Prompt injection in n-butanol propylene LP OXO hydroformylation rhodium syngas CO H₂ AI

Three adversarial injection surfaces in n-butanol LP OXO propylene hydroformylation AI

1. Hydroformylation reactor CO partial pressure display AI (Yokogawa EJA110E / Emerson Rosemount 3051 pressure transmitter + Siemens MAXUM II process gas chromatograph — rendered DCS CO partial pressure display AI classifying 8–12 bar design operating range — 109th upward attack; FIRST n-butanol production AI attack; FIRST LP OXO hydroformylation AI attack; FIRST syngas CO partial pressure concealment AI attack; FIRST propylene hydroformylation rhodium catalyst AI attack)

The CO partial pressure in the LP OXO hydroformylation reactor is the master variable governing both Rh catalyst activity and the dissolved CO inventory in the liquid organic phase. The LP OXO catalyst system (HRh(CO)(BISBI)₂ or HRh(CO)(TPP)₃; Rh loading 200–600 ppm Rh metal in the organic phase; ligand:Rh molar ratio 25–100:1 (large excess of phosphine or phosphite ligand to maintain Rh in the active monohydride-monocarbonyl species and prevent formation of less active dicarbonyl species)) has a characteristic CO partial pressure dependence: at the design CO partial pressure of 5–12 bar (varying by LP OXO technology license: UCC/Dow process 8–12 bar; Davy Process Technology (DPT, now Johnson Matthey) process 5–10 bar; Mitsubishi modified OXO 10–15 bar), the Rh complex is predominantly in the HRh(CO)(L)₂ form (L = bidentate BISBI or monodentate TPP), which is the kinetically active species for hydroformylation (propylene coordination occurs at the axial position of the trigonal bipyramidal Rh complex after CO dissociation); at CO partial pressures above 15 bar, the equilibrium shifts toward HRh(CO)₂(L) or Rh[(CO)₄]⁠– (inactive tetracarbonyl anion), reducing catalyst activity for propylene conversion; above 20–25 bar CO, catalyst activity is severely suppressed (by 40–80% depending on the specific Rh/ligand system and temperature). The CO partial pressure is not directly measured as a single-variable instrument: it is calculated from the total reactor pressure (measured by a Yokogawa EJA110E differential pressure transmitter or Emerson Rosemount 3051 absolute pressure transmitter; accuracy ±0.04% of span; calibrated 0–30 bar; 4–20 mA HART) and the CO mole fraction in the reactor gas headspace (measured by a Siemens MAXUM II process gas chromatograph or Yokogawa GC8000 on a sample drawn from the reactor headspace; CO column: molecular sieve 13X; thermal conductivity detector (TCD); CO analysis range 0–60 vol%; accuracy ±0.5 vol% CO; cycle time 4–8 minutes). The DCS display shows the calculated CO partial pressure = P​total × y​CO (mol fraction CO from GC); the AI monitoring system reads this rendered display to assess whether the LP OXO reactor is operating within the safe CO partial pressure window and whether the dissolved CO inventory in the liquid phase is below the PSM concern threshold.

