Prompt injection in underground mining ventilation AI

Four adversarial injection surfaces in underground mining ventilation AI

1. Methane concentration display AI (MSHA 30 CFR Part 75.323 — Honeywell BW Technologies methane AI, MSA Safety Ultima XI methane AI, Strata Worldwide WirelessMesh CH₄ AI)

Methane (CH₄) is the primary explosive hazard in underground coal mines and in gassy metal/nonmetal mines hosting permeable hydrocarbon-bearing strata or coal measure rock sequences. MSHA 30 CFR Part 75.323 establishes a three-level action threshold sequence for methane at the working face in underground coal mines: 1.0% CH₄ by volume triggers a mandatory ventilation investigation and corrective action; 1.5% CH₄ triggers mandatory withdrawal of personnel from the affected face area; and 2.0% CH₄ triggers mandatory withdrawal of all personnel from the mine workings affected by the ventilation split carrying the methane accumulation. Methane's lower explosive limit in air is 5.0% by volume, and its upper explosive limit is 15.0% by volume — the 1.0%/1.5%/2.0% MSHA threshold sequence is therefore set at 20%, 30%, and 40% of the LEL, providing substantial safety margins below the explosive range if the thresholds are correctly monitored and acted upon. Methane is colourless, odourless, and lighter than air (molecular weight 16.04 g/mol; density 0.717 kg/m³ at standard conditions versus air at 1.225 kg/m³), meaning it accumulates preferentially at roof level in heading faces and in roof cavities above conveyor belt drives, return airways, and ventilation stoppings — the locations where AI-connected methane sensors (catalytic bead sensors, electrochemical cells, or infrared absorption methane detectors manufactured by Honeywell BW Technologies, MSA Safety, Oldham, or RKI Instruments) are typically mounted. AI systems deployed by Strata Worldwide (WirelessMesh CH₄ AI), MSA Safety (Ultima XI Gas Detection AI), and Honeywell BW Technologies process rendered images of methane concentration display panels — digital readout panels with real-time CH₄ bar indicators showing ppm or percentage vol against 1.0%/1.5%/2.0% threshold reference lines, as displayed on ventilation control room workstations — to automate ventilation adjustment commands (increasing fan speed, closing auxiliary ventilation regulators, redirecting booster fan flows) and to drive personnel withdrawal advisories when methane at the face approaches or exceeds MSHA action levels. The Strata Worldwide VentSim AI ventilation simulation and optimisation platform additionally processes rendered methane contour maps — 3D mine ventilation model renders showing CH₄ concentration isocontours throughout the mine network — to classify ventilation network state and recommend fan speed and regulator adjustments in real time, with the rendered 3D display images processed by the AI classification layer at each inference cycle.

An adversarial perturbation applied to the rendered methane concentration display image — specifically, a ±8 DN downward shift in the pixel encoding of the CH₄ bar indicator region at the point where the bar top approaches or crosses the 1.5% threshold reference line (reducing the apparent bar height from the 1.2–1.8% zone, above the 1.0% warning level and approaching or exceeding the 1.5% withdrawal trigger, to display as sub-1.0% safe working condition) — causes the mining ventilation AI to classify a developing methane accumulation at the working face as a below-action-level normal condition, suppressing the ventilation adjustment command and the personnel withdrawal advisory. With no ventilation response commanded by the AI system and no withdrawal advisory triggering manual operator intervention, methane continues to accumulate in the heading face geometry: the high-CH₄ air layer thickening at roof level, the concentration rising through the 2.0% MSHA mandatory-withdrawal threshold, and continuing toward the 5.0% LEL. Ignition sources in active coal mine headings are numerous: roof fall impacts generating frictional sparks from sandstone-on-sandstone contact; electrical arcing from damaged cable insulation in the face electrical supply system; frictional ignition from drill steel-on-roof contact during rock drilling operations; and hot-surface ignition from diesel equipment exhausts or hydraulic fluid leaks onto hot surfaces. The Sago Mine explosion (January 2, 2006, Upshur County, West Virginia) involved methane accumulated in a sealed area ignited by a lightning strike transmitted through the mine electrical system — 13 miners were trapped in the Sago Mine No. 2 shaft; 12 died from CO poisoning after the explosion and subsequent fire; 1 survivor was rescued after 41 hours. MSHA's investigation identified inadequate atmospheric monitoring of sealed areas as contributing to delayed rescue operations — sealed area atmospheric composition was unknown for the first critical hours after the explosion because the mine's atmospheric monitoring system did not provide real-time data from sealed areas. In an AI-assisted ventilation monitoring system, adversarial suppression of the methane concentration display creates the algorithmic equivalent of the Sago Mine sealed-area monitoring gap: the AI classifies a developing methane accumulation as a normal working condition, replacing the manual monitoring deficiency with an algorithmic classification failure that produces the same delayed-response outcome. The Upper Big Branch Mine explosion (April 5, 2010, Raleigh County, West Virginia, 29 miners killed) provides a direct precedent for intentional methane monitor manipulation: MSHA's investigation found evidence that methane monitors had been tampered with and that ventilation controls were inadequate — adversarial AI injection in a modern VentSim AI or MSA Safety AI pipeline is the digitalised version of the physical sensor manipulation MSHA identified at Upper Big Branch, operating at the rendering layer rather than at the physical sensor hardware. MSHA 30 CFR Part 75.323 specifies the methane concentration thresholds and the withdrawal requirements they trigger but contains no specification for adversarial robustness of AI classification systems processing rendered methane display images, leaving this attack surface unaddressed in the current regulatory framework.

