Underground coal mine ventilation AI adversarial injection: how ±8 DN in the rendered methane monitor display suppresses a CH₄ reading above the MSHA 30 CFR 75.323 action level — and why MSHA has no adversarial robustness criterion for the sole-barrier methane detection AI

Sago Mine 2006: the methane accumulation timeline and what the AI classification boundary means

Understanding why the Sago Mine disaster is the structural anchor for underground coal mine ventilation AI adversarial injection requires understanding the event not as a sudden catastrophe, but as a sequence of monitoring and classification decisions at defined thresholds that determined whether the developing methane hazard was detected and acted upon before reaching ignition conditions.

Sago Mine produced coal from the Sago No. 2 seam, a gassy coal seam characteristic of the Central Appalachian coal fields. Methane desorbs continuously from the seam face and surrounding strata as mining advances, and the ventilation system — a primary fan at the mine portal with a designed quantity of approximately 14 m³/s — was responsible for diluting methane from active working sections to below the MSHA 30 CFR 75.323 1.0% action level and carrying it to the mine atmosphere. Active working sections were required to maintain continuous methane monitoring at defined sensor locations. The sealed areas — including Sealed Unit 1 — were off-limits to the primary ventilation circuit: they were sealed with permeable and impermeable overcasts to prevent methane migration into the active workings, and were required under 30 CFR 75.336 to be sampled for CH₄, CO, CO₂, and O₂ at defined intervals.

MSHA’s investigation concluded that methane in Sealed Unit 1 had accumulated above the lower explosive limit of 5% by volume. The accumulation was not detected and acted upon — not because the physical requirement for sealed area atmospheric sampling did not exist, but because the sampling frequency and sensor placement had not been sufficient to detect the developing hazard before ignition. On 2 January 2006, a lightning strike at the surface near the mine portal sent a transient electrical arc through an inadequately sealed overcost into Sealed Unit 1. The arc ignited the methane-air mixture, producing an explosion that propagated through the sealed area with a pressure wave sufficient to destroy the overcosts, blow out the stopping curtains separating the sealed area from the active workings, and structurally compromise two of the mine’s primary entries — severing all communications and blocking evacuation routes for the thirteen miners who could not reach the portal before rising CO concentrations forced them to seek refuge.

The critical insight for the AI adversarial injection threat model is this: the monitoring failure at Sago was at the boundary between “within specification” and “above threshold requiring action.” The physical sensors existed. The regulatory requirement to sample existed. The failure was in the classification decision — whether the sampled atmosphere in the sealed area was being detected as above-threshold for methane and acted upon with the interventions the threshold required. An AI system classifying rendered sealed-area tube-bundle sample displays or continuous methane sensor outputs, adversarially manipulated to classify an above-LEL reading as nominal, would produce exactly this monitoring gap — with the critical difference that the manipulation would be invisible to the monitoring system and to any inspection regime that tests the physical sensor (as Part 75.362 requires) without testing the AI classifier.

For the active working sections — where continuous AI-assisted methane monitoring is the current and emerging standard — the structural parallel is at the 1.0% action-level boundary: a ±8 DN adversarial shift normalises an apparent CH₄ reading of 1.2–1.4% to 0.7–0.8%, suppressing the de-energisation trigger and allowing electrical equipment — the dominant ignition source category in MSHA underground mining fatality investigations — to remain energised as the methane concentration climbs toward the LEL.

Westray 1992 and Upper Big Branch 2010: two additional precedents and the human-analogue of adversarial injection

The Sago Mine disaster is not an isolated precedent. Westray Mine (Pictou Coal, Plymouth, Nova Scotia, 9 May 1992) and Upper Big Branch Mine (Massey Energy, Montcoal, West Virginia, 5 April 2010) each provide a distinct documented instance of methane monitoring failure as a proximate cause of a multi-fatality underground coal mine explosion.

