GRIP — the decision engine

Why probabilistic, not deterministic

Pharma operational decisions are not exact: a tender outcome is not "yes" or "no" but a probability conditioned on demand, carrier capacity, lane history and competitor pricing. A supplier's reaction to a counter-offer is not certain but a distribution over Accept / Counter / Walk. A batch's remaining shelf-life after a temperature excursion is not a single number but a confidence interval driven by Arrhenius kinetics and historical stability data.

GRIP returns the full distribution — and recommends an action against it with the operator's risk tolerance applied. This means:

• You can budget against the 80^th-percentile worst case, not the mean.
• You can explain to auditors why a borderline batch was released or held — with the math, not "the model said so".
• You can A/B compare two negotiation strategies with quantified outcome variance.

The three pillars

Sibling: the Quality module's Arrhenius engine

The Quality module ships with a separate, dedicated probabilistic engine — not GRIP — built around Arrhenius kinetics for shelf-life impact assessment after temperature excursions. The math is the same general primitive (k(T) = A · exp(−E_a / RT) integrated over the actual sensor profile, with Monte Carlo uncertainty propagation), but the engine, the parameter library and the validation pack are independent. We split them deliberately: Quality runs on the QP's desk and is validated GAMP 5 Cat 4 against IQ/OQ scripts focused on excursion review, while GRIP runs on the operational data hub for Transport + Negotiation decisions with different acceptance criteria.

Full mathematical walkthrough lives in the Arrhenius shelf-life modelling whitepaper — activation-energy sourcing, sensor calibration error propagation, spatial inhomogeneity adjustments, when peak-temperature dominates vs cumulative exposure, and a worked 6-hour excursion at 14°C with full distribution output.

Example 1 — Bayesian price-elasticity in the Negotiation module

Each supplier has a latent price-elasticity distribution that GRIP estimates and updates after every interaction. Prior comes from category benchmarks (Fastmarkets, ICIS, LME, Argus); posterior is updated via Bayes' rule after each round of price exchange:

This is why the Negotiation module's "recommended counter-offer" comes with a calibrated expected acceptance probability — not a gut guess. After 3-5 rounds with a supplier, the posterior converges and recommendations become tighter.

Example 2 — Monte Carlo for tender coverage in the Transport module

Given the production plan and active tender lanes, what's the probability that you cover 95% of your transport demand over the next quarter without spot freight exposure?

GRIP runs 10,000 Monte Carlo simulations sampling from internal distributions:

• Production plan stochasticity (slip probability per SKU)
• Lane-level award acceptance from each carrier (historical, ML-trained on your tender outcomes)
• Demand volatility per temperature class (cold-chain vs ambient)
• Spot rate inflation distribution (carrier × month)

External signal layer

The Monte Carlo sampler also pulls live external feeds into the input distributions, so every recommendation reflects the actual current operating environment, not a stale baseline:

• Fuel-price indices — Brent / ULSD / diesel by lane region, hedging-window timing for fuel-surcharge exposure.
• Commodity benchmarks — ICIS, Fastmarkets, LME, Argus for accessorial and surcharge baselines.
• Geopolitical disruption events — port congestion (Rotterdam, Antwerp, Hamburg), Red Sea / Suez transit times, sanctions impact on specific trade routes, customs strike duration.
• Capacity-tightness signals — spot-vs-contract spread, carrier on-time-pickup rates, seasonal volume curves.
• Weather / mode-shift — North-Atlantic storm season effects, air-freight overflow when ocean lanes congest.

Output: distribution of "coverage rate × spot exposure cost × landed cost" outcomes, conditioned on the current external state. Transport then surfaces the lanes contributing most to the tail risk (worst-case 5%) and recommends a mini-tender for them before invoices reconcile. Every recommendation carries the input-data lineage hash, so the auditor can replay the exact decision with the exact external snapshot that informed it.

Auditability — model versions in every decision

Every decision package output by Terminus includes:

• GRIP engine version (semver)
• Model parameter set version (per-tenant)
• Rule set version (your SOPs)
• Policy layer version (your risk tolerance config)
• Input data lineage hash
• SHA-256 hash of the full PDF + JSON output

When an auditor asks "why did the model recommend X two years ago?" — we replay the exact GRIP version with the exact same inputs and get the exact same output. Bit-for-bit reproducibility is the operational requirement, not a nice-to-have.

What GRIP is NOT

• Not a large-language-model. No LLM is in the decision loop. LLMs are used elsewhere (e.g., parsing free-text excursion notes into structured fields) but their output never feeds the decision recommendation.
• Not a black box. Every probability distribution can be inspected, every rule is human-readable code, every recommendation traces back to its inputs.
• Not a replacement for QA / QP / procurement judgment. GRIP recommends; humans approve. The approval is part of the decision package.
• Not magic. Quality of recommendations is bounded by quality of inputs. Garbage in → calibrated-garbage out. We help you fix the inputs.