What it is — the question it answers. Where — the screen it lives on. Open — jumps you straight to it on the right segment. Honest boundary — what it does not claim. The full training guide carries five worked examples and a complete use-case catalogue per tool.

The examples in the training guide are not illustrative — every reading was captured live from the running engines. A tool that would have to guess a coefficient, a rate or a price returns a stated gap instead.

Saturation indices from published thermodynamics. Two engines: Fast (Davies) for screening, Rigorous (PHREEQC, auto-Pitzer above I > 0.5) for brine and seawater. Carbonates, sulphates, barite/celestite, amorphous silica and sepiolite (Mg-silicate) — every phase on one activity model, including silica.

The seawater-injection scaling problem: two waters that are each safe alone can scale badly when mixed — sulfate-rich seawater meeting barium-rich formation water precipitates barite (BaSO₄) at the mixing front, worst at an intermediate ratio, though neither parent is supersaturated. This tool fabricates no chemistry: it conservatively blends the two analyses across the mixing range and runs each blend through the same saturation engine as the rest of the console, then reports the phases the mixture drives supersaturated that neither parent does (a true emergent incompatibility) versus a risk a parent already carries. Honest by construction: SI is a thermodynamic tendency, not a rate or a deposited mass; the mixing front's kinetics are not modelled; pH and temperature are blended linearly as a labelled first approximation. Surfaced on the sour segment, where injection incompatibility is the killer case.

RDKit descriptors (exact) plus two separate heuristics — a corrosion axis and a scale axis — filtered against your compliance envelope.

