---
schema_version: "1.0.0"
edition_id: "mp-2026-09-03-morning-0056"
published_at: "2026-09-03T09:00:00.000-04:00"
modified_at: "2026-09-03T14:20:28.218-04:00"
canonical_url: "https://themachinepress.com/edition/2026-09-03"
story_count: 27
lead_story_id: "mp-2026-09-03-001"
---

# The Machine Press — Morning edition

Edition ID: `mp-2026-09-03-morning-0056`  
Published: 2026-09-03T09:00:00.000-04:00  
Canonical edition: https://themachinepress.com/edition/2026-09-03

Across 24 sensing conditions, a long measurement-free rollout chose a different estimator from the closed-loop optimum 18 times.

## 1. The Best Imagined Future Picked the Wrong Controller {#mp-2026-09-03-001}

- Story ID: `mp-2026-09-03-001`
- Type: `lead`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-001/the-best-imagined-future-picked-the-wrong-controller

**Dek:** Across 24 sensing conditions, a long measurement-free rollout chose a different estimator from the closed-loop optimum 18 times.

A controlled differential-drive study compared six state estimators by replay error, a 20-step measurement-free rollout and actual closed-loop path tracking under biased odometry and intermittent landmarks. Replay position error tracked closed-loop cross-track error more closely than rollout error, with Spearman correlations of 0.923 and 0.774; replay selected a different estimator from the closed-loop optimum in 5 of 24 conditions, while the rollout metric did so in 18. Long predictions remained useful when regular measurement corrections were preserved, making the sensing and update schedule part of the evaluation—not a detail to omit.

### Why it matters {#why-it-matters-mp-2026-09-03-001}

Across 24 sensing conditions, a long measurement-free rollout chose a different estimator from the closed-loop optimum 18 times.

### Limits and context {#limitations-mp-2026-09-03-001}

- Long predictions remained useful when regular measurement corrections were preserved, making the sensing and update schedule part of the evaluation—not a detail to omit.

### Claims and sources {#claims-mp-2026-09-03-001}

- Across 24 sensing conditions, a long measurement-free rollout chose a different estimator from the closed-loop optimum 18 times. [source-2026-09-03-001] — Qualification: Long predictions remained useful when regular measurement corrections were preserved, making the sensing and update schedule part of the evaluation—not a detail to omit.

## 2. Ten Video Sources Became One Long-Horizon World {#mp-2026-09-03-002}

- Story ID: `mp-2026-09-03-002`
- Type: `secondary`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-002/ten-video-sources-became-one-long-horizon-world

**Dek:** SolarWM unifies 1.43 million clips and adapts four 5B-to-33B video backbones behind shared training and inference interfaces.

SolarWM converts clips from ten datasets into a frame-aligned record carrying observations, metric camera geometry, captions, quality metadata, selection decisions and provenance. A backbone-native layer then applies one three-stage recipe across four models based on Wan2.2, LTX-2.5 and MiniMax-H3 without erasing their native representations. The authors say causal models trained on five-second sequences can support interactive rollouts from minutes to hours, and they are releasing the data, pipeline, recipes, weights and framework; those are preprint claims, not independent replication.

### Why it matters {#why-it-matters-mp-2026-09-03-002}

SolarWM unifies 1.43 million clips and adapts four 5B-to-33B video backbones behind shared training and inference interfaces.

### Limits and context {#limitations-mp-2026-09-03-002}

- The authors say causal models trained on five-second sequences can support interactive rollouts from minutes to hours, and they are releasing the data, pipeline, recipes, weights and framework; those are preprint claims, not independent replication.

### Claims and sources {#claims-mp-2026-09-03-002}

- SolarWM unifies 1.43 million clips and adapts four 5B-to-33B video backbones behind shared training and inference interfaces. [source-2026-09-03-002] — Qualification: The authors say causal models trained on five-second sequences can support interactive rollouts from minutes to hours, and they are releasing the data, pipeline, recipes, weights and framework; those are preprint claims, not independent replication.

## 3. The Web Agent Predicted Differences, Not Just Pages {#mp-2026-09-03-003}

- Story ID: `mp-2026-09-03-003`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-003/the-web-agent-predicted-differences-not-just-pages

**Dek:** A matching objective trains predicted states to separate the true result of an action from the states produced by alternatives.