At Westray, the Westray Mine Public Inquiry — the Richard Commission, which produced its final report in November 1997 — found that methane concentrations at or above the Canadian threshold of 1.25% had been documented in mine records on multiple occasions before the explosion and had not triggered the required personnel withdrawal and equipment de-energisation. The Commission found that the mine’s management culture systematically prioritised coal production over methane safety compliance: inspectors were not adequately empowered to enforce threshold actions; above-threshold readings documented in pre-shift examination books were not escalated; and the ventilation system was inadequate to maintain methane below threshold in the working sections. The Westray explosion killed 26 miners. In 2003, Canada enacted the Westray Law (Bill C-45, Criminal Code amendments establishing corporate criminal liability for workplace safety failures) specifically in response to the Commission’s findings. The Westray structural parallel for AI adversarial injection is the sustained above-threshold methane classification failure over time: an AI that persistently classifies above-threshold readings as nominal (whether from training distribution mismatch or adversarial manipulation) provides the same organisational cover as the management culture the Commission identified, without requiring active human concealment.

At Upper Big Branch, the MSHA investigation (concluded December 2011) found documented evidence that Massey Energy had maintained a dual-record safety system for methane monitoring: pre-shift examination books showing compliant readings for MSHA inspectors, and production records showing actual above-threshold readings during operation. MSHA investigators found that mine foremen instructed equipment operators to reduce methane monitor sensitivity, that above-threshold readings were erased from examination books before inspectors arrived, and that ventilation controls were configured to produce acceptable readings at inspection points while real-time readings in active sections were higher. This practice — manually manipulating methane gauge readings to conceal above-threshold concentrations — is the human-operated analogue of adversarial pixel injection at the rendered CH₄ monitor display boundary. The adversarial injection attack is technically superior to the Upper Big Branch human manipulation in every respect relevant to concealment: it requires no co-ordination among mine employees (a single adversarial file modification changes the rendered display AI’s input); it operates continuously rather than only during inspection visits; it is undetectable by the standard MSHA calibration testing requirement (which tests the physical sensor against a known gas concentration, not the AI classifier processing the rendered sensor output); and it produces a manipulated record that passes digital scrutiny because the physical sensor reading in the system log is unchanged. The adversarial manipulation occurs at the AI input boundary — the rendered display image — not at the sensor or data logger level.

Together, Sago (12 killed), Westray (26 killed), and Upper Big Branch (29 killed) — 67 miners killed in three documented multi-fatality methane disasters whose root-cause investigations each found methane monitoring classification failure as a proximate or contributing cause — establish the consequence envelope for underground mining ventilation AI adversarial injection as among the most consequential industrial safety AI threat models in the portfolio.

How underground coal mine ventilation AI works — and where the adversarial injection surface lives

Underground coal mines use integrated ventilation monitoring and control systems from vendors including Strata Worldwide (VentSim AI — ventilation network simulation, real-time atmospheric monitoring, and predictive gas hazard modelling), Howden (Ventilation on Demand AI — demand-based ventilation control that adjusts fan speed and regulators based on production activity and gas sensor readings), MSA Safety (fixed gas detection systems — catalytic bead and infrared CH₄ sensors, electrochemical CO sensors, and multi-gas monitors with SCADA integration), Honeywell BW Technologies (area monitoring instruments with rendered display outputs and AI-assisted trend classification), Epiroc Mobilaris (underground positioning and personnel tracking integrated with gas sensor mapping), and MineARC Systems (refuge chamber atmospheric monitoring with CO, CO₂, O₂, and CH₄ display AI).

These systems process real-time gas concentration data from sensor networks deployed throughout the mine — at working faces, at return airways, in sealed area tube-bundle sampling stations, in refuge chambers, and at strategic monitoring points required by the approved mine ventilation plan — and render the measurements into visual displays: digital readout panels showing real-time CH₄ concentration in percent-by-volume, time-series trend charts showing CO concentration over 1-hour, 8-hour, and 24-hour windows, geographic overlays showing methane concentration across the mine layout, and atmospheric monitoring dashboards in refuge chambers showing O₂, CO, and CO₂ in real time.

The adversarial injection surface is at the boundary between each rendered display image and the AI classifier that processes it — the structural pattern present in every underground mining ventilation AI monitoring context: accurate physical sensors measure the safety-critical atmospheric parameter; the measurement is rendered into a 2D visual display for operator monitoring and AI classification; the AI classifier generates automated alerts, action-level triggers, and ventilation control commands; and the AI has never been evaluated for adversarial robustness at its rendered-image ingestion boundary. The physical calibration testing requirement (Part 75.362: methane detectors tested with known gas concentrations every 31 days) verifies the physical sensor, not the AI classifier. The ventilation plan approval process (Part 75.220) specifies sensor locations and action levels, not AI robustness. The gap is structural.