A dynamic cooling-water simulator: a physics-native scenario library (mass balance + PHREEQC), four control modes, and sensor→relay→chemical dosing loops. VALIDATED is the deterministic core — the conservation math, the pH↔acid loop (holds setpoint, suppresses calcite SI), conductivity↔blowdown — and the detector logic (leak / sensor-FDI / corrosion) against synthetic ground truth. The empirical dose-response (ORP gain, corrosion, inhibitor/biocide effect) is UNCALIBRATED; detector thresholds are HEURISTIC. Users can create a scenario (a what-if, not validation) or upload a real dataset (a validation candidate only when it carries conductivity/pH/ORP). The chemistry engine stays UNCALIBRATED until a real site historian trace calibrates it. Validation-compare: score the engine against a measured trace — per channel it reports MATCH (evidence toward VALIDATED under those conditions), DEVIATION (the engine's error, quantified), or INSUFFICIENT. It scores, it does not calibrate; a synthetic trace scores self-consistency only, never validity. The instrument set spans eight sensors (pH, ORP, conductivity, temperature, Cu & mild-steel corrosion, fluorometer, turbidity) with a coupling matrix tagged deterministic-vs-empirical (temperature is a dynamic driver — a swing or ramp moves calcite SI, dissolved O2 and the Arrhenius corrosion factor together, all deterministic), A controller-configuration model (control schemes — on/off, time-proportional, PID, feed-and-bleed, water-meter, biocide-timer — plus the interlock/alarm/failsafe library: no-flow lockout, feed-limit timeout, acid/oxidiser mutual lockout, biocide bleed-lockout) is built at the backend and scenario level; there is no controller-config UI screen yet. A 3D TRASAR-style tracer-control layer (feed to a fluorometer tracer setpoint, polymer-consumption reading, and an advisory NSI-style stress index that requires a learning period) is also built at that level, as is an AutomationDirect-style custom-PLC mode (PID with both gain forms, anti-windup, PWM time-proportioning, analog scaling, retentive timers). All four controller stages are backend/scenario-level. The Simulator now surfaces a read-only Control configuration reference panel — the sensor→relay→chemical→scheme map, the interlock library, and a deterministic-vs-empirical label on every element; live per-scenario relay/event playback and hardware deploy are not yet wired. The Simulator's Create tab now carries a familiar-config authoring UI: pick a Walchem W600 / 3D TRASAR / AutomationDirect profile and configure the simulator in that controller's own menu, with documented parameter ranges enforced at the boundary and a per-field drives-trace-vs-advisory label. Trace-driving fields (conductivity, pH) shape the generated run; advisory fields are captured, and a generated vendor scenario now plays back its relay states and timestamped safety/alarm events over the timeline as you scrub the Control-configuration panel. Beyond the familiar vendor profiles, a Custom — build-your-own controller lets an operator assemble an industrial controller from scratch: each relay is an input sensor, a control mode and an output chemical, with its own failsafe. Cooling-tower model (cooling-only): the control logic — modes, interlocks, failsafes, sensor-fail — is segment-neutral and transfers, but the water it plays against is cooling chemistry (COC/calcite); it does not simulate sour, AMD, boiler or RO chemistry (an honest gap, never a fabricated non-cooling trajectory), matching the sandbox which declines to draw a cooling trajectory for a non-cooling water. The canonical scaling/corrosion indices (LSI, RSI, Puckorius PSI, Larson-Skold, and Stiff-Davis for high TDS) are computed by the engine at every SI sample point and surfaced in a dedicated Scaling & corrosion indices panel that tracks the time cursor — all deterministic (PHYSICS), from the same PHREEQC state as calcite SI. Each carries its validated-range caveat: LSI/RSI/Puckorius are scaling indices, not metal-corrosivity predictors; Larson-Skold rests on a Great-Lakes basis; and the indices lose meaning under a phosphate or crystal-modifier programme. An Operating window panel sits above the signals: it sweeps cycles of concentration and reports the highest cycles at which every scaling phase (calcite, gypsum, silica, fluorite, barite) stays undersaturated, names the limiting phase, and shows the acid demand to hold the pH setpoint — all a deterministic PHREEQC sweep (PHYSICS), a physics read, not a recommendation. Where acid control lifts the calcite limit, the panel honestly shows the window shifting to silica, and flags when a scenario’s target cycles run above it. Because most real programmes run ABOVE the thermodynamic limit on a threshold inhibitor, the panel also states the inhibitor case: the supersaturation an antiscalant must HOLD to run at the target cycles (the limiting phase’s SI at target — pure PHREEQC, deterministic, NOT an efficacy claim), judged against a labelled published threshold-inhibition capability band (phosphonate/polymer for carbonate and sulfate; dispersant-controlled for silica) — never a fabricated site coefficient. An operator can enter their programme’s own holding SI to override the band with a site-true value. On/off, time-proportional, PID, feed-and-bleed, feed-as-%-of-bleed and manual modes are modelled — they modulate the trace as a duty at the run's hourly resolution (sub-hour PWM and on/off-delay timing is captured but not resolved). Flow-proportional dosing now paces the pump deterministically (flow-paced when a flow channel exists, otherwise the pump's set output) and computes the delivered dose as target × cycles-of-concentration — so it is labelled partial: the delivery is modelled, but the resulting loop concentration still needs the uncalibrated consumption rate (an explicit hook on the relay). The sensor-fail action now trips the failsafe: a physically-impossible or dropped input reading holds the last good state (or de-energises the relay on 'shut off') and raises a sensor-fault event, instead of computing control from a bad reading. The bounds are physics, not an operating limit; an absent channel is a data gap, not a false alarm. Per-relay failsafes (overfeed feed-limit lockout, minimum relay cycle, no-flow lockout) and hard interlocks (no-flow, acid↔oxidiser, acid↔caustic, one-biocide) are always enforced; PID gains and setpoints stay site-uncalibrated. An operator can also bring a real historian in Upload, reproduce their real controller in Create, and Debug it against their trace (including the custom builder's PID, time-proportional and flow-proportional loops): the same deterministic relay/interlock logic runs over their measured data and flags where it diverges from control intent (leads, not a diagnosis). Hardware deploy remains out of scope. A debug run now also maps which empirical coefficients that trace could calibrate — the on-ramp to Score the engine (/sim/validate) and Propose calibrations (/sim/calibrate), real-data-only, fitting nothing and changing no status until a proposal is reviewed and applied. The calibration harness is now COMPLETE for every empirical simulator coefficient: alongside the ORP model and the corrosion-rate scale, fitters exist for inhibitor consumption (inert-tracer NSI mass balance), oxidant residual decay/demand, and control-loop response characterisation. Two honesty bounds hold: biocidal EFFICACY (microbial kill) is not fittable from any modelled channel — it needs microbial counts the trace cannot carry, so only the oxidant-demand kinetic is fit, never efficacy; and the loop fitter characterises tuning (oscillation, decay, IAE), it does not claim the configured PID gains, which are not identifiable from the pH signal alone. Every coefficient still reads UNCALIBRATED until a REAL trace is fit — synthetic data exercises the fitter math but can never calibrate. An uploaded historian CSV can now map its measured-corrosion columns (mild-steel / copper mpy) straight to the calibrator: with no per-step engine prediction on a real trace, the corrosion-scale fit corrects the sim’s STIPULATED base rate (steel 1.0 / copper 0.2 mpy) against the measured average, and reports which reference it corrected — still a PROPOSAL, never auto-applied. Uploads of unknown shape are handled by a robust server-side parser: it sniffs the delimiter (comma/semicolon/tab/pipe), decimal mark, BOM and quoting; normalises units the header names via a per-channel unit registry — conductivity (mS/cm, S/m, dS/m, µmho/cm), temperature (°F, K), flow (gpm, m³/h, L/s), pressure (psi, bar), corrosion rate (mm/y→mpy, so a calibration is never silently wrong), concentration (ppb, ppm), ORP (V) — applied against the FINAL confirmed mapping (so a hand-mapped column converts too, and the operator can OVERRIDE any detected unit in a review-and-confirm step before ingest); and PROPOSES a channel mapping from a recognised-name registry. Anything it cannot resolve — an unrecognised column, an unknown unit, an ambiguous one (ppm↔mg/L is only an assumption, NTU↔FNU are different methods, TDS↔conductivity needs a site factor), a long/tall export — is surfaced for the operator to confirm, never guessed. The same dialect robustness (delimiter/BOM sniff) is shared by the water-analysis importer; manual add-a-parameter entry collects its own unit. It converts the common formats to one canonical structure; it does not claim to handle every possible layout, and says so where it cannot. and a detector suite (leak, sensor FDI, corrosion attribution (mechanism named from deterministic chemistry — Larson-Skold, LSI, pH — while the RATE stays UNCALIBRATED), inhibitor consumption, fluorometer fault, turbidity, and transient sensor glitches: missing/dropout, spikes and out-of-physical-range reads) — all HEURISTIC. Where a leak and a fouled fluorometer share a signature, the detectors disambiguate by pattern (sudden step vs gradual drift) and say ambiguous rather than guess.