Most web-agent world models learn to reproduce the next HTML or accessibility-tree snapshot, even though the downstream ranker needs to tell candidate actions apart. The new objective uses branching WebArena Go-Browse trajectories with multiple actions and resulting states at each decision point. The authors report better predicted-state matching, action ranking on WebPRMBench and end-to-end success on WebArena-Lite than action-only or supervised-next-state comparisons.

### Why it matters {#why-it-matters-mp-2026-09-03-003}

A matching objective trains predicted states to separate the true result of an action from the states produced by alternatives.

### Limits and context {#limitations-mp-2026-09-03-003}

- The authors report better predicted-state matching, action ranking on WebPRMBench and end-to-end success on WebArena-Lite than action-only or supervised-next-state comparisons.

### Claims and sources {#claims-mp-2026-09-03-003}

- A matching objective trains predicted states to separate the true result of an action from the states produced by alternatives. [source-2026-09-03-003] — Qualification: The authors report better predicted-state matching, action ranking on WebPRMBench and end-to-end success on WebArena-Lite than action-only or supervised-next-state comparisons.

## 4. The Model's Explanation May Not Describe Its Computation {#mp-2026-09-03-004}

- Story ID: `mp-2026-09-03-004`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-004/the-model-s-explanation-may-not-describe-its-computation

**Dek:** A security argument says linguistic monitoring cannot supply complete guarantees when internal activation-space computation is only lossy-translated into words.

James Mickens names the gap “linguistic illegibility”: external text or language-like probes may fail to represent how a model produced an outcome. The paper argues that chain-of-thought monitoring, self-critique and linguistically defined activation probes therefore need a non-linguistic security floor. It proposes output taint tracking, robust virtualization and independent configuration audits as complementary isolation measures; this is a position and design argument, not a measured proof that every monitor fails.

### Why it matters {#why-it-matters-mp-2026-09-03-004}

A security argument says linguistic monitoring cannot supply complete guarantees when internal activation-space computation is only lossy-translated into words.

### Limits and context {#limitations-mp-2026-09-03-004}

- It proposes output taint tracking, robust virtualization and independent configuration audits as complementary isolation measures; this is a position and design argument, not a measured proof that every monitor fails.

### Claims and sources {#claims-mp-2026-09-03-004}

- A security argument says linguistic monitoring cannot supply complete guarantees when internal activation-space computation is only lossy-translated into words. [source-2026-09-03-004] — Qualification: It proposes output taint tracking, robust virtualization and independent configuration audits as complementary isolation measures; this is a position and design argument, not a measured proof that every monitor fails.

## 5. Four-Bit Training Dropped the Extra Rotation {#mp-2026-09-03-005}

- Story ID: `mp-2026-09-03-005`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-005/four-bit-training-dropped-the-extra-rotation

**Dek:** Unsigned E5M3 block scales replace randomized Hadamard transforms and BF16 final-layer exemptions in one FP4 pretraining recipe.

The recipe pairs E2M1 payloads with wider-range unsigned E5M3 scales, uses periodic tensor scaling and applies stochastic rounding selectively to backward gradients. In software-emulated training of an 8B Nemotron-H model for nearly 190 billion tokens, the authors report lower final-window and held-out validation loss than their Transformer Engine NVFP4 comparison. A separate ablation that removed the transform and final-block exemption raised measured model-body throughput by 21.2%, motivating native hardware support rather than proving a universal speedup.

### Why it matters {#why-it-matters-mp-2026-09-03-005}

Unsigned E5M3 block scales replace randomized Hadamard transforms and BF16 final-layer exemptions in one FP4 pretraining recipe.

### Limits and context {#limitations-mp-2026-09-03-005}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-09-03-005}

- Unsigned E5M3 block scales replace randomized Hadamard transforms and BF16 final-layer exemptions in one FP4 pretraining recipe. [source-2026-09-03-005]

## 6. The Table Model Interpolated Physics but Lost the Units {#mp-2026-09-03-006}

- Story ID: `mp-2026-09-03-006`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-006/the-table-model-interpolated-physics-but-lost-the-units

**Dek:** Four tabular foundation models beat six baselines on samples from 316 equations yet could not represent noiseless mechanisms or physical units.

The study treats table completion as a probe of what physics-like priors tabular foundation models acquire. Across in-domain and out-of-domain datasets sampled from 316 equations, the four tested systems led the baselines before and after tuning. The authors' central negative result is that those priors still cannot encode a deterministic mechanism or units, so strong interpolation should not be confused with a physical model.

