Critical-Event Delivery Under Bounded Communication Budgets
Event Importance · Link-Aware Redundancy · Equal-Budget Evaluation
Overview · Method · Status · Related System · 简体中文
EventGuard-LoRa studies critical-event delivery versus communication cost on constrained Sub-GHz links. Its frozen v1 policy combines event importance with link-aware redundancy and is evaluated in host simulation and on two ESP32-S3 boards with E220-400T22D radios. The v1 hardware evidence is a post-hoc balanced set of 96 completed runs; the earlier 300-run v0 pilot is a separate archive.
At an exact DATA-copy budget, with the same trace and channel calendar within each paired seed, does importance- and link-aware redundancy deliver more critical events than budget-matched policies?
Event importance combines normalized change, rate, baseline deviation, co-change, persistence, and a directional hazard rule. NORMAL, IMPORTANT, and CRITICAL map to 1, 2, and 3 copies on a good link. Link state uses a sliding window of first-copy accepted ACK outcomes and consecutive first-copy failures to avoid strategy-dependent observation bias. Copies are bounded, and physical attempts/ACK outcomes are recorded separately. See the frozen algorithm and configuration.
Python reference trace ──serial control──> Sensor ESP32-S3 ──E220──> Gateway ESP32-S3
│ ▲ │
└── shared run manifest └────── ACK ────────┘
host parser / metrics / plots / report
The Sensor classifies and transmits compact CRC-protected frames; the Gateway validates, deduplicates logical samples, records copies, and ACKs. Benchmark traces use a strict 10 s interval. Optional REAL_SENSOR_MODE is a separate demonstration, not the source of benchmark results.
| Strategy | Purpose |
|---|---|
NO_PROTECTION / FIXED_1 |
One DATA copy |
FIXED_REDUNDANCY, FIXED_2, FIXED_3 |
Uniform copies; configured redundancy defaults to two |
IMPORTANCE_ONLY |
Importance-driven 1/2/3 copies without link state |
LINK_ONLY |
Link-driven 1/2/3 copies without importance |
EVENTGUARD |
Importance and link state; default cap three |
UNIFORM_BUDGET, RANDOM_BUDGET |
Match EventGuard's exact DATA-copy total using round-robin or seeded hash allocation, without event labels or sensor values |
Pairing fixes model, rate, seed, trace, and loss calendar. DATA and ACK calendars are independent. RANDOM_COPY erases keyed copy opportunities; BURST_SAMPLE erases all copies of contiguous samples. BURST_COPY is an additional copy-level burst model. On hardware, configured losses are software drops after reception, not measured RF packet errors.
The set contains four strategies × two models (RANDOM_COPY, BURST_SAMPLE) × two loss rates (20%, 30%) × six seeds (31–36): 96 real-device runs. Each paired seed uses one shared 54-sample trace. This is the earliest contiguous complete subset of an interrupted 160-run Stage 1 campaign, selected by completion and execution order, not outcomes; six seeds were not preregistered.
| Comparison | Main finding |
|---|---|
IMPORTANCE_ONLY vs EVENTGUARD |
Identical critical delivery in every selected pair; EventGuard used 21–33 more DATA copies/run on average |
| Exact-budget blind baselines | EventGuard assigned a larger copy share to critical samples; delivery gains were small and condition-specific |
RANDOM_COPY at 30% |
EventGuard mean critical-delivery gain +0.0256 versus blind baselines: two higher seeds, four ties |
BURST_SAMPLE at 20% / 30% |
All strategies tied; redundancy could not recover erased samples |
IMPORTANCE_ONLY is on the observed critical-delivery/cost Pareto frontier in all four conditions; EventGuard is on none. These results do not establish a general benefit from the added link-state mechanism. With six paired seeds, comparisons are directional; descriptive Wilcoxon and exploratory Holm-adjusted p-values do not support confirmatory significance claims.
The offline audit passed 96/96 raw logs/manifests, with no uncontrolled physical DATA/ACK misses or host/firmware delivery differences in the selected set. The later UNIFORM_BUDGET / RANDOM_COPY / 30% / seed37 run103 stalled after a DATA frame was absent from Gateway pre-injection logs. Its cause remains unconfirmed; incomplete seed-37 attempts are excluded and their raw evidence is preserved.
See the hardware report, audit, paired tests, rendered byte-cost Pareto plot, paper draft, and research summary.