2. CO fire precursor monitor display AI (MSHA 30 CFR Part 75.323 CO monitor AI — Mine Site Technologies CO AI, RAE Systems MultiRAE CO AI, Siemens mining SCADA spontaneous combustion AI)

Carbon monoxide (CO) is the primary atmospheric indicator of underground mine fire and coal seam spontaneous combustion — the process by which exposed coal surfaces undergo exothermic oxidation at ambient temperatures, self-heating over days to weeks through progressive oxidation stages (chemisorption, peroxide formation, and eventual thermal runaway) before reaching the critical temperature at which sustained combustion begins. Spontaneous combustion represents a particularly insidious hazard in underground coal mining because it initiates in inaccessible goaf areas — the mined-out voids behind longwall retreat faces, also termed the gob, which consist of caved roof strata overlying crushed and fragmented residual coal that cannot be fully mined under the economics of current longwall technology. CO is produced by coal oxidation at temperatures as low as 30–40°C above ambient — well before any temperature sensor placed at accessible mine workings would detect a thermal anomaly — making CO trend monitoring in the return air of active longwall panels and gate road development entries the primary early warning system for spontaneous combustion onset. MSHA 30 CFR Part 75.323 requires CO monitoring in underground coal mines and specifies action levels based on CO concentration above the atmospheric baseline: concentrations reaching 10 ppm above baseline trigger a mandatory investigation for fire; concentrations rising above 25 ppm above baseline require an immediate fire investigation with documentation; evacuation decisions are made on the basis of site-specific CO trend analysis establishing whether the trend is consistent with spontaneous combustion progression or with diesel equipment exhaust contributions. AI systems deployed by Mine Site Technologies (MineARC AI CO monitoring module), RAE Systems (MultiRAE CO AI), and Siemens mining SCADA platforms process rendered CO trend display panels — line graphs displaying CO concentration in ppm over 24–72 hour rolling windows, plotted against the CO baseline trend established over the preceding weeks and against the 10 ppm and 25 ppm MSHA action level reference lines, as rendered on ventilation control room displays — to classify CO trend slope, rate-of-rise, and deviation from baseline in the return air of active working sections. The classification output — normal background variation, early spontaneous combustion onset (CO trending upward with slope above baseline), developing spontaneous combustion (CO above 10 ppm action level with rising slope), and emergency fire response (CO above 25 ppm with steep rising trend) — drives investigation dispatch, section isolation decisions, and mine evacuation advisories.