### Why it matters {#why-it-matters-mp-2026-09-03-006}

Four tabular foundation models beat six baselines on samples from 316 equations yet could not represent noiseless mechanisms or physical units.

### Limits and context {#limitations-mp-2026-09-03-006}

- The authors' central negative result is that those priors still cannot encode a deterministic mechanism or units, so strong interpolation should not be confused with a physical model.

### Claims and sources {#claims-mp-2026-09-03-006}

- Four tabular foundation models beat six baselines on samples from 316 equations yet could not represent noiseless mechanisms or physical units. [source-2026-09-03-006] — Qualification: The authors' central negative result is that those priors still cannot encode a deterministic mechanism or units, so strong interpolation should not be confused with a physical model.

## 7. The Compressed Model Never Needed Its Source Code {#mp-2026-09-03-007}

- Story ID: `mp-2026-09-03-007`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-007/the-compressed-model-never-needed-its-source-code

**Dek:** H3DNAS searches and rewrites ONNX graphs directly for edge deployment of 3D point-cloud networks.

A channel-dependency graph classifies ONNX operators and computes a topology-defined ceiling on how much of a model can be pruned. A two-stage search then selects channels by importance, ranks candidates by output fidelity without labels and mutates Pareto candidates with GhostConv. On ModelNet40, the paper reports parameter reductions of 43.2% to 65.5% and Jetson Orin Nano speedups of 1.29× to 1.99× across three architectures with negligible accuracy loss under its evaluation.

### Why it matters {#why-it-matters-mp-2026-09-03-007}

H3DNAS searches and rewrites ONNX graphs directly for edge deployment of 3D point-cloud networks.

### Limits and context {#limitations-mp-2026-09-03-007}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-09-03-007}

- H3DNAS searches and rewrites ONNX graphs directly for edge deployment of 3D point-cloud networks. [source-2026-09-03-007]

## 8. Two-Bit Weights Met the Cost of Their Decoder {#mp-2026-09-03-008}

- Story ID: `mp-2026-09-03-008`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-008/two-bit-weights-met-the-cost-of-their-decoder

**Dek:** A full multi-shell Leech-lattice decoder exposes the gap between compact on-disk weights and the bytes a serving kernel actually reads.

The implementation expands a 301-class codebook into GPU layouts and fuses dequantization with matrix-vector multiplication for batch-one decoding. Binary bit planes beat one-hot masks, but the served layout still read 2.40× more bytes and ran 2.27× slower than the compared trellis kernel. With a shared output head, end-to-end gains rose from 1.11× at 4B to 1.41× at 14B, while the 4B quality cost remained substantial, underscoring that nominal bit rate is not the whole serving story.

### Why it matters {#why-it-matters-mp-2026-09-03-008}

A full multi-shell Leech-lattice decoder exposes the gap between compact on-disk weights and the bytes a serving kernel actually reads.

### Limits and context {#limitations-mp-2026-09-03-008}

- With a shared output head, end-to-end gains rose from 1.11× at 4B to 1.41× at 14B, while the 4B quality cost remained substantial, underscoring that nominal bit rate is not the whole serving story.

### Claims and sources {#claims-mp-2026-09-03-008}

- A full multi-shell Leech-lattice decoder exposes the gap between compact on-disk weights and the bytes a serving kernel actually reads. [source-2026-09-03-008] — Qualification: With a shared output head, end-to-end gains rose from 1.11× at 4B to 1.41× at 14B, while the 4B quality cost remained substantial, underscoring that nominal bit rate is not the whole serving story.

## 9. Clean LiDAR Rankings Broke Under Deployment Conditions {#mp-2026-09-03-009}

- Story ID: `mp-2026-09-03-009`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-009/clean-lidar-rankings-broke-under-deployment-conditions

**Dek:** A protocol adds coarse safety labels, eight corruption types, cross-domain tests and embedded-device speed to semantic-segmentation evaluation.

The proposed evaluation asks whether LiDAR segmentation rankings survive safety-relevant label grouping, atmospheric and sensor degradation, and transfer to another dataset without adaptation. It also measures inference on a Jetson AGX Orin. The reported results show architecture-dependent degradation under every corruption, rankings that change with coarse labels and insufficient current domain generalization, framing a test protocol rather than a new production perception stack.

### Why it matters {#why-it-matters-mp-2026-09-03-009}

A protocol adds coarse safety labels, eight corruption types, cross-domain tests and embedded-device speed to semantic-segmentation evaluation.

