Engineering playground

Paper result. Adapted from “Local Multimodal Retrieval-Augmented Question Answering over Construction Drawing Sets”, Table 3; Section 4.2; Figure 11. Triple fusion reaches 65.8% Precision@1 over all 407 questions. An exploratory selective setting answers 241/407 questions (59.2% coverage) at 80.1% Precision@1. Its threshold was chosen on the same evaluation questions, so it is not a held-out deployment result. The two points are drawn separately; no curve is drawn between them.

Synthetic simulation. Seed 11 gives 64 synthetic queries from a fixed linear congruential generator. Scores sit on 21 levels from 0 to 1, so ties occur. Correctness is drawn independently of score and is not calibrated to the paper. Coverage counts accepted queries over all queries. P@1 counts correct first results over accepted queries. Every query that ties at τ is accepted together. With no accepted query, P@1 and risk are “Not defined”.

Modes. Switching mode resets τ to 0.50, the midpoint of the synthetic score grid. The paper's reported threshold is not applied, because its score definition and scale have not been verified against an export. Empirical mode stays disabled until approved per-query confidence scores, correctness and anonymized keys, or the complete author threshold sweep, are supplied.

This is a browser data explorer. It performs no retrieval and no model inference.

Paper result. Adapted from “Trustworthy validation of industrial energy assessment reports: a neuro-symbolic validator and seeded-error benchmark”, Abstract; Tables 4–6; Figure 1. Mutant-level exact-match on the held-out seeded-error evaluation: Rules plus calculator 96.2%; Hybrid 8B 43.0%; Direct LLM 8B 0.1%. Sample: 81 source reports; 3,249 seeded errors; excludes 27 calibration reports. Exact-match means the expected set of findings for a mutated report, not general document accuracy. These rates are not probabilities for the rows in this lab.

Synthetic simulation. Five authored rows. Expected amount = power (W converted to kW) × hours (h or h/year) × rate (USD/kWh), rounded to cents and compared with the claim in USD or USD/year. For a savings row, power is the load reduction. A row reports, in this order: missing operand (an empty field), incompatible unit, unresolved claim link, arithmetic mismatch, or passed. Reset restores the same rows.

Simulated checker output. Direct and hybrid findings are authored examples written for the reset rows. No model is called, no error is missed at random to imitate a rate, and edited rows are not re-evaluated by them.

Paper result. Adapted from “Skill Classification and Behavioral Analysis of Excavator Operators Using Wearable Sensors”, Table 4; Section 4.2.1; Figure 6; Table 5. Mean simultaneous joystick movement as a share of active time: experienced 43.2%, inexperienced 20.4%. Sample: usable imu trials, 19 experienced, 17 inexperienced; wearable classification and coordination branch. These are reported group means. They cannot reconstruct anyone's movements. “Experienced” and “inexperienced” are the paper's group names; the plan's expert and novice labels map to them.

Synthetic simulation. Each schedule has 1,000 ticks of 20 ms (20 s). 500 ticks have at least one stick active; 216 (experienced) or 102 (inexperienced) have both. A fixed permutation spreads them out. Simulated overlap = both-active time ÷ either-active time, so the schedules give 43.2% and 20.4% by construction, chosen to illustrate the reported shares. Loops are drawn from the masks with an exponential moving average; the default smoothing is 0.3. These are not participant traces. No observed-trace mode exists without approved, participant-safe traces.

Motion. Playback uses elapsed frame time. It stops while the tab is hidden or the canvas is offscreen, and on navigation. With reduced motion, the settled paths and full lanes show without looping.

Synthetic drawing. The 256 × 192 px fragment (walls, a door swing, a column grid, a stair core and a hatched section) is drawn in code from fixed shapes. It is not taken from any client, project or paper drawing. Nothing is uploaded.

Dithering. Standard Bayer rank matrices. Contrast is applied about mid-gray, threshold shifts brightness (0.50 is neutral), and a pixel takes the light tone when its adjusted brightness is at least (rank + 0.5) ÷ matrix area. The output uses exactly the two palette tones.

Illustrative patch grid. The grid splits the image into whole-pixel rectangles that cover every pixel once. It shows image partitioning only. It is not a reproduction of ColQwen2 or any model's preprocessing, and it shows no attention, embedding or retrieval score.