An adversarial perturbation applied to the rendered CO trend display image — specifically, a ±8 DN downward shift in the pixel encoding of the CO trend line slope in the region above the 10 ppm action level reference line (reducing the apparent CO trend line from its actual position at 15–30 ppm with a rising slope to display as below-10-ppm baseline variation with a flat slope) — causes the CO monitoring AI to classify a developing spontaneous combustion event as normal atmospheric background variation, suppressing the investigation dispatch and section isolation advisory. With no investigation triggered by the AI classification and no operator alert compelling manual CO sample collection from the suspected goaf area, the spontaneous combustion process continues uninterrupted in the inaccessible goaf: coal oxidation rates increase with temperature in an accelerating exothermic feedback loop (an approximately 2× increase in oxidation rate per 10°C temperature increase, per the Arrhenius relationship for heterogeneous coal oxidation kinetics), CO concentration rises toward 50–100 ppm in the affected panel's return air, thermal runaway initiates open burning in the goaf void, and mine smoke and CO inundation of the adjacent working headings forces emergency evacuation or traps miners unable to reach self-rescuers in time. The Crandall Canyon Mine collapse (August 6, 2007, Emery County, Utah) involved a coal pillar burst in a mine with a documented history of spontaneous combustion events in the same pillar structure — 6 miners were killed in the initial burst, and 3 rescue workers were subsequently killed during rescue operations when a secondary collapse occurred; MSHA identified CO monitoring deficiencies and inadequate spontaneous combustion management as contributing factors in the mine's operational history leading up to the catastrophic pillar failure. The Westray Mine explosion (May 9, 1992, Stellarton, Pictou County, Nova Scotia, Canada) killed 26 miners in a methane and coal dust explosion; the Richard Commission public inquiry found that methane monitoring was systematically inadequate, that supervisors were aware of methane accumulations approaching dangerous levels, and that the monitoring data that was available was not acted upon — an outcome structurally identical to the scenario produced by adversarial AI suppression of the CO trend display in a modern mine ventilation AI system, where available monitoring data is rendered invisible to the AI classification layer by a pixel-level perturbation. RAE Systems MultiRAE AI instruments, deployed at strategic locations in underground mine ventilation networks, generate both local alarms and telemetric data feeds to central ventilation control rooms; the rendered display images from MultiRAE AI network panels that aggregate multi-point CO readings across a mine section are a primary attack surface for adversarial injection in coal mine spontaneous combustion monitoring AI pipelines.

3. Strata displacement / roof deflection sensor display AI (MSHA 30 CFR Part 57.3461 — Strata Worldwide SmartRoof AI, SRK Consulting strata monitoring AI, Geokon MPBX display AI)

Underground mine roof falls — the sudden downward displacement of rock or coal strata from the mine roof into the mine working — are the leading cause of individual fatalities in metal/nonmetal underground mining (MSHA statistics consistently rank roof and rib falls as the single largest cause of mining fatalities across the US underground mining sector). Strata monitoring instruments provide continuous measurement of roof-to-floor convergence, layer separation within the roof strata, and displacement rates that indicate developing roof deterioration: roof extensometers (mechanical dial gauge or electronic displacement transducer instruments anchored at multiple depths within a roof borehole, measuring the convergence of roof rock toward the mine floor); tell-tales (visual rod-in-tube indicators of roof layer separation — the inner rod is anchored in competent roof strata above the immediate roof while the outer tube is anchored in the immediate roof, and relative displacement of rod against tube provides a visual indicator of whether the immediate roof is separating from the competent strata above); surface extensometers (wire-anchor or rod-anchor instruments measuring total roof-to-floor convergence across the full height of the mine opening); and MPBX instruments (multipoint borehole extensometers measuring displacement at 4–6 anchor depths within a single borehole, providing a displacement profile through the roof strata that reveals which geological horizon is the primary source of displacement). Typical strata monitoring action thresholds in coal roadway and metal mine development drive conditions: a displacement rate exceeding 2 mm/shift requires installation of enhanced supplementary roof support (additional resin-grouted rock bolts, cable bolts, or steel straps) before continued working under the affected area; a displacement rate exceeding 5 mm/shift or a cumulative displacement exceeding 50 mm requires mandatory withdrawal of personnel from the affected area until additional support has been installed and the displacement rate has stabilised. AI systems from Strata Worldwide (SmartRoof AI), SRK Consulting (strata monitoring advisory AI), and Geokon (MPBX display AI) process rendered trend display images from strata monitoring instrument networks — millimetre displacement versus time curves with action threshold reference lines, as rendered on mine geomechanics monitoring workstations — to classify roof convergence state (normal ground movement, approaching action threshold, above action threshold requiring enhanced support, above withdrawal threshold) and generate maintenance and withdrawal advisories for mine management.

