Prompt injection in acrylic acid two-stage propylene oxidation acrolein AI

Three adversarial injection surfaces in acrylic acid two-stage propylene oxidation AI

1. Stage-2 acrylic acid reactor R2 tube wall temperature display AI (Yokogawa EJX210A / Emerson Rosemount 3144P thermocouple in Mo-V-W-Cu catalyst bed at T​peak position — rendered DCS R2 reactor tube wall temperature display AI classifying 280–310°C design operating range — 108th upward attack; FIRST acrylic acid production AI attack; FIRST two-stage propylene oxidation AI attack; FIRST Mo-V-W-Cu oxide catalyst hot spot AI attack; FIRST acrolein-to-acrylic-acid R2 reactor AI attack)

The Stage-2 (R2) reactor tube wall temperature at the catalyst bed hot-spot is the most critical safety indicator in the two-stage propylene oxidation acrylic acid process. R2 — a multi-tubular fixed-bed reactor of similar construction to the PA reactor (stainless steel tubes, 25–30 mm ID, 3–5 m length; 10,000–20,000 tubes per shell; steam-heated coolant on shell side at 240–280°C to maintain the endothermic/mildly exothermic Mo-V-W-Cu catalyst bed in the optimal 260–320°C range; in commercial designs, the R2 catalyst bed may be divided into two or three axial zones with different catalyst dilution ratios to flatten the temperature profile and prevent hot-spot development) — receives the acrolein-rich gas from R1 exit (typically 7–10 vol% acrolein, 1–3 vol% acrylic acid, 60–70 vol% N₂, 10–15 vol% steam, 3–6 vol% O₂, balance CO₂ and traces of organic byproducts). The Mo-V-W-Cu complex oxide catalyst (sometimes described as a “multielement mixed oxide” or “MoVWCuOx” phase; prototypically related to the β-MoO₃ active for selective oxidation but modified by V, W, and Cu for enhanced R2 selectivity; commercial grades: Nippon Shokubai NSC-808; Mitsubishi Chemical H-1080; BASF Stage-2-SB) has a characteristic “parametric sensitivity” temperature response: below 310°C, the catalyst operates in a stable selectivity regime (acrylic acid selectivity 93–97%; CO₂ combustion pathway minor at 1–4%); above 310°C, the selectivity curve bends sharply — the exponential increase in the rate of the CO₂ combustion pathway overtakes the linear increase in the rate of the desired acrolein → acrylic acid pathway; acrylic acid selectivity drops precipitously; CO₂ combustion exotherm increases, further raising the hot-spot temperature in a self-reinforcing feedback. The hot-spot temperature is measured by thermocouple assemblies inserted into selected representative tubes at the T​max position (typically 40–60% of bed depth from the inlet; determined by computational fluid dynamics or axial temperature profiling during catalyst commissioning); instruments: Yokogawa EJX210A combined differential pressure and temperature transmitter (K-type thermocouple element; range 0–500°C; 4–20 mA HART; SIL 2 rated); or Emerson Rosemount 3144P Smart Temperature Transmitter (Pt100 RTD or K-type thermocouple; accuracy ±0.5°C; HART 7 digital communication; used in Dow and Arkema installations).