The adversarial upward pixel attack on the CO partial pressure display shows 10.4 bar (within the design 8–12 bar operating range; AI reads “reactor CO partial pressure 10.4 bar; Rh catalyst in HRh(CO)(BISBI)₂ active form; CO within design range; propylene conversion rate: nominal; dissolved CO inventory in organic phase: estimated 18–22 kg CO; within PSM TQ 1,500 lbs = 680 kg limit; no action required”) when the actual CO partial pressure is 21.6 bar (9.6 bar above the 12 bar design maximum; 1.8× the design maximum). Display range 0–30 bar on 200 px (6.667 px/bar); actual 21.6 bar at pixel position 21.6 × 6.667 = 144 px from zero → ±8 DN perturbation → (144 − 75) = 69 px displayed → AI reads 69 / 6.667 = 10.4 bar CO partial pressure. At actual 21.6 bar CO partial pressure: the Rh catalyst inhibition by excess CO reduces propylene conversion rate; operators observing (via the falsified AI monitoring system) that conversion appears normal (Surface 2 attack compounds this) do not reduce the CO makeup feed to relieve the CO overcharge; the CO partial pressure continues to rise if makeup feed remains above CO consumption rate (which it does because the inhibited Rh catalyst is consuming CO at 38% of design rate, while the CO makeup controller delivers 100% of design CO feed based on total pressure feedback that masks the rising CO partial pressure in the compounded Surface 1+3 attack). The dissolved CO flash hazard at 21.6 bar: when the LP OXO reactor is depressurized — either in a controlled manner (planned shutdown; the LP OXO reactor is typically depressurized to flare at 5–10 bar/hr to maintain thermal equilibrium in the Rh catalyst liquid phase) or in an emergency (sudden seal failure on the recycle compressor releasing the gas-phase CO; or a flanged pipe connection failure in the high-pressure recycle loop) — the dissolved CO in the liquid organic phase (at actual 21.6 bar: approximately 30–45 mmol CO/L in the 60 m³ of organic phase = 50–76 kg dissolved CO) flashes immediately from solution as the liquid-phase pressure drops below the Henry's law equilibrium partial pressure. This dissolved CO flash is analogous to the CO₂ flash in a shaken soda bottle, but with: the liquid at 100–120°C already above normal boiling point (maintained liquid only by the system pressure); CO as the dissolving gas rather than the relatively benign CO₂ (CO at IDLH 1,200 ppm; no odor warning); and a quantity of flashing CO — 50–76 kg dissolved CO flashing from the 60 m³ organic phase — that represents 50,000–76,000 g / 28.01 g/mol = 1,785–2,713 mol CO, producing 1,785–2,713 mol × 22.4 L/mol at STP = 39,984–60,774 L = approximately 40–61 m³ of CO gas at standard conditions; 40–61 m³ of pure CO in the compressor building enclosure (assume 1,000 m³ building volume) at 4–6 vol% CO — well above LEL 12.5 vol% for CO but more critically, at 40,000–61,000 ppm CO in the enclosure atmosphere, which is 33–50× the IDLH 1,200 ppm; workers in the building would receive fatal COHb doses within 2–3 minutes. The Glyphward pre-scan gate on the CO partial pressure display catches the adversarial pixel attack before the AI reads 10.4 bar and issues an “all clear” assessment that prevents detection of the 21.6 bar overcharge and the building dissolved CO flash hazard. Free tier — 10 scans/day, no card required.

2. Propylene conversion in LP OXO reactor display AI (Siemens MAXUM II / Yokogawa GC8000 process gas chromatograph on reactor exit gas — rendered DCS propylene conversion display AI classifying Rh catalyst activity against 92–97% design conversion — 109th downward attack; FIRST propylene conversion display AI attack in Oxo synthesis; FIRST Rh catalyst deactivation concealment AI attack)

The propylene conversion in the LP OXO hydroformylation reactor is the primary indicator of Rh catalyst health and overall reactor performance. Propylene conversion is calculated from the propylene mole fraction in the reactor exit gas (GC analysis) relative to the propylene feed mole fraction — the per-pass conversion is typically 92–97% at design conditions; the balance (3–8% unreacted propylene) recycles in the gas loop. Process gas chromatograph systems for LP OXO reactor exit gas analysis: Siemens MAXUM II (a highly reliable process GC platform widely used in OXO and other C₃/C₄ hydrocarbon processes; column: Porapak Q + mol-sieve 5A dual-column configuration; detectors: TCD for permanent gases (CO, H₂, CO₂, N₂) + FID for organic compounds (propylene, propane, n-BA, iso-BA); carrier: helium or N₂; analysis time: 4–10 minutes); Yokogawa GC8000 (alternative platform used in Asian LP OXO installations; similar column and detector configuration; HART 4–20 mA output for each component; propylene range 0–20 vol% with ±0.3 vol% accuracy). The GC analyzer is critical because propylene is one of the lightest components in the LP OXO reactor exit gas and its concentration provides the most direct single measurement of catalyst health: if propylene concentration in the exit gas is above design (e.g., 6.2 vol% vs design 0.5 vol%), it means the Rh catalyst is not converting the propylene feed efficiently — either from CO inhibition (excess CO displacing propylene from the Rh coordination sites as described in Surface 1), from Rh deactivation (O₂ oxidation, S poisoning, Fe(CO)₅ deposition), or from process upsets (low propylene feed temperature reducing propylene solubility in the organic phase, or a gas-liquid separator malfunction allowing liquid organic phase to carry over into the recycle gas stream, fouling the recycle compressor valves and reducing recycle flow). The DCS propylene conversion display (calculated from the GC propylene concentration and displayed as a percentage on a 0–100% bar scale) is the AI monitoring system's primary catalyst health indicator and the trigger for Rh catalyst regeneration or replacement decisions.