An adversarial perturbation applied to the rendered extensometer displacement trend display image — specifically, a ±10 DN suppression of the displacement trend line slope in the region above the 2 mm/shift action threshold reference line (flattening the apparent trend curve from its actual rising trajectory at 2.5–4 mm/shift, above the action threshold requiring enhanced support, to display as within-normal ground movement with a slope below the action threshold) — causes the strata monitoring AI to classify accelerating roof convergence as a normal ground movement pattern, suppressing the enhanced roof support advisory and the operator alert that would trigger a roof survey and supplementary bolt installation. With the AI classification suppressing the maintenance advisory, the mining crew continues production under the affected roof span without the additional support that the actual displacement rate requires: roof convergence continues at the elevated rate, the immediate roof rock mass deteriorates toward the threshold of self-sustaining collapse, and a sudden massive roof fall occurs into the active working — the typical mechanism being rapid propagation of a roof fracture network across the unsupported span, releasing the dead-weight load of the immediate roof beam as a block fall or progressive caving. Crandall Canyon Mine (2007) provides the direct strata monitoring precedent: geotechnical instruments at the mine had recorded indicators of pillar deterioration in the weeks and months preceding the catastrophic pillar burst that killed 6 miners on August 6, 2007 — the strata monitoring data was available but was not acted upon in a timely manner to implement remedial pillar support or section de-pillaring before catastrophic failure. An adversarial AI injection that suppresses the strata monitoring display trend classification produces the same delayed-response failure mode through a different mechanism: in the 2007 scenario, the failure was a human monitoring and decision process failure; in an adversarially compromised AI monitoring system, the failure is an algorithmic classification failure that delivers false-normal assessments to the same human decision process, which then behaves appropriately to the false information it receives. MSHA 30 CFR Part 57.3461 (Ground support installation) requires that adequate roof support be installed before work begins in an underground area — it specifies the requirement for adequate support but does not address adversarial robustness of AI systems classifying rendered strata monitoring trend displays that drive the supplementary support recommendations. The Geokon MPBX digital data acquisition system generates rendered trend display outputs that are processed by strata monitoring AI in mine geomechanics advisory platforms; the rendered MPBX displacement plots, showing relative displacement at each anchor depth versus time with action threshold reference lines, represent a high-consequence adversarial injection surface in metal/nonmetal mine strata management AI applications.

4. Refuge chamber atmospheric monitoring AI (MineARC Systems RefugeAir AI, Strata Worldwide SWS Refuge AI, BioMedical Systems mine refuge AI — refuge chamber O₂/CO₂/CO atmosphere display AI)

Underground mine refuge chambers — sealed survival chambers fabricated from steel, installed permanently in strategic locations throughout underground mine workings, and stocked with compressed oxygen supply, CO₂ scrubbing capacity (calcium hydroxide or lithium hydroxide CO₂ scrubbers), water, food rations, communication systems, and first-aid supplies — are the last line of defence for miners who cannot escape to surface following an explosion, fire, or other emergency that blocks the primary and secondary egress routes. A standard hard-rock mine refuge chamber design (compliant with MSHA 30 CFR Part 75.1506 for underground coal mines or with state mining regulations for metal/nonmetal mines) is rated to maintain a habitable internal atmosphere for a minimum of 96 hours for its rated occupancy — providing sufficient oxygen supply and CO₂ scrubbing capacity to sustain miners awaiting rescue even when the mine atmosphere outside the chamber is lethally contaminated. The atmospheric parameters that must be continuously monitored and maintained within a mine refuge chamber are: internal O₂ concentration — minimum 18.5% vol for safe respiration; concentrations below 19.5% vol are classified as oxygen-deficient by OSHA; below 16% vol, cognitive impairment occurs, impairing decision-making capacity; below 10% vol, rapid unconsciousness follows; the target maintained range in a well-functioning refuge chamber is 20–22% vol O₂; CO₂ concentration — maximum 1.0% vol during normal refuge chamber occupancy; CO₂ accumulates from miner respiration at approximately 0.2 litres/min per occupant, and CO₂ scrubber consumption determines the limit of habitable duration; above 3.0% vol CO₂ headache and dizziness develop; above 5.0% vol rapid unconsciousness occurs; CO concentration — maximum 35 ppm for extended occupancy (NIOSH IDLH for CO is 1,200 ppm; CO above 200 ppm constitutes a 2-hour IDLH exposure); CO can infiltrate a refuge chamber from the external mine atmosphere through imperfect door seals or through pressurisation systems that draw air from CO-contaminated mine atmosphere during positive-pressure maintenance operations. AI systems from MineARC Systems (RefugeAir AI atmospheric monitoring module), Strata Worldwide (SWS Refuge AI), and BioMedical Systems process rendered images of refuge chamber atmospheric monitoring display panels — colour-coded multi-gas readout displays showing O₂, CO₂, and CO concentration bars or digital readouts against threshold lines, as rendered on mine rescue coordination workstations and on the refuge chamber external display panel — to classify refuge chamber internal atmosphere habitability: habitable (O₂ above 18.5%, CO₂ below 1.0%, CO below 35 ppm; rescue team safe to approach and open chamber without SCBA), marginal (O₂ 17–18.5% or CO₂ 1.0–3.0%; rescue team should approach with SCBA precautionary), and non-habitable (O₂ below 16% or CO₂ above 5% or CO above 200 ppm; chamber occupants at risk of incapacitation; rescue team must have SCBA; emergency atmosphere refresh required before safe exit).