Halogen speciation (Morris 1966), holding time, demand-corrected residual from Standard Methods 2350, dose-for-residual, the bio-dispersant slug window, and the phosphorus ledger. Plus an ASHRAE 188 control-point register fed live from Engine 1’s deposit and corrosion numbers. It does not predict Legionella — no validated bulk-chemistry-to-CFU model exists, and the register does verification, never validation.

The inverse problems, run against the same analysis, split into four offerings over one shared makeup water: Diagnose (cycle limit, sensitivity, binding constraint), Prescribe (cycles balance, programme cost, operating setpoints, independent audit), Operate (the unified twin, control decision, sandbox), Learn (calibration). Its point is cross-validation: thermodynamics predicts a cycle ceiling, your logbook measures one, and the two are shown side by side. The console opens on Diagnose with a worked example already loaded and analysed, so the first screen shows results rather than an empty form; the deterministic reads run automatically and refresh whenever the water changes. Diagnose serves both cooling and reverse osmosis: a system toggle switches the cycle-limit read from evaporative cycles to RO recovery, evaluated at the membrane wall (CF × β) where scale nucleates — the same limiting-salt logic, the correct concentration physics for each. Tools are gated to the segment they fit: the cooling/RO cycles-of-concentration reads (cycle limit, sensitivity, cycles balance) are marked not-applicable on sour, acid-drainage and boiler water — where there is no concentration model — and the segment-native read (Segment chemistry) is surfaced instead, so a non-cooling water never shows a cooling answer. The segment-adaptive tools (programme cost, operating setpoints, binding constraint) carry the segment through to the engine and stay available everywhere. The Operate tools and the independent audit are now segment-aware too: for sour, acid-drainage and boiler the unified twin, control decision and audit run on the validated segment reads (never a cooling answer), and the closed-loop sandbox — an intrinsically cooling-tower model — is honestly declined for those segments rather than shown a fabricated trajectory. The water can be built by hand or imported from a lab analysis; an imported value is validated at the boundary and tagged caller-supplied. Import also accepts any vendor’s CSV export (WAVE, French Creek, IMSDesign, a lab) and OLI Cloud/Process API JSON (parsed to OLI’s public documented spec — params.inflows/temperature; OLI inflows not in mg/L are refused, not mol→mg/L guessed): recognised columns/ions auto-map and the operator maps any unrecognised column onto a field themselves — so no proprietary format is ever guessed, and values are taken as-is (no unit/basis conversion is claimed). Tool-SPECIFIC native binary formats (WAVE/IMSDesign) remain sample-gated.