Four adversarial injection surfaces in underground coal mine ventilation AI

1. Methane (CH₄) monitor display AI (Strata Worldwide VentSim AI, MSA Safety ALTAIR 5X AI, Honeywell BW Technologies GasAlertMax XT II AI — working section methane action-level classification AI)

The primary methane hazard in active coal mine working sections is methane desorbing from the freshly cut coal face. Continuous mining machines and longwall shearers cut at rates that can release 0.1–1.5 m³/tonne of methane from the seam, and the ventilation airflow at the working face must dilute this release to below the 1.0% MSHA action level on a continuous basis. Strata Worldwide VentSim AI, integrated with fixed CH₄ sensor networks, classifies the rendered methane concentration display to generate automated action-level alerts: at 1.0% CH₄, the AI triggers de-energisation of electrical mining equipment; at 1.5%, it triggers personnel withdrawal; at 2.0%, it triggers complete section clearance.

An adversarial perturbation on the rendered CH₄ monitor display that suppresses the apparent concentration — applying a ±8 DN downward shift to the pixel region encoding the concentration value, normalising a displayed reading of 1.2–1.4% CH₄ to an apparent reading of 0.7–0.8% — causes the methane detection AI to classify an above-threshold section atmosphere as compliant ventilation, suppressing the equipment de-energisation trigger. Electrical equipment — continuous mining machines, conveyors, pumps, and lighting systems — remains energised in a methane-enriched atmosphere. As the methane concentration continues rising toward the LEL (5%), the probability of ignition from any electrical arc increases. The MSHA statistics from Underground Big Branch and other post-1980 disasters show that the average time from a 1.0% CH₄ reading to a 5% (LEL) condition at a poorly ventilated face can be as short as 8–15 minutes under high gassing rates — within the latency of a single adversarial display manipulation cycle.

The Westray structural parallel is the most direct: the Westray Commission found methane readings at or above the 1.25% threshold that were not acted upon over days and weeks before the explosion. A persistently adversarially suppressed methane display AI replicates this sustained monitoring failure with pixel-level precision, without requiring the organisational culture documented at Westray or the active human concealment documented at Upper Big Branch.

2. Carbon monoxide trend display AI (Strata Worldwide Mine Sensor AI, MSA Safety CO area monitor AI, Detectogen underground CO AI — spontaneous combustion and post-explosion CO trend classification AI)

Carbon monoxide (CO) in underground coal mines has two primary sources: spontaneous combustion (exothermic oxidation of coal that develops in sealed goaf areas, broken coal, or exposed coal ribs over days to weeks) and post-explosion combustion (CO generated by the methane-air explosion itself, which permeates the mine atmosphere). Spontaneous combustion is the more insidious pre-explosion hazard: it generates CO concentrations of 5–100 ppm in the return airway for days or weeks before the heated coal body reaches the temperature at which it will ignite independently or release sufficient additional methane to create an explosive atmosphere. CO trend display AI classifies the 8-hour and 24-hour CO trend at strategic monitoring points — return airways, ventilation control points, and sealed area tube-bundle stations — and generates alerts when the trend indicates rising CO concentration above 25 ppm (the NIOSH recommended exposure limit), above 50 ppm (the MSHA action level for spontaneous combustion investigation), or when the Graham Ratio (CO/O₂ depletion index for spontaneous combustion staging) indicates advancing spontaneous combustion.

An adversarial perturbation on a rendered CO trend display that suppresses a rising trend — applying a ±8 DN shift to the pixel region encoding the trend line slope, normalising a rising trend from 20 ppm → 60 ppm → 140 ppm over 72 hours to a flat trend at 25–30 ppm — causes the CO trend AI to classify a developing spontaneous combustion event as normal mine atmosphere variation. The heating coal body continues developing, the CO output increases, and the surrounding strata begin to desorb additional methane as the coal temperature rises. The Westray Mine parallel: the Commission found evidence that elevated CO readings in the Foord seam had been documented in pre-explosion inspection records and not acted upon. A mine CO trend AI adversarially suppressed to classify elevated readings as normal produces the same non-response without the human active concealment that MSHA inspectors can investigate.