### Limits and context {#limitations-mp-2026-09-03-009}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-09-03-009}

- A protocol adds coarse safety labels, eight corruption types, cross-domain tests and embedded-device speed to semantic-segmentation evaluation. [source-2026-09-03-009]

## 10. The Robot Reconsidered Intent Only When the Task Changed {#mp-2026-09-03-010}

- Story ID: `mp-2026-09-03-010`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-010/the-robot-reconsidered-intent-only-when-the-task-changed

**Dek:** HINT separates sparse semantic decisions from continuous object-hand tracking in long-horizon manipulation.

The framework invokes semantic reasoning at manipulation-pattern transitions to select the subtask and target, then holds that commitment through multiview grounding and visual tracking. It can pass the tracked intent through either image-space highlighting or attention-prior injection without training the foundation action model. Across three long-horizon tasks and out-of-distribution variants, the authors report higher intent understanding, progress and completion for two policies while preserving low-latency control.

### Why it matters {#why-it-matters-mp-2026-09-03-010}

HINT separates sparse semantic decisions from continuous object-hand tracking in long-horizon manipulation.

### Limits and context {#limitations-mp-2026-09-03-010}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-09-03-010}

- HINT separates sparse semantic decisions from continuous object-hand tracking in long-horizon manipulation. [source-2026-09-03-010]

## 11. The Data-Center Plan Learned From the Price It Changed {#mp-2026-09-03-011}

- Story ID: `mp-2026-09-03-011`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-011/the-data-center-plan-learned-from-the-price-it-changed

**Dek:** Differentiating through electricity-market clearing lets a planner search where to place 50 megawatts across six candidate buses.

Each forward pass clears a synthetic electricity market; reverse-mode differentiation then sends the resulting price effects back into the siting plan. Across 36 operating states in two networks, the continuous allocations nearly matched exhaustive enumeration, with worst-case gaps equal to 2.3% and 8.5% of the best-to-worst single-site cost spread. The smooth relaxation closed sites too late near discrete boundaries, a limitation the authors identify explicitly.

### Why it matters {#why-it-matters-mp-2026-09-03-011}

Differentiating through electricity-market clearing lets a planner search where to place 50 megawatts across six candidate buses.

### Limits and context {#limitations-mp-2026-09-03-011}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-09-03-011}

- Differentiating through electricity-market clearing lets a planner search where to place 50 megawatts across six candidate buses. [source-2026-09-03-011]

## 12. Five Percent Exposure Changed What Zero-Shot Meant {#mp-2026-09-03-012}

- Story ID: `mp-2026-09-03-012`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-012/five-percent-exposure-changed-what-zero-shot-meant

**Dek:** ZETA separates truly unseen robot hardware from embodiments glimpsed during pretraining across 14 held-out targets.

The benchmark holds tasks, scenes and protocols steady while changing robot embodiment, distinguishing strict zero-shot transfer from pretrain-exposed transfer. Local end-effector representations, source-embodiment diversity and auxiliary co-training improved average transfer by about 15, 18 and 7 percentage points in the reported analysis. Adding only 5% target-embodiment data during pretraining raised target progress by 13.4 points, showing why the two zero-shot conditions should not share one label.

### Why it matters {#why-it-matters-mp-2026-09-03-012}

ZETA separates truly unseen robot hardware from embodiments glimpsed during pretraining across 14 held-out targets.

### Limits and context {#limitations-mp-2026-09-03-012}

- Adding only 5% target-embodiment data during pretraining raised target progress by 13.4 points, showing why the two zero-shot conditions should not share one label.

### Claims and sources {#claims-mp-2026-09-03-012}

- ZETA separates truly unseen robot hardware from embodiments glimpsed during pretraining across 14 held-out targets. [source-2026-09-03-012] — Qualification: Adding only 5% target-embodiment data during pretraining raised target progress by 13.4 points, showing why the two zero-shot conditions should not share one label.

## 13. The Humanoid Looked Beyond the Next Foothold {#mp-2026-09-03-013}

- Story ID: `mp-2026-09-03-013`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-013/the-humanoid-looked-beyond-the-next-foothold

**Dek:** A recurrent predictive feature guides a locomotion policy across gaps, stepping stones and narrow stairs from one depth stream.