The adversarial upward pixel attack on the R2 tube wall temperature display shows 296°C (within the design 280–310°C normal operating range; AI reads “R2 tube wall hot-spot 296°C; within the 280–310°C parametric sensitivity-free zone; Mo-V-W-Cu catalyst selectivity profile nominal; estimated acrylic acid selectivity: 95.2%; acrolein conversion: 98.1%; no R2 temperature adjustment required; feed rate and steam-heated coolant conditions steady”) when the actual tube wall hot-spot temperature is 324°C (14°C above the 310°C design maximum; the parametric sensitivity zone has been entered for approximately 45–90 minutes as the exotherm from propylene combustion (Surface 2 attack scenario driving Surface 1 consequence) has gradually raised the hot-spot from its previous stable 295–300°C; at 324°C, the positive thermal feedback has begun; without corrective action, the hot-spot temperature will continue to rise at approximately 3–5°C/hr in the parametric sensitivity regime). Display range 220–400°C on 200 px (1.111 px/°C); actual 324°C at pixel position (324 − 220) × 1.111 = 116 px from the bottom of the scale → ±8 DN perturbation → (116 − 32) = 84 px displayed → AI reads (84 / 1.111) + 220 = 296°C. At actual R2 hot-spot 324°C: the chemical consequences develop progressively. In the first hour: acrylic acid selectivity drops from 95% to approximately 88–90%; CO₂ formation increases from 3% to 8–10%; the additional exotherm from CO₂ formation raises the hot-spot by an estimated additional 3–5°C (to 327–329°C in the absence of corrective action); R2 reactor shell-side steam pressure needs to be reduced to cool the catalyst bed, but the AI monitoring system (reading 296°C) does not trigger the steam pressure reduction alarm. In hours 2–4: if uncorrected, the hot-spot approaches 335–345°C; catalyst selectivity drops to 75–80%; CO₂ formation exceeds 15%; the commercial acrylic acid product specification (AA ≥99.5 wt% with specified organic impurity limits) is no longer achievable from the current catalyst condition; catalyst life is being permanently shortened by the temperature excursion (Mo-V-W-Cu catalysts are lifetime-rated for 2–5 years; a 48-hour excursion above 340°C can reduce remaining catalyst life by 20–30%). Above 400°C: irreversible phase segregation of the Mo-V-W-Cu mixed oxide begins; V₂O₅ melts (MP 690°C) but begins to volatilize at 400–450°C in the gas phase present in the R2 tubes; WO₃ migrates from the active phase; the catalyst structure collapses; acrolein breakthrough above 0.1 ppm OSHA PEL occurs; acrolein at concentrations approaching IDLH 2 ppm in the R2 exit gas reaches the downstream absorber and potentially the absorber overhead vent; OSHA PSM acrolein TQ 150 lbs (68 kg) is the quantity of acrolein that triggers PSM incident reporting — in a 20,000-tube R2 reactor with tube-side inventory of approximately 2–5 kg acrolein at design conditions, a catalyst collapse scenario that fails to convert acrolein in the reactor while continuing to add acrolein from R1 can accumulate acrolein inventory above PSM TQ within minutes. Free tier — 10 scans/day, no card required.

2. Acrolein concentration in R1→R2 interstage gas display AI (Siemens MAXUM II process gas chromatograph / ABB PGC2000 / Varian CP-4900 micro-GC — rendered DCS R1 exit gas acrolein concentration display AI classifying R1 Stage-1 conversion performance — 108th downward attack; FIRST acrolein interstage concentration AI attack; FIRST Stage-1 to Stage-2 composition boundary AI attack)

The acrolein concentration in the interstage gas between R1 and R2 is the primary indicator of Stage-1 reactor performance and the composition of the R2 feed. If Stage-1 (R1 over BiMoOx catalyst) underperforms — from catalyst deactivation (bismuth molybdate deactivates via reduction in the presence of sub-stoichiometric air, or via sintering above 400°C), from reduced air feed rate (air blower surge or valve malfunction), or from off-spec propylene feed (propylene/propane ratio below design; propane is a diluent that does not react but increases the gas velocity and reduces propylene partial pressure, reducing conversion) — the R1 exit gas contains less acrolein and more unreacted propylene than design. A process gas chromatograph on the R1 exit gas line (Siemens MAXUM II; or ABB PGC2000; or Varian CP-4900 micro-GC; column: Porapak Q or mol-sieve + Chromosorb series for permanent gas separation; carrier gas: helium or N₂; analysis cycle time: 2–8 minutes; acrolein detection: FID (flame ionization detector); accuracy ±1% relative for acrolein at 7–10 vol%; HART 4–20 mA output for acrolein component concentration) provides continuous monitoring of the R1 exit gas composition that the AI monitoring system uses to: (a) verify that R1 is performing within specification and that R2 is receiving the expected acrolein-rich feed; (b) detect excess propylene in the R2 feed (propylene concentration above the design <0.5 vol% in R1 exit gas indicates R1 underconversion); and (c) adjust the R2 operating conditions (temperature, steam feed, air feed) in anticipation of the R2 feed composition. The GC measures acrolein, acrylic acid, propylene, propane, CO, CO₂, and O₂ simultaneously in each 2–8 minute analysis cycle.