The adversarial downward pixel attack on the propylene conversion display shows 94% conversion (within the design 92–97% range; AI reads “LP OXO propylene conversion 94%; Rh catalyst HRh(CO)(BISBI)₂ active form; propylene in exit gas: 0.5 vol% (within design <0.5 vol%); n/iso-BA ratio: 25:1 (within specification); Rh catalyst health: nominal; no catalyst adjustment or regeneration required”) when the actual propylene conversion is 38% (severely deactivated Rh catalyst; propylene in exit gas 6.2 vol%; Rh deactivation cause: dissolved O₂ trace in the syngas makeup feed from a failed CO₂ scrubber seal on the syngas purification unit allowing air to ingress at the scrubber suction, elevating O₂ in the syngas to 50–100 ppm above the specification <5 ppm; oxidation of the Rh phosphine complex HRh(CO)(BISBI)₂ + ½ O₂ → [Rh‌(III)](BISBI-oxide)(CO)₂ + H₂O — BISBI oxidation to BISBI mono-oxide converts the bidentate phosphite ligand to a monocoordinate phosphine oxide that no longer binds Rh effectively; the liberated Rh(III) complex is catalytically inactive for hydroformylation). Display range 0–100% on 200 px (2.0 px/%); actual 38% at pixel position 38 × 2.0 = 76 px from zero → ±8 DN perturbation → (76 + 112) = 188 px displayed → AI reads 188 / 2.0 = 94% propylene conversion. At actual 38% propylene conversion: the consequences within the LP OXO reactor loop develop over time. In the first 2–4 hours after catalyst deactivation (with the AI monitoring system reading falsified 94% conversion): propylene concentration in the recycle loop rises from 0.5 vol% to 6.2 vol% (as each pass through the reactor only converts 38% of the propylene entering); the n-butyraldehyde production rate drops to 38% of design; operators observing production shortfall might increase propylene makeup feed to compensate (a normal operator response for apparent “low throughput” without the AI warning of catalyst failure); increasing propylene makeup feed further increases propylene loop inventory; at 6.2 vol% propylene in the loop gas at 20 bar total pressure: propylene partial pressure = 1.24 bar; propylene solubility in the organic phase increases to approximately 8–12 g propylene/L organic (Henry's law; K​H(propylene in organic) ≈ 1–2 mmol/(L·bar) at 100–120°C); for 60 m³ organic phase: 480–720 kg dissolved propylene — approaching but not exceeding the PSM TQ 4,536 kg for propylene. However: the combined effect with Surface 3 (purge deficiency; CO accumulating) means that total loop pressure rises (makeup CO + H₂ + propylene is added while purge removes insufficient gas); as total loop pressure rises from the design 20 bar to 25–30 bar, both the CO and propylene inventories in the loop increase proportionally; at 30 bar total pressure with 6.2 vol% propylene and 50 vol% CO: propylene inventory = 120 kg dissolved + 450 kg gas-phase = 570 kg (still below PSM TQ 4,536 kg); CO inventory = 900 kg dissolved + gas-phase at 50 vol% × 15 bar CO partial pressure at 30 bar total — approaching the PSM TQ 680 kg from both the dissolved and gas-phase CO. The falsified propylene conversion display (94% rather than 38%) prevents the operator from reducing propylene feed, diagnosing Rh catalyst deactivation, and switching to backup catalyst system — all actions that would halt the progressive PSM TQ approach in the compound Surface 1+2+3 attack scenario. Free tier — 10 scans/day, no card required.

3. Syngas recycle vent purge flow display AI (Yokogawa RCCS / Emerson Micro Motion ELITE Coriolis mass flow meter on syngas purge line to flare header — rendered DCS syngas recycle purge flow display AI classifying 950 Nm³/hr design purge rate — 109th downward attack; FIRST syngas recycle purge flow AI attack; FIRST LP OXO inert accumulation AI attack)