The post-explosion CO monitoring context is equally critical: CO generated by an underground explosion — as at Sago — permeates the mine entries at concentrations of 500–5,000 ppm, with an IDLH of 1,200 ppm and a lethal concentration of approximately 1,600 ppm over 1–3 hours of continuous exposure. If the refuge chamber CO monitoring display AI is adversarially suppressed, the miners sheltering in the refuge do not receive the indication that CO has reached IDLH concentrations — they remain in the refuge without SCBA protection until CO asphyxiation incapacitates them, exactly the fate of the twelve Sago miners.

3. Strata extensometer and roof displacement display AI (Epiroc Mobilaris roof monitoring AI, Trimble MineEdge geotechnical AI, Strata Worldwide TDR AI — geotechnical failure precursor classification AI)

Underground coal mines in thick seam and multiple-seam environments face significant roof fall hazards from inadequate support of the immediate roof strata, delayed failure of overpressured pillars, and stress redistribution as the longwall face advances. Roof and rib falls account for approximately 35–40% of non-explosion, non-transport underground coal mining fatalities in the MSHA database. Strata extensometers — instruments measuring the relative displacement between anchor points at different depths in the roof strata above a mine entry — provide continuous data on roof deformation rate, and their rendered display outputs are classified by AI to determine whether displacement rates exceed geotechnical trigger action response (TAR) levels requiring miners to be withdrawn from unsupported ground.

An adversarial perturbation on a rendered strata extensometer display that suppresses a rising displacement rate — applying a ±10 DN shift to the pixel region encoding the displacement rate indicator, normalising a reading of 8 mm/day (above the TAR Level 2 trigger of 5 mm/day for a roof cut height of 3.5 m) to an apparent reading of 2 mm/day — causes the roof monitoring AI to classify an actively deforming roof section as stable, suppressing the withdrawal of miners from the affected entry. In a coal mine environment where roof strata are under elevated stress from adjacent goaf loading or multiple seam interaction, a displacement rate of 8 mm/day represents an accelerating roof deformation trajectory that typically leads to roof fall within 4–12 hours. The adversarially suppressed AI keeps miners in the entry through this window.

The geotechnical monitoring AI surface is distinct from the methane explosion surfaces in its failure mode — roof fall rather than explosion — but shares the structural pattern: the physical sensor is calibrated and qualified; the AI classifying the rendered display output is the unspecified layer at the regulatory gap boundary.

4. Refuge chamber atmospheric monitoring display AI (MineARC Systems LifeSaver AI, Strata Worldwide refuge monitor AI, MSA Safety refuge atmospheric AI — post-disaster survivor atmosphere management AI)

The MINER Act 2006 — enacted directly in response to Sago — requires underground coal mines to provide refuge alternatives capable of sustaining miners in a survivable atmosphere for at least 96 hours following a mine emergency. MSHA regulations (30 CFR 75.1714-3) specify that refuge alternatives must maintain: O₂ concentration between 18.5% and 23%, CO concentration below 25 ppm, CO₂ below 1%, and CH₄ below 1%. MineARC Systems LifeSaver chambers and equivalent refuge alternatives are equipped with atmospheric monitoring instruments — electrochemical O₂ sensors, electrochemical CO sensors, infrared CO₂ sensors, and catalytic bead CH₄ sensors — with rendered display panels showing real-time concentrations and trend data. AI systems classify these displays to determine whether the refuge atmosphere is within safe limits, whether CO₂ scrubber cartridges need replacement, and whether CO ingress from the mine atmosphere is approaching IDLH concentrations.

An adversarial perturbation on a rendered refuge chamber atmospheric monitoring display that suppresses a deteriorating atmosphere — applying a ±8 DN shift that moves an O₂ reading from 17.8% (below the 18.5% MSHA minimum) to 20.4% (normal), or that suppresses a CO reading from 45 ppm (above 25 ppm limit) to 6 ppm — causes the refuge atmospheric AI to classify an unsafe refuge atmosphere as within normal parameters. Miners in the refuge — who rely on the displayed atmospheric readings to determine whether to remain in the refuge, whether to don SCBA, and whether to evacuate — receive false assurance of atmospheric safety. As O₂ continues to decline from CO₂ accumulation (exhaled breath in a sealed refuge will raise CO₂ above 2% within 4–6 hours without active scrubbing) or as CO infiltrates from the mine atmosphere, the displayed readings no longer reflect the actual hazard. The Sago parallel is the most direct: the twelve miners who died were in a makeshift refuge (a dead-end entry sealed with curtain) with no reliable atmospheric monitoring, unable to determine whether the air they were breathing was safe. A MINER Act refuge chamber equipped with adversarially manipulated AI atmospheric monitoring recreates this uncertainty with the appearance of monitored safety.