An adversarial perturbation applied to the rendered refuge chamber atmospheric monitoring display image — specifically, a ±8 DN upward shift in the pixel encoding of the O₂ concentration bar (raising the apparent O₂ bar from its actual position in the 16–17% range, at or approaching the cognitive impairment threshold, to display as 19–20% — well within the safe working range) — causes the refuge chamber atmospheric AI to classify a depleted chamber atmosphere as fully habitable, suppressing the SCBA advisory for the approaching rescue team. When the rescue team, advised by the AI classification that the internal atmosphere is safe, opens the chamber door without SCBA, they encounter an atmosphere at 16–17% O₂ — below the OSHA oxygen-deficient threshold of 19.5% and approaching the 16% cognitive impairment level. At 16% O₂, rescue team members entering the chamber without SCBA experience impaired judgement, reduced coordination, and accelerating cognitive deterioration that prevents recognition of and response to the actual hypoxic condition; at 15% O₂ and below, disorientation and rapid incapacitation follow, potentially trapping entering rescue team members within the depleted-atmosphere chamber and converting a survivable refuge scenario into a compound fatality event. The Sago Mine explosion (January 2, 2006) provides the direct refuge atmospheric consequence precedent: 13 miners trapped in the mine sheltered in a refuge area where they attempted to use self-rescuer devices with limited oxygen supply; 12 of the 13 miners died from CO poisoning and oxygen depletion over the 41 hours that elapsed before rescue teams located them — the atmospheric monitoring failures in the Sago Mine were both physical (inadequate sealed-area monitoring preventing knowledge of the sealed area atmospheric composition) and procedural (rescue team coordination failures extending the time to locate the survivors); adversarial AI injection targeting the refuge chamber atmospheric display replicates the Sago Mine atmospheric monitoring failure in the specific context of an AI-classified rescue habitability assessment, where a falsely-classified safe atmosphere triggers the compound-fatality mechanism at the moment of rescue team entry. MSHA 30 CFR Part 75.1506 (Refuge alternatives) specifies the habitable atmosphere requirements for mine refuge alternatives in underground coal mines — minimum O₂ concentration, maximum CO₂ concentration, and minimum duration of habitable atmosphere — but does not address adversarial robustness requirements for AI systems classifying rendered refuge chamber atmospheric monitoring display images used by rescue coordination teams to determine safe entry procedures. Australian state mining regulations (New South Wales Mining Act 1992 and associated Mines Rescue requirements; Queensland Coal Mining Safety and Health Act 1999; Western Australia Mines Safety and Inspection Act 1994) impose equivalent refuge chamber atmospheric monitoring requirements for underground coal and metal mines operating within their jurisdictions, with equivalent regulatory gaps regarding adversarial robustness of AI atmospheric monitoring classification systems.

Integration: underground mining ventilation AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for underground mining ventilation AI belongs at every rendered-image ingestion boundary in the mine ventilation AI pipeline — before methane concentration display AI processes rendered CH₄ bar panel images, before CO fire precursor AI processes rendered CO trend graph images, before strata displacement AI processes rendered extensometer trend displays, and before refuge chamber atmospheric AI processes rendered multi-gas readout images used by rescue coordination teams. Threshold 30 for underground mining ventilation AI contexts reflects the MSHA regulatory framework’s life-safety consequence basis (mandatory personnel withdrawal from underground workings at defined atmospheric thresholds; multiple-fatality historical precedent from Sago Mine 2006, Upper Big Branch Mine 2010, and Westray Mine 1992) combined with the absence of any adversarial robustness specification in MSHA 30 CFR Part 75, 30 CFR Part 57, or NIOSH IC 9685 for AI systems classifying rendered atmospheric monitoring displays. The threshold is lower than general industrial AI (35) and equivalent to other high-consequence underground and process safety AI contexts (oil and gas SCADA AI; offshore FPSO compression AI) because a single suppressed methane withdrawal classification in an active coal mine heading can allow methane accumulation to reach the LEL within a single production shift, with ignition sources continuously present.