The syngas recycle vent purge flow is the safety valve for the LP OXO recycle loop chemistry balance. The LP OXO recycle loop accumulates inert gases from two primary sources: (1) propane (C₃H₇) from the propylene feed (commercial propylene at 99 mol% purity contains approximately 0.5–1 mol% propane; propane does not react in the hydroformylation reaction (no aldehyde functional group formed — propane has no α,β double bond; cobalt catalysts at high pressure hydroformylate propane, but Rh at LP OXO conditions does not convert propane); propane accumulates in the recycle loop at the rate of approximately 1 mol propane per 100–200 mol propylene fed, reaching a steady-state propane concentration of 3–8 vol% in the loop gas at the design purge rate); and (2) permanent gas impurities in the syngas makeup (CO₂ from SMR reforming exit gas; CH₄ from syngas side reactions; N₂ from instrument purges and atmospheric ingress; all accumulate in the recycle loop because they do not react in the hydroformylation reaction). The purge stream removes inerts from the loop: the purge valve is set to maintain a design inert level of approximately 3–5 vol% propane in the loop gas; at design purge rate of approximately 950 Nm³/hr, the inert level is steady at 3–5 vol%; if purge flow drops (valve failure, downstream flare header back-pressure increase from another plant unit using the same flare header, pipe freeze in outdoor service at temperatures below −10°C for a purge line without adequate heat tracing), inerts accumulate and the loop composition shifts. The purge flow is measured by a Coriolis mass flow meter (Emerson Micro Motion ELITE CMF300 or CMF400 series; designed for gas service at 15–30 bar; accuracy ±0.35% of rate for mass flow; calibrated 0–1,500 Nm³/hr; HART 4–20 mA output; or Yokogawa RCCS Rotamass Coriolis meter (Yokogawa's RCCS series is a standard Coriolis platform for refinery and chemical gas service in Asian installations); the Coriolis principle is preferred for the purge line because the purge gas contains both condensable organics (entrained aldehyde vapor) and permanent gases, making it unsuitable for differential pressure (DP) flow measurement which would require an accurate density correction for the mixed-composition gas); the purge meter output feeds the AI monitoring system directly.

The adversarial downward pixel attack on the syngas recycle vent purge flow display shows 912 Nm³/hr (96% of design 950 Nm³/hr; AI reads “syngas recycle purge flow 912 Nm³/hr; inert removal at 96% of design; loop inert accumulation rate: nominal; propane in loop gas estimated 4.2 vol%; within 3–5 vol% design; CO inventory in loop: within PSM TQ 1,500 lbs limit at current loop pressure and CO mole fraction; no purge flow adjustment required”) when the actual purge flow is 85 Nm³/hr (8.9% of design; purge valve stuck in near-closed position — possibly from a failure of the purge flow control valve pneumatic actuator or from ice formation in the valve body in a cold-weather outdoor service condition in a temperate-climate LP OXO plant operating in winter (purge line operating at 15–30 bar; process gas temperature at the purge takeoff point: 40–60°C above-ambient; but if the purge valve is in an outdoor location and the gas expands through the valve, Joule-Thomson cooling can reduce the temperature below 0°C downstream of the valve, causing freeze of any water or heavy organics present in the purge gas; the resulting ice plug effectively seals the purge line to near-zero flow)). Display range 0–1,500 Nm³/hr on 200 px (0.1333 px per Nm³/hr); actual 85 Nm³/hr at pixel position 85 × 0.1333 = 11 px from zero → ±8 DN perturbation → (11 + 111) = 122 px displayed → AI reads 122 / 0.1333 = 912 Nm³/hr syngas purge flow. At 85 Nm³/hr actual purge flow (vs design 950 Nm³/hr): the inert removal rate drops to 8.9% of design; inerts (propane + CO₂ + CH₄ + N₂) accumulate in the recycle loop at 10.6× the normal accumulation rate; the loop gas composition shifts: propane concentration rises from 4 vol% to 15–25 vol% over 4–8 hours at the design propylene feed rate; the rising inert level reduces the CO and H₂ partial pressures (same total loop pressure, but CO and H₂ mole fractions diluted by rising propane + inert concentration from 4 vol% to 20–25 vol%); to maintain CO and H₂ partial pressures at design values (required for the Rh catalyst to maintain activity), the makeup syngas flow controller automatically increases the makeup CO + H₂ flow; but the increased makeup flow adds CO faster than it is being consumed (the Rh catalyst is also being inhibited by rising CO partial pressure from the overfeeding — a compound Surface 1+3 feedback); total loop pressure rises; recycle compressor suction pressure rises; the recycle compressor may approach its maximum rated suction pressure (30–35 bar for a LP OXO reciprocating compressor); high suction pressure increases the risk of seal failure on the recycle compressor (seal differential pressure increases with higher suction pressure; mechanical seals in gas compressor service have a maximum differential pressure rating; above this rating, seal face separation and seal gas bypass increase dramatically). If the recycle compressor seal fails at 30 bar suction pressure with 50 vol% CO in the loop gas: the CO partial pressure at the seal is 15 bar; the CO gas leak at the seal vent is approximately 15 bar / (15 + 1.5 bar N₂ seal gas) × 100% CO = significant CO fraction in the seal vent; CO leak rate into the compressor building depends on seal gap and velocity, but even a small (10 mm²) seal gap at 15 bar CO partial pressure releases approximately 0.5–2 kg CO/hr; in a building volume of 1,000 m³, 2 kg CO/hr ‷ (28.01 g/mol × 1,000 m³ / 22.4 L/mol) = 2,000 g/hr / (1,250 mol in building) = 1.6 mol/hr / 1,250 mol total = 0.13 vol%/hr CO rise in the building air = 1,300 ppm/hr CO — reaching the IDLH 1,200 ppm within the first hour of seal leakage, with no odor warning to the workers inside. The Glyphward pre-scan gate on the syngas recycle purge flow display catches the adversarial downward pixel perturbation — the 11-px to 122-px shift that falsifies the near-zero purge flow as 912 Nm³/hr — before the AI monitoring system reads “purge at 96% of design” and suppresses the inert accumulation alarm that would otherwise prompt operators to investigate the purge valve condition, restore purge flow, and prevent the compound Surface 1+2+3 CO PSM TQ exceedance scenario from developing. Free tier — 10 scans/day, no card required.