See also: underground coal mine ventilation AI prompt injection — CH₄, CO, refuge, and sealed area monitoring context for the technical classification boundary specification at each sensor type.

MSHA 30 CFR Part 75, the MINER Act 2006, and the adversarial robustness gap

The US federal underground coal mine safety regulatory framework is among the most detailed industrial safety specification regimes in the world. MSHA 30 CFR Part 75 comprises 74 subparts specifying requirements for ventilation, roof support, electrical systems, explosives, fire protection, emergency response, and atmospheric monitoring. Part 75.323 specifies the methane action levels. Part 75.362 specifies 31-day calibration testing intervals for methane detectors. Part 75.220 requires an approved mine ventilation plan specifying sensor locations, minimum airflow quantities, and response actions for above-threshold readings. Part 75.336 through Part 75.340 specify sealed area atmosphere sampling requirements. Part 75.1714-3 specifies refuge alternative atmospheric maintenance requirements.

The MINER Act of 2006 (P.L. 109-236, enacted 15 June 2006) substantially expanded these requirements in direct response to Sago and Aracoma. Section 2 required wireless two-way communications for all underground coal mines within 3 years. Section 3 required electronic or other tracking to determine the location of all miners underground within 3 years. Section 4 required mine operators to provide self-rescuers capable of a minimum 1-hour supply and refuge alternatives capable of sustaining a miner for at least 96 hours, with approval from MSHA within 3 years. Section 6 required enhanced mine rescue team availability. These requirements are operationally rigorous and technologically demanding. They do not address adversarial manipulation of AI systems classifying rendered atmospheric monitor outputs.

The post-Upper Big Branch enforcement changes — MSHA’s enhanced Pattern of Violations (POV) enforcement under 30 CFR Part 104 (2013 amendments), which allow MSHA to order withdrawal of all miners from a mine where the operator has demonstrated a pattern of significant and substantial violations — addressed the documented Massey Energy practice of concealing methane readings from inspectors through a heightened inspection regime. But the POV framework operates on the assumption that inspection visits, underground walkarounds, and review of examination books will detect methane monitoring deficiencies. An adversarial AI manipulation operating at the rendered-display boundary — leaving physical sensor readings unchanged in the data logger, affecting only the pixel-level input to the AI classifier — would not be detectable by the physical calibration testing the 31-day cycle requires, the ventilation plan review, or the MSHA inspector walkaround. The regulatory gap for adversarial AI in underground coal mining is not a gap in regulatory intent: the three major underground coal mine disasters of the past 35 years have each generated comprehensive regulatory reforms that have substantially reduced the incidence of methane explosions in US underground coal mining. The gap is that none of these reforms contemplate an AI classifier layer operating on rendered display images — because none of the disasters involved such a layer. The adversarial injection threat model is not a retrospective on how Sago, Westray, or UBB could have been different if AI had been present. It is a prospective threat model for how the monitoring systems that replace the practices documented in those disasters will fail if the AI classification layer is not evaluated for adversarial robustness.

The structural parallel extends to oil refinery APC AI adversarial injection under OSHA PSM 29 CFR 1910.119, where the Texas City BP 2005 ISOM unit raffinate splitter overflow — 15 killed, caused by a misleading level gauge display that operators classified as normal operation — demonstrates the same monitoring classification failure mode in a downstream process context: accurate physical sensors, rendered display output, human or AI classification at the boundary, and a regulatory framework (OSHA PSM) that specifies process hazard analysis but not adversarial robustness testing for the AI layer. The pattern across underground coal mining (MSHA Part 75) and downstream process (OSHA PSM) is the same: the physical monitoring obligation is specified, the action levels are specified, the AI classification layer at the rendered boundary is not.

The methane explosion physics of ignition probability and intervention timing: why the 1.0% action level is the only practical intervention window

The MSHA 30 CFR 75.323 methane action level of 1.0% — a 5x safety margin below the LEL of 5% — exists because the relationship between methane concentration and ignition probability is not a step function but a probabilistic one that depends on concentration uniformity, ignition source energy, and the spatial distribution of the methane plume relative to energised equipment. Understanding the physics establishes why the adversarial suppression of the 1.0% action level trigger — rather than the 1.5% or 2.0% trigger — is the critical target for threshold-30 adversarial scan gating.