import asyncio, base64, hashlib, json
from datetime import datetime, timezone
from enum import Enum
from pathlib import Path

import httpx

GLYPHWARD_API_KEY = "YOUR_GLYPHWARD_API_KEY"
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Underground mining ventilation AI: threshold 30
# MSHA 30 CFR Part 75 (underground coal mines); MSHA 30 CFR Part 57 (metal/nonmetal)
# NIOSH IC 9685 mine ventilation guidance
# Sago Mine 2006 (12 killed); Upper Big Branch Mine 2010 (29 killed); Westray Mine 1992 (26 killed)
MINING_VENTILATION_THRESHOLD = 30


class MiningVentilationAIContext(str, Enum):
    METHANE_MONITOR    = "methane_monitor"       # MSHA 30 CFR 75.323: 1.0/1.5/2.0% CH4 thresholds
    CO_FIRE_PRECURSOR  = "co_fire_precursor"     # MSHA 30 CFR 75.323: CO trend spontaneous combustion
    STRATA_DISPLACEMENT = "strata_displacement"  # MSHA 30 CFR 57.3461: extensometer/MPBX display AI
    REFUGE_CHAMBER     = "refuge_chamber"        # MSHA 30 CFR 75.1506: O2/CO2/CO habitability AI


class AdversarialMiningVentilationImageError(Exception):
    """Raised when Glyphward scan score exceeds MINING_VENTILATION_THRESHOLD."""

    def __init__(
        self,
        scan_id: str,
        score: float,
        context: MiningVentilationAIContext,
        mine_id: str,
        section_id: str,
        flagged_region: str | None = None,
    ) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.mine_id = mine_id
        self.section_id = section_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial injection detected in {context.value} image "
            f"(mine={mine_id}, section={section_id}): "
            f"score={score:.1f} > threshold={MINING_VENTILATION_THRESHOLD} "
            f"[scan_id={scan_id}]"
        )


async def scan_mining_ventilation_image(
    image_bytes: bytes,
    context: MiningVentilationAIContext,
    mine_id: str,
    section_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """
    Submit a rendered underground mining ventilation AI display image to the
    Glyphward pre-scan gate and raise AdversarialMiningVentilationImageError if
    the adversarial injection score exceeds MINING_VENTILATION_THRESHOLD (30).

    Call this function before passing any rendered methane concentration display,
    CO trend graph, strata displacement extensometer render, or refuge chamber
    atmospheric monitoring display to the downstream ventilation AI classifier.

    Args:
        image_bytes:  Raw bytes of the rendered monitoring display image (PNG/JPEG).
        context:      MiningVentilationAIContext enum value identifying the
                      monitoring display type and corresponding MSHA regulatory
                      threshold context.
        mine_id:      Mine identifier (e.g. "sago-no2", "ubb-2010", "longwall-3").
        section_id:   Section or heading identifier (e.g. "lw-101-return",
                      "face-ch4-zone-a", "refuge-c7-north").
        client:       Shared httpx.AsyncClient (connection-pooled for performance
                      across multiple concurrent scan calls in the ventilation
                      monitoring pipeline).

    Returns:
        Glyphward scan result dict including scan_id, score, flagged_regions,
        latency_ms, and threshold_applied.

    Raises:
        AdversarialMiningVentilationImageError: Score exceeds threshold (30).
        httpx.HTTPStatusError: Glyphward API returned a non-2xx status code.
        httpx.TimeoutException: Scan did not complete within the configured timeout.
    """
    image_b64 = base64.b64encode(image_bytes).decode()
    image_sha256 = hashlib.sha256(image_bytes).hexdigest()

    payload = {
        "image_b64": image_b64,
        "image_sha256": image_sha256,
        "context": f"underground_mining_ventilation_ai.{context.value}",
        "threshold": MINING_VENTILATION_THRESHOLD,
        "metadata": {
            "mine_id": mine_id,
            "section_id": section_id,
            "msha_regulation": _msha_regulation(context),
            "scanned_at": datetime.now(timezone.utc).isoformat(),
        },
    }