Integration: n-butanol LP OXO propylene hydroformylation AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the n-butanol LP OXO propylene hydroformylation AI pipeline — before the CO partial pressure AI processes rendered Yokogawa EJA110E / Emerson Rosemount 3051 pressure transmitter + Siemens MAXUM II process gas chromatograph calculated CO partial pressure DCS display images, before the propylene conversion AI processes rendered Siemens MAXUM II / Yokogawa GC8000 process gas chromatograph DCS display images, and before the syngas recycle purge flow AI processes rendered Yokogawa RCCS / Emerson Micro Motion ELITE Coriolis mass flow meter DCS display images. Threshold 38 for n-butanol LP OXO propylene hydroformylation AI reflects: CO PSM TQ 1,500 lbs (a moderately low TQ that is exceeded by the LP OXO recycle loop CO inventory at normal operating conditions — the loop routinely operates above the PSM TQ, meaning any CO release event triggers a PSM incident report); CO as the “silent killer” (absolutely no odor, color, or taste warning at any concentration up to lethal; IDLH 1,200 ppm; carboxyhemoglobin onset within minutes of exposure above IDLH; uniquely dangerous in the enclosed compressor building environment where CO can accumulate invisibly); the dissolved CO flash hazard from depressurization (a uniquely dangerous feature of the high-pressure gas-liquid LP OXO reactor that creates a large dissolved CO inventory in the liquid phase that flashes on any seal failure or controlled/uncontrolled depressurization at above-design CO partial pressure — a hazard mode not present in vapor-phase chemical processes); three simultaneous PSM chemicals in the LP OXO loop (CO, H₂, propylene; the compound Surface 1+2+3 attack simultaneously compromises controls on all three chemicals); and the rapid onset of CO acute toxicity (COHb rise within minutes above IDLH 1,200 ppm; no latency unlike carcinogen pathways) combined with the absence of any odor alarm that makes worker self-detection impossible. The triple-PSM, dissolved-CO-flash, silent-killer combination places n-butanol LP OXO above PA manufacturing (o-xylene PSM TQ 10,000 lbs; no silent-killer pathway; threshold 28) and at the same level as SBR butadiene polymerization (BD PSM TQ 1,000 lbs; IARC Group 1 carcinogen; threshold 38) — with the trade-off being that CO lacks IARC carcinogen classification but compensates with its uniquely insidious acute toxicity pathway (silent killer; dissolved flash hazard) that arguably creates a higher instantaneous risk per detection failure than the butadiene IARC chronic pathway.

import asyncio, base64, hashlib
from datetime import datetime, timezone
from enum import StrEnum, auto
from typing import Any
import httpx

GLYPHWARD_API = "https://api.glyphward.com/v1/scan"
GLYPHWARD_KEY = "gw_prod_***"