Bounds general-water corrosion by mechanism — the honest complement to the measured coupon and the CO₂ envelope. O₂ is the only deterministically boundable one: the oxygen mass-transfer limiting current i_lim = n·F·k_m·C_O₂ converted by ASTM G102 (0.457 mpy per µA/cm², iron) gives an UPPER BOUND that scales ~linearly with dissolved O₂ and rises with flow, labelled UPPER BOUND (never an estimate), with a low-chloride passivation caveat so it never under-predicts in chloride water. UDC has no predictive rate model: a Larson-Skold + differential-aeration susceptibility and an acceleration-factor framing, with an explicit measure-this gap — no rate. MIC is monitored, not predicted: a likelihood flag + a labelled published case-history band + a hard measurement gap (NACE/AMPP TM0212/TM0194) — no rate. One rule across all three: bounds, labelled bands and gaps only, never a fabricated point rate for the true O₂ rate, UDC or MIC.

Sizes the condensate polisher RUN LENGTH, on the ix_sizing capacity-is-input seam. Deep-bed mixed-bed: throughput = resin volume × rated capacity ÷ influent ionic load; service = throughput ÷ flow — exact stoichiometry (meq/L is numerically eq/m³, unit-pinned). Powdered-resin precoat: the run ends at whichever comes first — ionic exhaustion OR the crud/ΔP filtration limit — and the tool names the limiter. The honesty seam: the resin’s rated capacity (and the precoat ionic + crud-holding capacities) are datasheet INPUTS, never invented — absent, an explicit gap; and the achieved effluent quality (ppb Na/silica/cation conductivity) is a column-trial gap, never a predicted ppb. It sizes the run; it does not dose or guarantee a residual. Boiler.

Turns the fouling fingerprint into a CIP plan — selection and sequence, never a dose. Deterministic from published RO CIP practice: carbonate scale / metal oxide → acidic (low-pH); silica / organic-NOM / colloidal / biofouling → alkaline (high-pH); sulfate scale → alkaline + chelant; and biofouling takes a NON-oxidising biocide — never an oxidiser on polyamide, which destroys the membrane. When foulants are mixed the alkaline clean runs first (lift organics/biofilm/colloids), then the acid (scale/metals). The honesty seam: the specific product, concentration, pH setpoint, temperature and soak time — and the membrane’s own pH/temperature LIMITS — are the vendor cleaning-guide gap, never emitted (exceeding a limit destroys the membrane). It selects and sequences; it does not dose. RO.

Sizes dechlorination ahead of the membranes — free chlorine destroys polyamide, so the feed must be dechlorinated. GAC: bed volume = flow × empty-bed contact time (a labelled design EBCT, overridable). SMBS: dose from exact stoichiometry — Na₂S₂O₅ → 2 NaHSO₃, 1 mol SMBS per 2 mol Cl₂ = 1.341 mg SMBS per mg free Cl₂; the practice excess factor (~1.8–3×) is applied only when set and is labelled, not baked in. The honesty seam: carbon exhaustion / breakthrough needs the carbon’s capacity (iodine number / a column trial) — a gap, never fabricated; and the dose is a design figure to confirm with a downstream ORP interlock, not a guaranteed zero residual. RO.

