter **Dek:** Suspends a print head from tensioned lines anchored around a room, replacing a rigid gantry with cable geometry so an open RepRap can work across an unusually large build space. Suspends a print head from tensioned lines anchored around a room, replacing a rigid gantry with cable geometry so an open RepRap can work across an unusually large build space. ### Why it matters {#why-it-matters-mp-2026-09-11-020} An independent builder is turning an improbable idea into a working project. ### Limits and context {#limitations-mp-2026-09-11-020} - A Desk Pick is an editorial selection, not a product endorsement. ### Claims and sources {#claims-mp-2026-09-11-020} - This Invention Desk entry makes no independently sourced news claim. ## 23. Precious Plastic {#mp-2026-09-11-021} - Story ID: `mp-2026-09-11-021` - Type: `invention_desk` - Classification: `editorial` - Content status: `carried_over` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-021/precious-plastic **Dek:** Publishes replicable shredders, presses, workspace plans, and shared know-how so small local teams can sort waste plastic and turn it into reusable flakes and sheet material. Publishes replicable shredders, presses, workspace plans, and shared know-how so small local teams can sort waste plastic and turn it into reusable flakes and sheet material. ### Why it matters {#why-it-matters-mp-2026-09-11-021} An independent builder is turning an improbable idea into a working project. ### Limits and context {#limitations-mp-2026-09-11-021} - A Desk Pick is an editorial selection, not a product endorsement. ### Claims and sources {#claims-mp-2026-09-11-021} - This Invention Desk entry makes no independently sourced news claim. ## 24. Watchy {#mp-2026-09-11-022} - Story ID: `mp-2026-09-11-022` - Type: `invention_desk` - Classification: `editorial` - Content status: `carried_over` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-022/watchy **Dek:** Pairs a square e-paper display with an ESP32-S3 and publishes the hardware, software, documentation, and case files so owners can build and program their own watch faces. Pairs a square e-paper display with an ESP32-S3 and publishes the hardware, software, documentation, and case files so owners can build and program their own watch faces. ### Why it matters {#why-it-matters-mp-2026-09-11-022} An independent builder is turning an improbable idea into a working project. ### Limits and context {#limitations-mp-2026-09-11-022} - A Desk Pick is an editorial selection, not a product endorsement. ### Claims and sources {#claims-mp-2026-09-11-022} - This Invention Desk entry makes no independently sourced news claim. ## 25. Ploopy Classic 2 {#mp-2026-09-11-023} - Story ID: `mp-2026-09-11-023` - Type: `invention_desk` - Classification: `editorial` - Content status: `carried_over` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-023/ploopy-classic-2 **Dek:** Turns a desktop trackball into an inspectable kit by publishing its mechanical and electrical design files, assembly documentation, and programmable QMK firmware. Turns a desktop trackball into an inspectable kit by publishing its mechanical and electrical design files, assembly documentation, and programmable QMK firmware. ### Why it matters {#why-it-matters-mp-2026-09-11-023} An independent builder is turning an improbable idea into a working project. ### Limits and context {#limitations-mp-2026-09-11-023} - A Desk Pick is an editorial selection, not a product endorsement. ### Claims and sources {#claims-mp-2026-09-11-023} - This Invention Desk entry makes no independently sourced news claim. ## 26. The First Paid Slot {#mp-2026-09-11-024} - Story ID: `mp-2026-09-11-024` - Type: `invention_desk` - Classification: `house_example` - Content status: `carried_over` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-024/the-first-paid-slot **Dek:** A transparent preview of paid placement with one verified link and no claim of endorsement. A transparent preview of paid placement with one verified link and