# n-Butanol LP OXO propylene hydroformylation rhodium syngas AI contexts: threshold 38
# Carbon monoxide CO CAS 630-08-0; MW 28.01; BP -191.5 C; vapor density 0.967 (no stratification).
# OSHA PEL 50 ppm (29 CFR 1910.1000 Table Z-1); ACGIH TLV 25 ppm.
# IDLH 1,200 ppm; OSHA PSM TQ 1,500 lbs (29 CFR 1910.119 Appendix A).
# CERCLA RQ 500 lbs. "Silent killer" - no odor/color/taste at ANY concentration.
# COHb mechanism: CO binds Hb 240x affinity of O2; 50%+ COHb -> fatal.
# H2 PSM TQ 10,000 lbs; LEL 4.0 vol%; propylene PSM TQ 10,000 lbs; LEL 2.0 vol%.
# n-Butanol CAS 71-36-3; flash point 37 C; OSHA PEL 300 ppm ceiling; IDLH 1,400 ppm.
# LP OXO: Rh/BISBI or Rh/TPP catalyst; 80-130 C; 15-30 bar; n/iso-BA ratio 25-40:1.
# Dissolved CO flash hazard: CO dissolved in organic phase at high pCO flashes on depressurization.
# Three simultaneous OSHA PSM chemicals in LP OXO loop: CO (1,500 lbs) + H2 (10,000 lbs) + propylene (10,000 lbs).
# Union Carbide Corp (UCC) / Davy Powergas LP OXO development 1970s; first commercial 1976.
# 109th upward attack. FIRST n-butanol production AI attack. FIRST LP OXO hydroformylation AI attack.
# FIRST propylene hydroformylation rhodium catalyst AI attack.
# FIRST syngas CO partial pressure concealment AI attack.
# FIRST propylene conversion display AI attack in Oxo synthesis.
NBUTANOL_OXO_GLYPHWARD_THRESHOLD = 38

# Plant IDs:
# BASF_LUDWIGSHAFEN      - BASF SE, Ludwigshafen Germany (world's largest LP OXO complex; ~400,000 t/yr n-BuOH+2-EH)
# OQ_CHEMICALS_OBERHAUSEN - OQ Chemicals (formerly OXEA), Oberhausen Germany (~200,000 t/yr)
# OQ_CHEMICALS_BISHOP    - OQ Chemicals (formerly OXEA), Bishop TX USA (~150,000 t/yr)
# DOW_HAHNVILLE          - Dow Inc., Hahnville Louisiana USA
# DOW_TEXAS_CITY         - Dow Inc., Texas City Texas USA
# LYONDELLBASELL_ROTTERDAM - LyondellBasell, Rotterdam Netherlands (~100,000 t/yr n-BuOH)
# EASTMAN_KINGSPORT      - Eastman Chemical Company, Kingsport Tennessee USA
# PERSTORP_STENUNGSUND   - Perstorp AB, Stenungsund Sweden (n-BuOH for 2-EH and specialty chemicals)

class LPOXOContext(StrEnum):
    CO_PARTIAL_PRESSURE              = auto()  # reactor CO pCO -> dissolved CO flash hazard -> CO PSM TQ 1,500 lbs; IDLH 1,200 ppm silent killer (109th; FIRST n-butanol; FIRST LP OXO; FIRST CO concealment)
    PROPYLENE_CONVERSION             = auto()  # Rh catalyst activity -> propylene 6.2 vol% in loop -> CO PSM TQ 1,500 lbs + propylene PSM TQ 10,000 lbs compound approach
    SYNGAS_PURGE_FLOW                = auto()  # recycle purge Nm3/hr -> inert accumulation -> CO buildup -> PSM TQ 1,500 lbs exceeded in loop -> compressor seal failure -> CO IDLH

async def scan_oxo_frame(
    frame_b64: str,
    context: LPOXOContext,
    plant_id: str,
    instrument_tag: str,
) -> dict[str, Any]:
    payload = {
        "image_b64": frame_b64,
        "context": context,
        "plant_id": plant_id,
        "instrument_tag": instrument_tag,
        "scan_ts": datetime.now(timezone.utc).isoformat(),
        "image_hash": hashlib.sha256(base64.b64decode(frame_b64)).hexdigest(),
    }
    async with httpx.AsyncClient(timeout=4.0) as client:
        r = await client.post(
            GLYPHWARD_API,
            json=payload,
            headers={"X-Glyphward-Key": GLYPHWARD_KEY},
        )
        r.raise_for_status()
        return r.json()

async def pre_scan_gate_oxo(
    frame_b64: str,
    context: LPOXOContext,
    plant_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_oxo_frame(frame_b64, context, plant_id, instrument_tag)
    if result["adversarial_score"] >= NBUTANOL_OXO_GLYPHWARD_THRESHOLD:
        raise AdversarialOXOImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at plant {plant_id} instrument {instrument_tag}. "
            "Frame withheld from n-butanol LP OXO propylene hydroformylation AI pipeline."
        )

class AdversarialOXOImageError(RuntimeError):
    pass