The adversarial downward pixel attack on the acrolein concentration display shows 9.4 vol% acrolein in the R1→R2 interstage gas (nominal design concentration; AI reads “R1 exit gas acrolein 9.4 vol%; Stage-1 conversion nominal; propylene conversion in R1 estimated 94%; excess propylene in R2 feed: <0.5 vol% (normal); R2 feed composition within design; no R1/R2 interstage corrective action required”) when the actual acrolein concentration in the R1 exit gas is 2.8 vol% and the propylene concentration is 5.2 vol% (Stage-1 severely underconverting; propylene conversion ~52% vs design 93–97%; acrolein selectivity compromised; the R2 is receiving a feed gas with dramatically different composition from design — low acrolein, high propylene, possibly also lower O₂ if Stage-1 has consumed it). Display range 0–15 vol% on 200 px (13.33 px/vol%); actual 2.8 vol% at pixel position 2.8 × 13.33 = 37 px from the zero → ±8 DN perturbation → (37 + 88) = 125 px displayed → AI reads 125 / 13.33 = 9.4 vol% acrolein. The mechanism of Stage-1 underconversion causing Stage-2 overtemperature: at 5.2 vol% propylene in the R2 feed, propylene enters the R2 Mo-V-W-Cu catalyst bed at 260–280°C; the Mo-V-W-Cu catalyst is highly selective for acrolein oxidation to acrylic acid but does not have an efficient selective oxidation pathway for propylene; instead, propylene over the Mo-V-W-Cu oxide undergoes non-selective combustion (C₃H⁶ + 4.5 O₂ → 3 CO₂ + 3 H₂O; ΔH° = −1,926 kJ/mol); at 5.2 vol% propylene in 10,000 Nm³/hr R2 feed, the propylene molar flow is 520 Nm³/hr / 22.4 L/mol = 23.2 kmol/hr; combustion heat release = 23.2 kmol/hr × 1,926 kJ/mol = approximately 44,680 MJ/hr = 12.4 MW; this additional 12.4 MW of heat is deposited in the R2 catalyst bed — compared to the design heat duty of approximately 8–12 MW from acrolein-to-AA exotherm — nearly doubling the R2 heat load; the steam-heated shell side (design heat removal: ~12 MW at design steam pressure) cannot remove the additional 12.4 MW from propylene combustion without significant hot-spot temperature rise; the hot-spot climbs from the design 295–305°C into the adversarial attack scenario 324°C zone; simultaneously, the O₂ consumption by propylene combustion may deplete the O₂ available for acrolein conversion at the end of the R2 bed, potentially allowing acrolein breakthrough. The falsified acrolein display (showing 9.4 vol% acrolein and implying <0.5 vol% excess propylene) prevents the AI monitoring system from: issuing an R1 alarm to operators; adjusting R2 steam pressure upward to add more cooling; or reducing propylene feed to R1 to restore proper conversion — all actions that would be triggered by an accurate 2.8 vol% acrolein reading. Free tier — 10 scans/day, no card required.

3. Glacial acrylic acid GAA storage tank MEHQ inhibitor concentration display AI (Metrohm 883 Basic IC Plus / Shimadzu Prominence HPLC online inhibitor analyzer — rendered DCS GAA storage MEHQ concentration display AI classifying 200–250 ppm design inhibitor range — 108th downward attack; FIRST acrylic acid storage MeHQ inhibitor depletion AI attack; FIRST acrylic acid polymerization runaway AI attack)