no claim of endorsement. House example - no advertiser paid. Payment will buy placement, never endorsement. ### Why it matters {#why-it-matters-mp-2026-09-11-024} This placement explains how builders can appear in The Invention Desk without purchasing editorial endorsement. ### Limits and context {#limitations-mp-2026-09-11-024} - House example - no advertiser paid. Payment will buy placement, never endorsement. ### Claims and sources {#claims-mp-2026-09-11-024} - This Invention Desk entry makes no independently sourced news claim. ## 27. Put Your Project on the Desk {#mp-2026-09-11-025} - Story ID: `mp-2026-09-11-025` - Type: `invention_desk` - Classification: `house_example` - Content status: `carried_over` - Permanent URL: https://themachinepress.com/story/mp-2026-09-11-025/put-your-project-on-the-desk **Dek:** One manually reviewed placement stays active for seven days and remains separate from Desk Picks. One manually reviewed placement stays active for seven days and remains separate from Desk Picks. Manual intake only. Payment buys placement, never endorsement, and every submission is reviewed. ### Why it matters {#why-it-matters-mp-2026-09-11-025} This placement explains how builders can appear in The Invention Desk without purchasing editorial endorsement. ### Limits and context {#limitations-mp-2026-09-11-025} - Manual intake only. Payment buys placement, never endorsement, and every submission is reviewed. ### Claims and sources {#claims-mp-2026-09-11-025} - This Invention Desk entry makes no independently sourced news claim. ## Normalized sources - **source-2026-09-11-001:** [arXiv preprint 2609.11060](https://arxiv.org/abs/2609.11060) — arXiv; primary_research - **source-2026-09-11-002:** [arXiv preprint 2609.11922](https://arxiv.org/abs/2609.11922) — arXiv; primary_research - **source-2026-09-11-003:** [arXiv preprint 2609.10712](https://arxiv.org/abs/2609.10712) — arXiv; primary_research - **source-2026-09-11-004:** [arXiv preprint 2609.10964](https://arxiv.org/abs/2609.10964) — arXiv; primary_research - **source-2026-09-11-005:** [arXiv preprint 2609.11061](https://arxiv.org/abs/2609.11061) — arXiv; primary_research - **source-2026-09-11-006:** [arXiv preprint 2609.11065](https://arxiv.org/abs/2609.11065) — arXiv; primary_research - **source-2026-09-11-007:** [arXiv preprint 2609.11490](https://arxiv.org/abs/2609.11490) — arXiv; primary_research - **source-2026-09-11-008:** [arXiv preprint 2609.11542](https://arxiv.org/abs/2609.11542) — arXiv; primary_research - **source-2026-09-11-009:** [arXiv preprint 2609.10706](https://arxiv.org/abs/2609.10706) — arXiv; primary_research - **source-2026-09-11-010:** [arXiv preprint 2609.10895](https://arxiv.org/abs/2609.10895) — arXiv; primary_research - **source-2026-09-11-011:** [arXiv preprint 2609.10951](https://arxiv.org/abs/2609.10951) — arXiv; primary_research - **source-2026-09-11-012:** [arXiv preprint 2609.11382](https://arxiv.org/abs/2609.11382) — arXiv; primary_research - **source-2026-09-11-013:** [arXiv preprint 2609.11445](https://arxiv.org/abs/2609.11445) — arXiv; primary_research - **source-2026-09-11-014:** [arXiv preprint 2609.11553](https://arxiv.org/abs/2609.11553) — arXiv; primary_research - **source-2026-09-11-015:** [arXiv preprint 2609.11775](https://arxiv.org/abs/2609.11775) — arXiv; primary_research - **source-2026-09-11-016:** [arXiv preprint 2609.10664](https://arxiv.org/abs/2609.10664) — arXiv; primary_research - **source-2026-09-11-017:** [arXiv preprint 2609.10698](https://arxiv.org/abs/2609.10698) — arXiv; primary_research - **source-2026-09-11-018:** [arXiv preprint 2609.11004](https://arxiv.org/abs/2609.11004) — arXiv; primary_research - **source-2026-09-11-019:** [arXiv preprint 2609.11723](https://arxiv.org/abs/2609.11723) — arXiv; primary_research - **source-2026-09-11-020:** [arXiv preprint 2609.10719](https://arxiv.org/abs/2609.10719) — arXiv; primary_research - **source-2026-09-11-021:** [arXiv preprint 2609.11879](https://arxiv.org/abs/2609.11879) — arXiv; primary_research