In an underground coal mine entry, methane does not mix uniformly with ventilation air. Stratification — the tendency of CH₄, which has a density of 0.68 kg/m³ relative to air at 1.20 kg/m³, to accumulate at the top of the entry near the roof — produces peak concentrations 1.5–3x higher than the average measured concentration at the sensor location (typically mounted at standing height or on the continuous mining machine body). A face-area average CH₄ reading of 1.0% may correspond to roof-level stratification concentrations of 1.5–3.0% — above the 1.5% personnel-withdrawal level and approaching the 5% LEL. The 1.0% action level is calibrated to this stratification factor: de-energising electrical equipment at 1.0% average removes the most likely ignition source category (electrical arcing) before the stratified roof concentration has reached the ignition range.

The suppression of the 1.0% action level by adversarial CH₄ display AI manipulation therefore removes the primary ignition-source-elimination control at the moment when roof-level stratification may already be above the LEL at the location of roof-mounted shearer cable connections and ventilation fan motor housings. The relevant intervention timing: in a high-gassing working section producing 0.5 m³/tonne methane at a cutting rate of 150 tonnes/hour from a continuous miner with 14 m³/s face ventilation, the steady-state CH₄ concentration at the face is approximately 0.9%. A 15-minute ventilation disruption (a stopped auxiliary fan, a ventilation curtain displacement, a regulator failure) raises the face concentration from 0.9% to 1.4–2.1% in 8–12 minutes. If the CH₄ display AI suppresses the reading from 1.4% to 0.8% during this disruption window, the de-energisation that would have removed the ignition source does not occur. The roof-level stratification concentration during this window may be at 2–4% — within the 5–15% explosive range if the disruption extends another 4–6 minutes before an un-adversarially-manipulated sensor in the return airway provides a corrective reading.

The early-intervention window is therefore the 1.0% action-level trigger: de-energise equipment at 1.0%, before stratification reaches the LEL. AI adversarial injection that suppresses this trigger — presenting 0.8% while the actual reading is 1.2% — eliminates the only practical electrical-ignition-removal window in the short disruption scenario.

Glyphward threshold 30 for underground mining ventilation AI

Glyphward’s adversarial detection API operates as a pre-classification gate at each rendered-image ingestion boundary in the underground mine ventilation AI pipeline: before the methane display AI processes each rendered CH₄ monitor output, before the CO trend AI processes each trend display frame, before the strata extensometer AI processes each displacement rate output, and before the refuge chamber atmospheric AI processes each atmospheric monitoring panel. Each rendered image receives a risk score (0–100) in 8–15 ms. At or above threshold 30, Glyphward gates the AI classification and generates an alert that triggers manual verification against the physical sensor output — the direct CH₄ sensor reading at the data logger, the raw CO electrochemical sensor reading, the direct extensometer measurement — none of which can be adversarially manipulated at the pixel level of the rendered display image.

Threshold 30 for underground mining ventilation AI contexts reflects the documented multi-fatality consequence of methane monitoring failure at three major incidents (Sago 12 killed, Westray 26 killed, Upper Big Branch 29 killed), the demonstrated human-analogue of adversarial methane reading manipulation at Upper Big Branch (where MSHA documented active concealment of above-threshold readings), and the structural gap between MSHA Part 75 calibration requirements (which test the physical sensor) and the adversarial robustness of AI classifying rendered display outputs (which is not tested by any current MSHA requirement or industry standard).

The false positive cost at threshold 30 is a manual verification of the physical sensor reading — the same verification that MSHA’s 31-day calibration requirement already makes a routine part of mine atmospheric monitoring practice. The false negative cost is the suppression of the 1.0% action level trigger that removes the primary ignition source at the moment when methane stratification may already be above the LEL — producing the conditions documented at Westray, Sago, and Upper Big Branch. The proportionality is not close.

Free tier — 10 scans/day, no card required. Submit a rendered CH₄ monitor display image or CO trend chart from your mine ventilation management system to the Glyphward scanner to generate a baseline adversarial risk score for your underground mining ventilation AI classification inputs.