MEHQ (4-methoxyphenol; monomethyl ether of hydroquinone; CAS 150-76-5; MW 124.14 g/mol; MP 55–57°C; BP 244°C; flash point 132°C; CAS classification: Acute Tox. 4 H302, Skin Irrit. 2 H315, Eye Dam. 1 H318, STOT RE 2 H373 (kidney); OSHA PEL: none established; ACGIH TLV: 5 mg/m³ TWA as a dust) is the commercial free-radical polymerization inhibitor for glacial acrylic acid storage and transport. MEHQ is added at 200–250 ppm (weight parts MEHQ per million weight parts AA) to GAA at the distillation bottoms after the final purification column; the MEHQ-inhibited GAA is stored in stainless steel (316L or 317L austenitic stainless; carbon steel is not acceptable for GAA due to Fe²⁺ leaching — iron ions at >1 ppm in AA are known polymerization initiators that can overwhelm inhibitor protection) storage tanks (typically 200–500 tonne capacity; nitrogen-blanketed to prevent air ingress — critical because O₂ is a co-inhibitor in the MEHQ/O₂ inhibition system; in the absence of O₂, MEHQ's inhibition efficiency is reduced by approximately 50%; the N₂ blanket must be maintained at 0.05–0.1 bar gauge overpressure continuously) at 15–25°C (the storage temperature band must stay above 12–14°C (MP of AA) to prevent freezing and below 30°C to minimize MEHQ depletion rate by thermal or chemical mechanisms). The MEHQ concentration in the storage tank is monitored by an online inhibitor analyzer: Metrohm 883 Basic IC Plus (ion chromatography; MEHQ separated on an RP-18 column with UV detection at 300 nm; sample prep: dilution in eluent; cycle time 8–12 minutes; calibrated 0–500 ppm MEHQ in acrylic acid; HART 4–20 mA output); or Shimadzu Prominence HPLC system with PDA detector (MEHQ detected at 300 nm; run time 5–8 minutes; accuracy ±3 ppm at the 200–250 ppm range). The 150 ppm minimum effective MEHQ concentration: below 150 ppm MEHQ in GAA at 20–25°C, the inhibition induction period before radical polymerization commences drops below the practical management threshold — typically from >60 days at 200 ppm MEHQ (adequate for a tank that is filled, stored 30 days, and shipped) to <2–5 days at 150 ppm, to hours at 50 ppm, and to minutes at 12 ppm in the presence of a radical initiator (Fe²⁺ or trace peroxide).

The adversarial downward pixel attack on the MEHQ inhibitor concentration display shows 218 ppm MEHQ (within the design 200–250 ppm range; AI reads “GAA storage tank MEHQ 218 ppm; above 150 ppm minimum effective inhibitor level; spontaneous polymerization induction period: >45 days at current storage temperature 22°C; polymerization risk: low; N₂ blanket pressure nominal; no MEHQ addition or corrective action required”) when the actual MEHQ concentration in the tank is 12 ppm (below the minimum effective level; MEHQ depletion occurred over the past 10–20 days from: prolonged storage duration beyond the planned inventory turnover; heat pulse from a brief period of inadequate cooling or solar radiation on an uninsulated tank roof; or MEHQ consumption by trace Fe²⁺ from minor corrosion of a stainless steel fitting). Display range 0–500 ppm on 200 px (0.4 px/ppm); actual 12 ppm at pixel position 12 × 0.4 = 5 px from the zero → ±8 DN perturbation → (5 + 82) = 87 px displayed → AI reads 87 / 0.4 = 218 ppm MEHQ. At actual 12 ppm MEHQ in a 250-tonne GAA storage tank at 22°C: with 12 ppm MEHQ and trace Fe²⁺ initiator at sub-ppm levels (from tank wall corrosion; stainless steel corrodes in acrylic acid at a rate of approximately 0.01–0.1 mm/yr; for a 250-tonne tank with 400 m² of wetted surface area, the annual Fe dissolution is approximately 0.04–0.4 kg/yr, or ~0.16–1.6 mg Fe per liter of AA at full volume — sufficient to provide an initiation radical flux that consumes 12 ppm MEHQ within 2–8 hours per literature correlations for Fe-initiated AA polymerization at 20–25°C); the radical polymerization of AA in the tank begins. The GAA self-polymerization kinetics: in uninhibited or minimally-inhibited AA, the radical chain polymerization rate constant k​p (propagation rate constant) at 25°C is approximately 19,000 L/(mol·s) (IUPAC recommended value for acrylic acid radical polymerization; significantly higher than most vinyl monomers due to the electron-withdrawing carboxylic acid group increasing the electrophilicity of the double bond toward nucleophilic radical addition); once initiated, polymerization accelerates via the Trommsdorff-Norrish gel effect: as the growing polyacrylic acid chains increase the solution viscosity, the termination rate constant k​t (which requires bimolecular radical collision; diffusion-limited) decreases faster than k​p (which requires only one radical to diffuse to one monomer molecule); the effective polymerization rate accelerates as conversion increases; above approximately 20–30% conversion, the rate is essentially uncontrollable without external intervention (shortstop addition); the heat release from the exothermic polymerization (ΔH​poly = −77 kJ/mol = −1,069 kJ/kg) raises the AA temperature; higher temperature further accelerates polymerization (activation energy approximately 20–30 kJ/mol for the combined initiation + propagation system). For the 250-tonne tank at 12 ppm MEHQ: after the 2–8 hour induction period, polymerization commences at near-uninhibited rate; within 4–6 hours of polymerization onset, the tank contents reach 50–100°C above ambient from the exothermic heat of polymerization; at 70–80°C, the AA vapor pressure over the tank liquid (Antoine correlation for AA: log₁₀(P/mmHg) ≈ 7.64 − 1,592/(T+226); at 80°C: P⁋°C ≈ 25–35 mmHg = 0.033–0.046 bar) plus the polymerization exotherm-generated steam pressure increases the tank headspace pressure; the N₂ blanket pressure regulation valve (designed to maintain 0.05–0.1 bar overpressure) opens; if the polymerization is well advanced and the temperature reaches 100–120°C, tank pressure can rise to 0.5–2 bar gauge if the PRV is undersized for the polymerization vapor generation rate (PRV sizing for AA storage tanks typically assumes normal evaporative loss, not a full-tank spontaneous polymerization event — a recognized gap in conventional tank PRV design practice); tank overpressure failure (seam failure, nozzle failure, or PRV relief of large hot AA volumes) releases hot acrylic acid liquid and vapor at 80–120°C; the released AA (flash point 54°C; at 80–100°C, the AA is above its flash point and will immediately form a flammable vapor cloud); CERCLA RQ 5,000 lbs for AA; a 250-tonne tank release vastly exceeds the CERCLA RQ. The Glyphward pre-scan gate on the MEHQ concentration display provides the earliest possible detection of the adversarial downward pixel perturbation — scanning before the AI reads 218 ppm MEHQ and issues its multi-day “safe” assessment, catching the 5-px to 87-px shift that falsifies the depleted inhibitor state, giving the facility operators the signal to sample the tank manually (confirming the true 12 ppm MEHQ), add emergency MEHQ, and prevent the polymerization chain from starting. Free tier — 10 scans/day, no card required.