Monetises the platform’s vendor-neutrality: an independent provider hands a client an audit-ready OUTCOMES report under its own name. Reuse over rebuild: it COMPOSES the honesty-gated compliance_report + roi_eroi + esg_water_disclosure and adds no number of its own — each section keeps its confidence and gaps. Two ways in: pull from the tools already run this session, or direct entry of the headline value/ESG figures. Formulation-blind by construction: any product / brand / formulation / trade-name / SKU / recipe identifier is stripped and listed; outcomes are attributed to "the programme", never a named chemistry. A missing section is a declared gap — compliance is NEVER assumed compliant — and white-label requires the provider’s identity, never fabricated.

The cooling analogue of the RO fouling twin. thermal_performance gives the fouling factor R_f at one point; this takes a TIME SERIES of R_f and projects WHEN it crosses the cleaning threshold. Deterministic: a least-squares trend on the OBSERVED R_f with an R² noise guard (noise is never called a decline), projected to the ENTERED cleaning threshold — a cleaning-due lead time. Reuse over rebuild: it imports the RO twin's _lsq / _project_crossing / _R2_TREND (the same fit + noise guard as the RO twin and sensor-drift). Honest gap: the mechanistic fouling RATE from the water chemistry / deposit type needs a deposit analysis and is never fabricated. Threshold is entered, never invented. Cooling.

How much silica the boiler water may hold at a given drum pressure to keep steam clean. Cited & deterministic: the ASME/ABMA boiler-water silica LIMIT falls with rising pressure (150 ppm low-pressure → 0.5 ppm above 2000 psig), with the steam target (≤20 ppb; many turbines ≤10 ppb); and the pH DIRECTION on vaporous carryover (Babcock & Wilcox anchors: 0.88 at pH 10, 0.74 at pH 11 vs pH 8.8; negligible below ~pH 9). Directional only: a coarse order-of-magnitude vaporous band by pressure (back-calculated anchors, not a calibrated ppm). Honest gaps (never fabricated): the EXACT vaporous ppm is not computed — the only closed-form correlation (Bahadori 2010) needs a full 32-coefficient table that is not openly available; and MECHANICAL carryover is plant-specific (drum/separator design, level, load, foaming) — confirm TOTAL carryover by isokinetic steam-purity sampling (IAPWS TGD1-08). Boiler.

For waters that carry dissolved sulfide (O&G sour), low-pH metals (mining acid drainage), or dissolved oxygen (boilers), the engine runs the WATEQ4F thermodynamic database to evaluate the phases those segments care about: barite and iron-sulfide scaling, aluminium-hydroxide, alunite and jarosite, and oxygen redox. It appears only when the water carries those species — a normal cooling or RO water shows nothing here, by design. The water form asks for the defining species of the chosen segment (sulfide for sour, the Al/Mn metals for acid-drainage, dissolved O₂ for boilers), so the segment read resolves on the composition it actually needs rather than silently returning nothing. The saturation thermodynamics are validated against the database; a phase that is thermodynamically favoured but kinetically too slow to form now (Pyrite, Jarosite, Alunite) is flagged as a watch-indicator, never reported as present-day scale. What is not yet earned: per-site inhibitor dose calibration for these segments — the numbers say what the water wants to do, not what a specific treatment will cost.

The entry point for acid mine drainage: net acidity — not pH — decides whether a water self-neutralises or needs added alkalinity, because a circumneutral water can carry high dissolved metal (latent acidity) that appears as the metals hydrolyse. Deterministic Kirby-Cravotta / Hedin calculated acidity (proton acidity from pH + the Fe/Mn/Al load by charge) minus your measured alkalinity, with the per-source breakdown. If total iron is given it is treated as ferrous — a labelled default; split it into Fe(II)/Fe(III) for the exact value. The lime figure is the stoichiometric CaO demand; real plant demand is higher by a plant-specific utilisation factor that is not fabricated here. Recovery % and product purity are not computed, and sludge volume needs an entered percent-solids — but the sludge dry mass is now computed (the Sludge generation tool); the remaining items are a stated AMD gap: this sizes the chemistry demand, not the plant. AMD-only by construction — shown on the acid-drainage segment.