The locked pre-hardware matrix contains 8,000 runs (eight strategies, five rates, two models, 100 evaluation seeds). Seeds 1–30 are reserved for calibration and 31–130 for evaluation. C/Python parity checks cover policy and loss/budget decisions.
CRITICAL classifier precision/recall/F1 are 0.8125 / 1.0000 / 0.8966, with NORMAL→CRITICAL false rate 0 on the simple synthetic trace. IMPORTANCE_ONLY matches EventGuard's critical delivery in every paired host run with no more DATA copies; under BURST_SAMPLE, copies do not improve delivery. Expanded analysis finds EventGuard on 0 of 10 within-condition critical-delivery/byte Pareto frontiers.
Reports and retained figures: validation, per-run data, matched-budget plot, ablation, paired tests, Pareto frontiers, sensitivity, failure cases, and research status. Sensitivity varies critical threshold 2.0–4.0, window 8/12/16/24, and cap 2/3; cap four is unsupported by v1's calendar.
Two ESP32-S3/E220-400T22D devices run ESP-IDF v5.4.4 with native UART drivers. The Sensor configuration includes SHT30, BH1750, capacitive soil sensor, and optional OLED; the Gateway uses ESP32-S3/E220. The recorded radio profile uses UART1 9600 baud, TX/RX/AUX GPIO 17/16/15, M0/M1 GPIO 13/14, address 0x0000, REG0 0x62, and channel 0x17. Firmware checks settings at boot without rewriting them.
Real UART and radio frames were exercised. estimated_communication_time_ms is a UART-time proxy, (DATA bytes + ACK bytes) × 10 / 9600 seconds, not RF PHY airtime. Energy in joules was not measured. The unresolved run103 receive-path stall remains a hardware limitation.
Host-only analysis requires Python dependencies and no device:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python tools/run_all_experiments.py --simulate
python tools/run_research_analysis.pyThe second command adds 11,000 host-only sensitivity runs. To regenerate final hardware summaries and plots from preserved logs without contacting hardware:
python tools/finalize_hardware_dataset.pyOptional checks: python -m unittest discover -s tests -v. Physical reruns use tools/run_all_experiments.py and require both boards; preserved run evidence and the hardware report document the campaign.
Smart-Agriculture-Edge-AI draws on EventGuard-LoRa's communication-policy and evaluation methods. The complete frozen redundancy policy is not directly integrated; this remains an independent link-policy study.
One device pair, short synthetic traces, software-injected erasures, and six hardware seeds per condition limit generalization. Controlled RF fading/interference, range, physical channel error rate, energy, regulatory airtime, and field classification remain untested. Real RF experiments are the next priority; a fourth copy would require a new policy/calendar, and the v0 pilot cannot be pooled with v1.
Expand the separate 300-run pilot and its retained figures
The original pilot used three strategies, five configured loss rates, two loss models, ten seeds, and an old 18-sample trace with duplicate timestamps (8 NORMAL, 6 IMPORTANT, 4 CRITICAL). It does not validate frozen v1.
| Strategy | Critical delivery mean | Overall delivery mean | Mean bytes/run | Mean latency (ms) |
|---|---|---|---|---|
| No protection | 0.800 | 0.837 | 664 | 415 |
| Fixed redundancy (2 copies) | 0.927 | 0.929 | 1,340 | 492 |
| EventGuard (up to 3 copies) | 0.968 | 0.945 | 1,579 | 544 |
EventGuard's mean critical delivery was highest at about 17.8% more bytes than fixed redundancy. Its historical “similar cost” comparison can mix loss rates and is not equal-budget evidence. Application-layer injected loss is separate from RF error rate. Pilot firmware used 16,383/16,384 available IRAM bytes, leaving a tight constraint.
Original records: report, CSV, JSON, firmware sizes, and paper outline. Rendered figures remain linked: critical delivery, overall delivery, delivery vs bytes, and latency.
This repository is publicly available for research review, portfolio demonstration, reproducibility inspection, and academic citation.
It is not released under an open-source license.
Copyright © 2026 Sver0411. All rights reserved.
Reuse, modification, redistribution, republication, or commercial use of substantial project code, firmware, datasets, figures, or documentation requires prior permission from the author.
See NOTICE.md for details.
If you reference EventGuard-LoRa in academic or technical work, please use the
repository citation metadata in CITATION.cff.
The frozen research release is v1.0.0.