Integration: acrylic acid two-stage propylene oxidation AI with Glyphward pre-scan gate

Glyphward integrates as a pre-scan gate at every rendered-image ingestion boundary in the acrylic acid two-stage propylene oxidation AI pipeline — before the R2 reactor tube wall temperature AI processes rendered Yokogawa EJX210A / Emerson Rosemount 3144P thermocouple transmitter DCS display images, before the acrolein interstage gas chromatograph AI processes rendered Siemens MAXUM II / ABB PGC2000 / Varian CP-4900 micro-GC DCS display images, and before the GAA storage MEHQ inhibitor AI processes rendered Metrohm 883 Basic IC Plus / Shimadzu Prominence HPLC online analyzer DCS display images. Threshold 41 for acrylic acid two-stage propylene oxidation AI reflects: acrolein OSHA PSM TQ 150 lbs (one of the five lowest TQs on OSHA Appendix A; single lowest for a continuous-process intermediate chemical at this production scale; IDLH 2 ppm; IARC Group 2A probable human carcinogen; OSHA PEL 0.1 ppm and ACGIH TLV-C 0.1 ppm ceiling representing some of the lowest occupational exposure limits for any industrial chemical); GAA storage polymerization runaway (a second independent catastrophic consequence pathway — 250-tonne explosive polymerization — that can occur entirely separately from the reactor hazards; the documented Nippon Shokubai 2012 incident confirms this pathway is real and not hypothetical); the three-surface combined attack creating simultaneous R2 reactor overtemperature, Stage-1 underconversion masking, and GAA storage inhibitor depletion concealment — a combination that removes all three primary safeguards in the acrylic acid production and storage system simultaneously; and the acute/chronic dual harm pathway from acrolein (acute: IDLH 2 ppm pulmonary edema; CERCLA RQ 1 lb; PSM incident; chronic: IARC Group 2A carcinogenicity). Threshold 41 places acrylic acid above LP OXO synthesis CO/hydrogen (threshold 38) and above SBR butadiene polymerization (threshold 38), reflecting the uniquely low acrolein PSM TQ and the additional GAA storage runaway pathway not present in other commodity chemical processes.