WM-LOCO jointly trains a recurrent world model with a PPO policy so near-future observations and rewards can shape foot placement without explicit foothold labels. In simulation it succeeded on gaps and stepping stones where the matched baseline failed, while matching stair success and improving stride efficiency and pelvis acceleration. The same policy ran on a Unitree G1 and averaged 93.3% success across the three terrain classes in the authors' physical tests.

### Why it matters {#why-it-matters-mp-2026-09-03-013}

A recurrent predictive feature guides a locomotion policy across gaps, stepping stones and narrow stairs from one depth stream.

### Limits and context {#limitations-mp-2026-09-03-013}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-09-03-013}

- A recurrent predictive feature guides a locomotion policy across gaps, stepping stones and narrow stairs from one depth stream. [source-2026-09-03-013]

## 14. Task Success Fell When Safe Contact Counted {#mp-2026-09-03-014}

- Story ID: `mp-2026-09-03-014`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-014/task-success-fell-when-safe-contact-counted

**Dek:** A bathing benchmark pairs a deformable simulated person with region and force screening calibrated against a care manikin.

The protocol freezes a vision-only scorer and adds physics-aware contact measures to ordinary completion, reducing opportunities for evaluator leakage. Across 140 runs per method, the LLM-augmented state machine completed 72.9% of tasks but only 56.4% survived correct-region and force-safety checks; VoxPoser completed 27.9%, and zero-shot pi0.5 completed 0.7%. The benchmark argues that task completion alone is an unsafe proxy for contact-rich assistive behavior.

### Why it matters {#why-it-matters-mp-2026-09-03-014}

A bathing benchmark pairs a deformable simulated person with region and force screening calibrated against a care manikin.

### Limits and context {#limitations-mp-2026-09-03-014}

- The protocol freezes a vision-only scorer and adds physics-aware contact measures to ordinary completion, reducing opportunities for evaluator leakage.
- Across 140 runs per method, the LLM-augmented state machine completed 72.9% of tasks but only 56.4% survived correct-region and force-safety checks; VoxPoser completed 27.9%, and zero-shot pi0.5 completed 0.7%.

### Claims and sources {#claims-mp-2026-09-03-014}

- A bathing benchmark pairs a deformable simulated person with region and force screening calibrated against a care manikin. [source-2026-09-03-014] — Qualification: The protocol freezes a vision-only scorer and adds physics-aware contact measures to ordinary completion, reducing opportunities for evaluator leakage.

## 15. Plausible Pictures Failed the Global Rule {#mp-2026-09-03-026}

- Story ID: `mp-2026-09-03-026`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-026/plausible-pictures-failed-the-global-rule

**Dek:** RIG-BENCH tests whether generators can infer a hidden visual rule and render a logically constrained answer across 2,000 samples.

The benchmark spans concept, transformation, pattern-and-structure, and scenario reasoning rather than asking only whether an image matches surface events. Evaluations of unified generative models and image/video systems found a recurring gap: outputs could look locally plausible while violating the global constraint. The result is a diagnostic dataset and preprint claim, not evidence that every visual generator fails every reasoning task.

### Why it matters {#why-it-matters-mp-2026-09-03-026}

RIG-BENCH tests whether generators can infer a hidden visual rule and render a logically constrained answer across 2,000 samples.

### Limits and context {#limitations-mp-2026-09-03-026}

- The benchmark spans concept, transformation, pattern-and-structure, and scenario reasoning rather than asking only whether an image matches surface events.
- The result is a diagnostic dataset and preprint claim, not evidence that every visual generator fails every reasoning task.

### Claims and sources {#claims-mp-2026-09-03-026}

- RIG-BENCH tests whether generators can infer a hidden visual rule and render a logically constrained answer across 2,000 samples. [source-2026-09-03-015] — Qualification: The benchmark spans concept, transformation, pattern-and-structure, and scenario reasoning rather than asking only whether an image matches surface events.

## 16. Speech Interfaces Got a Common Information Scale {#mp-2026-09-03-027}

- Story ID: `mp-2026-09-03-027`
- Type: `dispatch`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-027/speech-interfaces-got-a-common-information-scale

**Dek:** Open-vocabulary mutual information compares neural speech decoders even when their datasets, vocabularies and recording setups differ.

Accuracy and word error rate can overstate communication when they score only the words a brain-computer interface already supports. OVMI instead measures conveyed information against a reference distribution of what a user may wish to say, putting coverage and decoding accuracy on the same scale. Applied to existing systems, it exposed vocabulary tradeoffs and found that selecting a vocabulary to maximize OVMI improved accuracy by as much as 16.3% across three speech domains.

