Forethought · Pillar 2, Foresight
Decide what to test next.
Simulate consumer panels before you commission them. Generate 20 to 200 demographically-specified agents, run a full probabilistic unfolding, and explore the concept space with a draggable 3D puck. The fieldwork still matters. The bet you place on which fieldwork to run improves.
Method · Synthetic Consumer Panel Simulation
20 to 200 synthetic respondents, each one demographically specified.
You set the panel composition: age bands, income tiers, geography, category usage, dietary constraints, whatever screens you'd write in a real study. Forethought generates the agents, scores each product through four sensory modalities, then runs in-character follow-up interviews so you see what the panel thought about each option, not just how they rated it.
Panel composition
Specify the panel like a real screener: 80 respondents, 25-54, household income $75K plus, weekly category buyers, balanced gender, regional spread. Forethought generates against the spec.
Four-modality scoring
Every product is scored on visual appeal, texture, flavor, and aroma as separate channels. The panel's overall liking rolls up from the modality scores so you can see which axis drove the response.
In-character follow-up
After scoring, agents respond to open-ended probes in character. The output reads like a real focus group transcript, not a model dump. Every quote traces to the agent's specified profile.
Cross-product comparisons
Panel-level statistics on differences between products: mean liking, modality split, segment-level preferences, JAR distributions, and the rest of the standard sensory output set.
Grounded against your data
When you bring a THEUS workspace or an internal data set, the synthetic panel anchors against your real evidence base, not a generic consumer training corpus. Your panel sounds like your category.
Reproducible
Same panel spec, same product spec, same seed: same panel. Forethought's panels are reproducible artifacts you can re-run, share, and cite in a stage-gate review.
Method · Interactive Concept Space Explorer
A draggable puck inside a 3D tetrahedron, with sensory descriptions on the fly.
Once HUSK has fit the panel, the resulting concept space is interactive. Set four anchor products at the corners of a tetrahedron, drag a probe inside, and the system blends a sensory description, a predicted liking score, and a panel-level segment breakdown in real time. Then ask Forethought to find the optimum for a target segment, and the optimizer runs in seconds.
Four-anchor tetrahedral space
Place any four real products at the tetrahedron's vertices. The interior is the convex blend of those anchors, mapped into the unfolded sensory space.
Barycentric interpolation
The probe's position in the tetrahedron determines a barycentric coordinate. Sensory profile and liking are interpolated continuously, not snapped to the nearest grid point.
Nelder-Mead "Find Optimal"
Ask for the optimum within a segment, a price band, or a constraint set. Nelder-Mead search runs over the interior and returns a concrete location with a predicted profile.
Gemini-grounded sensory descriptions
The blended sensory description is generated by a Gemini model grounded in the interpolated profile vector. The descriptive language matches the category lexicon, not generic adjective salad.
URL-first state sharing
Every position, segment, and anchor configuration encodes into the URL. Share a state with a colleague by sending a link, no export, no screenshot.
QDA CSV import
Bring a real QDA dataset, drop the CSV, and the anchors fill from your data. Forethought reads standard panel formats out of the box.
Method · Attribute Vector Overlays
QDA attributes, mapped into the hedonic space your buyers actually navigate.
A descriptive analysis panel produces sixty attributes. Your consumer panel produces one number: did they like it. The bridge between those two worlds is a vector overlay, and Forethought builds it with the statistical care you'd expect from a methods paper.
- Ridge regression with LOOCV validation
The overlay regression uses leave-one-out cross-validation to choose the ridge penalty, which is the right answer for small-N sensory data where holdouts are expensive.
- Bootstrap direction stability
The attribute direction vectors are bootstrapped. Stability across resamples is reported. An attribute that points one way in 70% of bootstraps and the opposite way in 30% gets flagged, not plotted with a bold arrow.
- Benjamini-Hochberg correction
With dozens of attributes tested, false discovery is real. The overlays report Benjamini-Hochberg corrected significance, not the naive p-values that get researchers in trouble at a defense.
- Plotted into the unfolded space
The output is a clean overlay on the HUSK-fit perceptual map: which attributes drive liking, which drive disliking, which split segments, which are noise. The plot is publication-ready.
- Drives the optimizer
"Find the spot where creamy is high and bitter is low" works because the attribute vectors are in the space the optimizer searches. The same vectors feed the Concept Explorer's optimization.
Real Client Outcome · Anonymized
How a global dairy producer used Forethought.
A multinational dairy company commissioned Forethought to design and analyze a cheddar category appraisal. Eight prototypes, two benchmarks, and a competitor needed to fit inside a single panel budget that previously would have only covered three. Stage 0 selected the spread-optimal six, the synthetic panel produced the segment structure, and the formulation team had a target sensory profile for the next round of prototyping inside the same quarter.
"We compressed a four-month upstream screening sequence into six weeks. The decisions we walked into the stage-gate with were better, not just faster."
Pricing · Productized Engagements
Five points along a spectrum, priced so you can pick without negotiating.
Forethought is sold as a sequence of productized engagements, each with three-alternative pricing. Alt 1 is the lean version, Alt 2 is the typical scope, Alt 3 is the expanded version. The same three-alt structure runs through our enterprise engagements too. Pricing is in USD, signed via standard MSA, 50-25-25 milestone payments with a 10-business-day acceptance window per milestone.
Two ways to start.
Request a demo and we'll walk you through the concept space with a sample data set, or book a Pilot call and we'll scope a real engagement against your category and your candidate set.