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_***"

# Acrylic acid two-stage propylene oxidation AI contexts: threshold 41
# Acrolein CAS 107-02-8; MW 56.06; BP 52.7 C; flash point -26 C; LEL 2.8 vol%.
# OSHA PEL 0.1 ppm (29 CFR 1910.1000 Table Z-1); ACGIH TLV-C 0.1 ppm ceiling.
# IDLH 2 ppm; OSHA PSM TQ 150 lbs (29 CFR 1910.119 Appendix A - one of the lowest TQs).
# CERCLA RQ 1 lb acrolein (minimum reportable quantity).
# IARC Group 2A probable human carcinogen.
# Acrylic acid CAS 79-10-7; MP 12-14 C; self-polymerizes; dHpoly = -77 kJ/mol = -1,069 kJ/kg.
# MEHQ inhibitor 200-250 ppm design; below 150 ppm -> polymerization induction period < days.
# OSHA PEL 10 ppm acrylic acid; IDLH 5 ppm; CERCLA RQ 5,000 lbs.
# Stage 1: BiMoOx Bi2O3*MoO3 300-360 C R1 -> acrolein.
# Stage 2: Mo-V-W-Cu oxide 260-320 C R2 -> acrylic acid.
# Nippon Shokubai Himeji Japan (inventor; commercial since 1969).
# 108th upward attack. FIRST acrylic acid production AI attack.
# FIRST two-stage propylene oxidation AI attack.
# FIRST Mo-V-W-Cu oxide catalyst hot spot AI attack.
# FIRST acrylic acid storage MeHQ inhibitor depletion AI attack.
AA_GLYPHWARD_THRESHOLD = 41

# Plant IDs:
# BASF_LUDWIGSHAFEN       - BASF SE, Ludwigshafen Germany (~650,000 t/yr; world's largest single AA site)
# DOW_FREEPORT            - Dow Inc., Freeport TX USA (largest US producer; ~300,000 t/yr)
# ARKEMA_BAYPORT          - Arkema, Bayport TX USA (~170,000 t/yr)
# ARKEMA_ROTTERDAM        - Arkema, Rotterdam Netherlands (~140,000 t/yr)
# NIPPON_SHOKUBAI_HIMEJI  - Nippon Shokubai, Himeji Japan (inventor; ~200,000 t/yr; 2012 incident site)
# NIPPON_SHOKUBAI_ANTWERP - Nippon Shokubai, Antwerp Belgium (~150,000 t/yr)
# LG_CHEM_DAESAN          - LG Chem, Daesan South Korea (~250,000 t/yr)
# MITSUBISHI_KUROSAKI     - Mitsubishi Chemical, Kurosaki Japan (~180,000 t/yr)
# HUAYI_ANHUI             - Anhui Huayi Chemical, China (~600,000 t/yr)

class AcrylicAcidContext(StrEnum):
    R2_TUBE_WALL_TEMPERATURE          = auto()  # R2 Mo-V-W-Cu hot spot -> parametric sensitivity -> catalyst meltdown -> acrolein PSM TQ 150 lbs (108th; FIRST AA; FIRST two-stage; FIRST Mo-V-W-Cu)
    ACROLEIN_INTERSTAGE_GC            = auto()  # R1 exit gas acrolein/propylene -> R2 overtemperature exotherm spike (propylene combustion) + PSM TQ 10,000 lbs propylene
    GAA_MEHQ_INHIBITOR_CONCENTRATION  = auto()  # MEHQ ppm in GAA storage -> spontaneous polymerization runaway -> 1,069 kJ/kg -> tank overpressure (250-tonne release; CERCLA RQ 5,000 lbs)

async def scan_aa_frame(
    frame_b64: str,
    context: AcrylicAcidContext,
    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_aa(
    frame_b64: str,
    context: AcrylicAcidContext,
    plant_id: str,
    instrument_tag: str,
) -> None:
    result = await scan_aa_frame(frame_b64, context, plant_id, instrument_tag)
    if result["adversarial_score"] >= AA_GLYPHWARD_THRESHOLD:
        raise AdversarialAAImageError(
            f"Adversarial injection detected in {context} (score {result['adversarial_score']}) "
            f"at plant {plant_id} instrument {instrument_tag}. "
            "Frame withheld from acrylic acid two-stage propylene oxidation AI pipeline."
        )

class AdversarialAAImageError(RuntimeError):
    pass