### Why it matters {#why-it-matters-mp-2026-09-03-027}

Open-vocabulary mutual information compares neural speech decoders even when their datasets, vocabularies and recording setups differ.

### Limits and context {#limitations-mp-2026-09-03-027}

- Accuracy and word error rate can overstate communication when they score only the words a brain-computer interface already supports.

### Claims and sources {#claims-mp-2026-09-03-027}

- Open-vocabulary mutual information compares neural speech decoders even when their datasets, vocabularies and recording setups differ. [source-2026-09-03-016] — Qualification: Accuracy and word error rate can overstate communication when they score only the words a brain-computer interface already supports.

## 17. Robot Decisions Got Four Audit Layers {#mp-2026-09-03-015}

- Story ID: `mp-2026-09-03-015`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-015/robot-decisions-got-four-audit-layers

**Dek:** TRACE connects sensor evidence, belief state, planning and execution verification in one reconstructable chain.

In 500 simulated warehouse decision cycles, the paper reports 98.6% evidence traceability, 99.0% temporal continuity and 98.1% reconstructability. The figures describe the evaluated framework, not field deployment.

### Why it matters {#why-it-matters-mp-2026-09-03-015}

TRACE connects sensor evidence, belief state, planning and execution verification in one reconstructable chain.

### Limits and context {#limitations-mp-2026-09-03-015}

- The figures describe the evaluated framework, not field deployment.

### Claims and sources {#claims-mp-2026-09-03-015}

- TRACE connects sensor evidence, belief state, planning and execution verification in one reconstructable chain. [source-2026-09-03-017] — Qualification: The figures describe the evaluated framework, not field deployment.

## 18. Safety Experience Rewrote Both Harness and Policy {#mp-2026-09-03-016}

- Story ID: `mp-2026-09-03-016`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-016/safety-experience-rewrote-both-harness-and-policy

**Dek:** SafeEvolve turns trajectory evidence into reversible harness updates, then trains the agent to use them.

On AgentDojo with Qwen3.5-4B, the authors report a threefold attack-success reduction while benign utility rose from 59.79% to 61.86%. This is a benchmark result for one co-evolution setup.

### Why it matters {#why-it-matters-mp-2026-09-03-016}

SafeEvolve turns trajectory evidence into reversible harness updates, then trains the agent to use them.

### Limits and context {#limitations-mp-2026-09-03-016}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-09-03-016}

- SafeEvolve turns trajectory evidence into reversible harness updates, then trains the agent to use them. [source-2026-09-03-018]

## 19. A Thousand Repositories Became Five Thousand Skills {#mp-2026-09-03-017}

- Story ID: `mp-2026-09-03-017`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-017/a-thousand-repositories-became-five-thousand-skills

**Dek:** DisCo distills operational knowledge from ML codebases into compact, verified task context.

With backbone, harness and execution budget held fixed, the paper reports gains from 9.2% to 134.3% across four research benchmarks. The comparison supports the tested library, not every repository-to-skill conversion.

### Why it matters {#why-it-matters-mp-2026-09-03-017}

DisCo distills operational knowledge from ML codebases into compact, verified task context.

### Limits and context {#limitations-mp-2026-09-03-017}

- The comparison supports the tested library, not every repository-to-skill conversion.

### Claims and sources {#claims-mp-2026-09-03-017}

- DisCo distills operational knowledge from ML codebases into compact, verified task context. [source-2026-09-03-019] — Qualification: The comparison supports the tested library, not every repository-to-skill conversion.

## 20. The LLM Judge Was Demoted to Adviser {#mp-2026-09-03-018}

- Story ID: `mp-2026-09-03-018`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-018/the-llm-judge-was-demoted-to-adviser

**Dek:** PROCTOR makes deterministic checks outrank the model grading an agent's self-improvement.

The authors catalog eleven evaluation failures, including a perfect score that hid 68% true capability after cached answers leaked. Their proposed roles, sandboxes, holdouts and canaries are production case-study lessons, not a universal benchmark.

### Why it matters {#why-it-matters-mp-2026-09-03-018}

PROCTOR makes deterministic checks outrank the model grading an agent's self-improvement.

### Limits and context {#limitations-mp-2026-09-03-018}

- Their proposed roles, sandboxes, holdouts and canaries are production case-study lessons, not a universal benchmark.

