Ambient Multimodal Cognitive Adjutants: Low-Latency Audio-Visual Memory Streams, Proactive Intent Anticipation and Real-Time Decision Support: chapter answers
Four popular beliefs surround personal AI adjutants: speak up early and often; record everything; the better the prediction, the more the adjutant should act ahead of the person; and confirm every risky action with a dialog. This book falsifies each in turn and rewrites it as a proposition a ledger can overturn. The net value of proactive prompts is bounded by a hit-rate threshold: 0.417 for a definition hint on a chart desk but only 0.073 for a gate red-line reminder. An overconfident adjutant that prompts every time loses 4,610 seconds a week, while a calibrated one that waits for task boun…
DOI pending · monograph page goes live after publication · 5DL-B-409 · 中文版
- Chapter 1 · The Hit-Rate Threshold: When Speaking Up Pays
Section A has four rows, one for each kind of prompt the book returns to. The follow-up hint on the chart desk, where the user has asked about chart structure and the adjutant wants to add a definition, is worth only 20… - Chapter 2 · Voice Activity Detection: the Endpoint Wait as a Latency Floor
The track holds 498,384 samples, cut into 3,114 frames. The noise floor is estimated at -32.7 dB. There are 1,442 background frames, sitting at the floor (-0.2 dB) with a zero-crossing rate of 0.143. The 1,073 voiced fr… - Chapter 3 · End-to-End Latency Budgets: Percentiles, Not Slogans
Section A is the itemised bill for the baseline architecture. The largest item is generation of the first token, with a median of 379 milliseconds, a 95th percentile of 797 and a 99th percentile of 1,073; it also has th… - Chapter 4 · Speculative Prefetch: Read-Only, Break-Even and Server Load
Without prefetch the median is 967 milliseconds, the 95th percentile 1,416 and the 99th percentile 1,697, with 0.564 of requests finished within one second. At a hit rate of 0.30 the median falls to 891 but the 95th per… - Chapter 5 · The Rolling Memory Window: Forgetting Curves, Bystander Exposure and One-Tap Deletion
With no price on bystanders ($k = 0$) the optimal window remains 50 minutes and holds 12.50 bystander-minutes at any moment; with on-device redaction the same 50-minute window holds only 1.10. At a price of 0.50 the opt… - Chapter 6 · Risk-Tiered Confirmation: Fatigue and the Undo Window
Holding for 5 seconds catches almost nothing (0.014) and has a negative net value. Ten seconds catch 0.062, 30 seconds 0.290, 60 seconds 0.493, 120 seconds 0.648 and 300 seconds 0.734, but by five minutes the cost of de… - Chapter 7 · Seeing the First Sign and the Palaces of Memory: an East-West Source Critique
If only 50 minutes are kept, there are 300 segments: an unindexed search inspects 150 on average, a time index 101, and a trail of curated notes 92. Keep one day, 5,760 segments, and the unindexed search inspects 2,880,… - Chapter 8 · The Adjutant Event Ledger: Field Tools and Checker Specifications
Chapter 8 gives the event ledger, the Prompt Threshold and Prefetch Latency Budgeter, a checker table grouped by the four kits and a pre-registered test needing 1,110 events per arm.
Illustrative model calculation, not a live quote