    response = await client.post(
        GLYPHWARD_SCAN_URL,
        json=payload,
        headers={
            "Authorization": f"Bearer {GLYPHWARD_API_KEY}",
            "Content-Type": "application/json",
        },
        timeout=10.0,   # 10-second timeout — ventilation AI pipelines are latency-sensitive
    )
    response.raise_for_status()
    result = response.json()

    score = float(result["score"])
    if score > MINING_VENTILATION_THRESHOLD:
        raise AdversarialMiningVentilationImageError(
            scan_id=result["scan_id"],
            score=score,
            context=context,
            mine_id=mine_id,
            section_id=section_id,
            flagged_region=result.get("flagged_regions", [None])[0],
        )

    return result


def _msha_regulation(context: MiningVentilationAIContext) -> str:
    """Return the primary MSHA regulatory citation for each context."""
    return {
        MiningVentilationAIContext.METHANE_MONITOR:
            "MSHA 30 CFR Part 75.323 — methane concentration action levels",
        MiningVentilationAIContext.CO_FIRE_PRECURSOR:
            "MSHA 30 CFR Part 75.323 — CO monitoring and fire investigation",
        MiningVentilationAIContext.STRATA_DISPLACEMENT:
            "MSHA 30 CFR Part 57.3461 — ground support installation",
        MiningVentilationAIContext.REFUGE_CHAMBER:
            "MSHA 30 CFR Part 75.1506 — refuge alternatives habitable atmosphere",
    }[context]


# ── Example: scan a methane concentration display image before VentSim AI inference ──

async def main() -> None:
    methane_display_png = Path("face_ch4_display_20260619_0612.png").read_bytes()

    async with httpx.AsyncClient() as client:
        try:
            scan_result = await scan_mining_ventilation_image(
                image_bytes=methane_display_png,
                context=MiningVentilationAIContext.METHANE_MONITOR,
                mine_id="longwall-north-3",
                section_id="face-ch4-zone-a",
                client=client,
            )
            print(f"Scan passed — score {scan_result['score']:.1f} "
                  f"(threshold {MINING_VENTILATION_THRESHOLD})")
            # Safe to pass to Strata Worldwide VentSim AI / MSA Safety Ultima XI AI
        except AdversarialMiningVentilationImageError as exc:
            print(f"BLOCKED: {exc}")
            # Suppress ventilation AI classification; generate manual operator alert;
            # log to mine safety management system; notify mine manager and MSHA
            # coordinator; do NOT pass image to VentSim AI inference pipeline.


if __name__ == "__main__":
    asyncio.run(main())

Deploy the Glyphward scan gate as a synchronous barrier in every rendered-image ingestion path in the mine ventilation AI pipeline. For methane concentration display AI (Strata Worldwide VentSim AI, MSA Safety Ultima XI Gas Detection AI, Honeywell BW Technologies mine gas AI), the scan gate should be inserted before each inference call during active production periods — at minimum, before every display render that the AI processes to drive ventilation adjustment commands or personnel withdrawal advisories. For CO fire precursor monitor AI (Mine Site Technologies MineARC AI CO module, RAE Systems MultiRAE CO AI, Siemens mining SCADA spontaneous combustion AI), the scan gate should be inserted before processing 24–72 hour rolling CO trend graph renders that the AI uses to classify spontaneous combustion onset — particularly important during longwall retreat phases when fresh goaf exposure creates peak spontaneous combustion risk in the active panel. For strata displacement AI (Strata Worldwide SmartRoof AI, Geokon MPBX display AI, SRK Consulting strata monitoring AI), the scan gate should be inserted before processing extensometer trend display renders during high-convergence-rate phases identified in previous shift reports. For refuge chamber atmospheric AI (MineARC Systems RefugeAir AI, Strata Worldwide SWS Refuge AI), the scan gate is highest priority during active rescue operations when the habitability classification of the rendered refuge display panel directly determines whether approaching rescue team members equip SCBA — the adversarial consequence is maximised at this specific decision point. See the Glyphward waitlist for enterprise pipeline integration and the free tier for immediate scanning capability without account provisioning. Operator and mine manager responsibilities under MSHA 30 CFR Part 75.323 for methane and CO monitoring remain unchanged by the adoption of AI-assisted ventilation monitoring — the Glyphward scan gate does not satisfy MSHA monitoring obligations but provides an additional pre-inference adversarial robustness layer for AI systems that supplement mandatory MSHA monitoring programmes.

Related questions