### Claims and sources {#claims-mp-2026-09-03-018}

- PROCTOR makes deterministic checks outrank the model grading an agent's self-improvement. [source-2026-09-03-020] — Qualification: Their proposed roles, sandboxes, holdouts and canaries are production case-study lessons, not a universal benchmark.

## 21. Each Attention Head Got Its Own Memory Window {#mp-2026-09-03-019}

- Story ID: `mp-2026-09-03-019`
- Type: `ticker`
- Classification: `editorial`
- Content status: `new`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-019/each-attention-head-got-its-own-memory-window

**Dek:** HeadWiseKV assigns static per-head histories under a total cache budget without retraining the model.

On Qwen3.6-27B, the fixed-model study reduced sampled peak device memory 8.59% at 112K context and extended the largest verified context from 114K to 161K while retaining near-full-cache quality on the tested suites.

### Why it matters {#why-it-matters-mp-2026-09-03-019}

HeadWiseKV assigns static per-head histories under a total cache budget without retraining the model.

### Limits and context {#limitations-mp-2026-09-03-019}

- No additional limitation was separately recorded.

### Claims and sources {#claims-mp-2026-09-03-019}

- HeadWiseKV assigns static per-head histories under a total cache budget without retraining the model. [source-2026-09-03-021]

## 22. MNT Reform Next {#mp-2026-09-03-020}

- Story ID: `mp-2026-09-03-020`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-020/mnt-reform-next

**Dek:** Reworks a laptop into public, swappable modules: processor, port boards, keyboard, trackpad, and user-serviceable battery packs can evolve without sealing the whole machine.

Reworks a laptop into public, swappable modules: processor, port boards, keyboard, trackpad, and user-serviceable battery packs can evolve without sealing the whole machine.

### Why it matters {#why-it-matters-mp-2026-09-03-020}

An independent builder is turning an improbable idea into a working project.

### Limits and context {#limitations-mp-2026-09-03-020}

- A Desk Pick is an editorial selection, not a product endorsement.

### Claims and sources {#claims-mp-2026-09-03-020}

- This Invention Desk entry makes no independently sourced news claim.

## 23. Maslow 4 {#mp-2026-09-03-021}

- Story ID: `mp-2026-09-03-021`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-021/maslow-4

**Dek:** Pulls a compact router sled across full sheets with four measured belts, trading a bulky gantry for corner anchors and community-developed control software.

Pulls a compact router sled across full sheets with four measured belts, trading a bulky gantry for corner anchors and community-developed control software.

### Why it matters {#why-it-matters-mp-2026-09-03-021}

An independent builder is turning an improbable idea into a working project.

### Limits and context {#limitations-mp-2026-09-03-021}

- A Desk Pick is an editorial selection, not a product endorsement.

### Claims and sources {#claims-mp-2026-09-03-021}

- This Invention Desk entry makes no independently sourced news claim.

## 24. FarmBot Genesis {#mp-2026-09-03-022}

- Story ID: `mp-2026-09-03-022`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-022/farmbot-genesis

**Dek:** Moves an interchangeable tool head across a raised bed to place seeds, water plants, and measure soil, backed by published hardware, software, data, and documentation.

Moves an interchangeable tool head across a raised bed to place seeds, water plants, and measure soil, backed by published hardware, software, data, and documentation.

### Why it matters {#why-it-matters-mp-2026-09-03-022}

An independent builder is turning an improbable idea into a working project.

### Limits and context {#limitations-mp-2026-09-03-022}

- A Desk Pick is an editorial selection, not a product endorsement.

### Claims and sources {#claims-mp-2026-09-03-022}

- This Invention Desk entry makes no independently sourced news claim.

## 25. OpenBikeSensor {#mp-2026-09-03-023}

- Story ID: `mp-2026-09-03-023`
- Type: `invention_desk`
- Classification: `editorial`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-023/openbikesensor

**Dek:** Combines a DIY bicycle distance sensor, GPS, and a shared portal so volunteer riders can map close passes and study where street design needs attention.

Combines a DIY bicycle distance sensor, GPS, and a shared portal so volunteer riders can map close passes and study where street design needs attention.

### Why it matters {#why-it-matters-mp-2026-09-03-023}

An independent builder is turning an improbable idea into a working project.

### Limits and context {#limitations-mp-2026-09-03-023}

- A Desk Pick is an editorial selection, not a product endorsement.

### Claims and sources {#claims-mp-2026-09-03-023}

- This Invention Desk entry makes no independently sourced news claim.

## 26. The First Paid Slot {#mp-2026-09-03-024}

- Story ID: `mp-2026-09-03-024`
- Type: `invention_desk`
- Classification: `house_example`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-024/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-2026-09-03-024}

This placement explains how builders can appear in The Invention Desk without purchasing editorial endorsement.

### Limits and context {#limitations-mp-2026-09-03-024}

- House example - no advertiser paid. Payment will buy placement, never endorsement.

### Claims and sources {#claims-mp-2026-09-03-024}

- This Invention Desk entry makes no independently sourced news claim.

## 27. Put Your Project on the Desk {#mp-2026-09-03-025}

- Story ID: `mp-2026-09-03-025`
- Type: `invention_desk`
- Classification: `house_example`
- Content status: `carried_over`
- Permanent URL: https://themachinepress.com/story/mp-2026-09-03-025/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-2026-09-03-025}

This placement explains how builders can appear in The Invention Desk without purchasing editorial endorsement.

### Limits and context {#limitations-mp-2026-09-03-025}

- Manual intake only. Payment buys placement, never endorsement, and every submission is reviewed.

### Claims and sources {#claims-mp-2026-09-03-025}

- This Invention Desk entry makes no independently sourced news claim.

## Normalized sources

- **source-2026-09-03-001:** [arXiv preprint 2609.02811](https://arxiv.org/abs/2609.02811) — arXiv; primary_research
- **source-2026-09-03-002:** [arXiv preprint 2609.02886](https://arxiv.org/abs/2609.02886) — arXiv; primary_research
- **source-2026-09-03-003:** [arXiv preprint 2609.02885](https://arxiv.org/abs/2609.02885) — arXiv; primary_research
- **source-2026-09-03-004:** [arXiv preprint 2609.02852](https://arxiv.org/abs/2609.02852) — arXiv; primary_research
- **source-2026-09-03-005:** [arXiv preprint 2609.02846](https://arxiv.org/abs/2609.02846) — arXiv; primary_research
- **source-2026-09-03-006:** [arXiv preprint 2609.02766](https://arxiv.org/abs/2609.02766) — arXiv; primary_research
- **source-2026-09-03-007:** [arXiv preprint 2609.02684](https://arxiv.org/abs/2609.02684) — arXiv; primary_research
- **source-2026-09-03-008:** [arXiv preprint 2609.02652](https://arxiv.org/abs/2609.02652) — arXiv; primary_research
- **source-2026-09-03-009:** [arXiv preprint 2609.02830](https://arxiv.org/abs/2609.02830) — arXiv; primary_research
- **source-2026-09-03-010:** [arXiv preprint 2609.02653](https://arxiv.org/abs/2609.02653) — arXiv; primary_research
- **source-2026-09-03-011:** [arXiv preprint 2609.02646](https://arxiv.org/abs/2609.02646) — arXiv; primary_research
- **source-2026-09-03-012:** [arXiv preprint 2609.02546](https://arxiv.org/abs/2609.02546) — arXiv; primary_research
- **source-2026-09-03-013:** [arXiv preprint 2609.02542](https://arxiv.org/abs/2609.02542) — arXiv; primary_research
- **source-2026-09-03-014:** [arXiv preprint 2609.02402](https://arxiv.org/abs/2609.02402) — arXiv; primary_research
- **source-2026-09-03-015:** [arXiv preprint 2609.02864](https://arxiv.org/abs/2609.02864) — arXiv; primary_research
- **source-2026-09-03-016:** [arXiv preprint 2609.02887](https://arxiv.org/abs/2609.02887) — arXiv; primary_research
- **source-2026-09-03-017:** [arXiv preprint 2609.02861](https://arxiv.org/abs/2609.02861) — arXiv; primary_research
- **source-2026-09-03-018:** [arXiv preprint 2609.02786](https://arxiv.org/abs/2609.02786) — arXiv; primary_research
- **source-2026-09-03-019:** [arXiv preprint 2609.02749](https://arxiv.org/abs/2609.02749) — arXiv; primary_research
- **source-2026-09-03-020:** [arXiv preprint 2609.02246](https://arxiv.org/abs/2609.02246) — arXiv; primary_research
- **source-2026-09-03-021:** [arXiv preprint 2609.02029](https://arxiv.org/abs/2609.02029) — arXiv; primary_research