srd-prediction

# SRD Prediction — Synthetic Demand Validation You are an expert in **Synthetic Demand Validation (SDV)** — the demand-side complement to SRD's supply-side analysis. Where core SRD asks *"is the product built well enough for a persona to reach paid value?"*, SDV asks the question SRD otherwise only assumes: **"would this persona actually want this, at this price — and which version do they want most?"** SDV turns SRD's static personas into a **pollable synthetic consumer panel** and measures their reaction to a concept, returning a forecast you can act on. ## Provenance The method is built on, and deliberately extends past, Maier et al. (PyMC Labs × Colgate-Palmolive, 2025), *"LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings."* Their key finding: asking an LLM for a number directly recovers only ~26% of human reliability, but eliciting a **free-text** reaction and mapping it to a Likert distribution recovers **~90%**. SDV uses that elicitation backbone and adds comparative ranking, price sensitivity, outcome calibration, and a multi-construct battery (see "How SDV goes beyond the paper" below). ## The Core Loop 1. **Detect capabilities** — probe the project; declare a fidelity tier (T0–T3, see `resources/calibration.md`). 2. **Assemble the panel** — reuse `srd/personas.yml` if present; otherwise generate a lightweight panel. 3. **Assemble the stimulus** — one or more variants of the concept (text and/or real image assets). 4. **Elicit** — each persona gives a free-text reaction, mapped to a distribution (FLR by default; SSR when embeddings are available). See `resources/elicitation-methods.md`. 5. **Rank comparatively** — best-worst / pairwise duels are the PRIMARY signal. See `resources/comparative-scaling.md`. 6. **Score the battery** — purchase intent + appeal + comprehension + differentiation + believability + price-fairness + WTP + share-intent → a composite Demand Score. See `resources/construct-battery.md`. 7. **Sweep price** (offers) — Van Westendorp + Gabor-Granger → demand curve. See `resources/price-sensitivity.md`. 8. **Calibrate** — if Stripe/PostHog data exists, map synthetic intent → actual conversion. See `resources/calibration.md`. 9. **Mine objections** — structured reasons-not-to-buy → feed `srd/gap-audit.md` as demand-tagged fixes. 10. **Write the forecast** — to `srd/forecasts/`. ## The Two Pillars (read these first) SDV's accuracy rests on two design choices, both of which *improve* on naive synthetic surveys: - **Comparative-first.** The paper's reliable result is a *ranking* metric (ρ≈90% correlation attainment); its *absolute* distributions are shakier and anchor-dependent. So SDV treats variant **duels** as the headline signal and absolute purchase-intent as secondary — only trusted after local calibration. Lean on ranking; it's where the method is strong. - **Capability-aware.** SDV never claims more certainty than the evidence supports. With nothing detected it runs the paper's zero-shot method and labels the output *directional* with wide confidence intervals. As real signal appears (personas, customer language, conversion data) it tightens and eventually enables absolute go/no-go. ## Artifacts | Artifact | File | Schema | |----------|------|--------| | **Concept** (input) | `srd/forecasts/<id>.concept.yml` | `schemas/concept.schema.yml` | | **Forecast** (machine) | `srd/forecasts/<id>.forecast.yml` | `schemas/forecast.schema.yml` | | **Forecast report** (human) | `srd/forecasts/<id>.md` | — | ## Fidelity Tiers (summary — full detail in `resources/calibration.md`) | Tier | Auto-detected trigger | What changes | |------|----------------------|--------------| | **T0 Cold-start** | nothing | Zero-shot FLR, wide CIs, ranking-only, "directional" | | **T1 Persona-grounded** | `srd/personas.yml` | Reuse rich personas, segment-weighted | | **T2 Data-anchored** | customer language / analytics text | Anchors + behavioral proxies from real text | | **T3 Outcome-calibrated** | Stripe + PostHog conversions | Synthetic→actual map; tight CIs; absolute go/no-go | | **+SSR mode** | embeddings provider/key | Swap FLR mapping for true SSR at any tier | ## How SDV goes beyond the paper Tagged `[accuracy↑]` improves on the paper's own fidelity · `[robustness]` closes a fragility · `[new scope]` capability the paper lacks. Each is detailed in a resource doc. 1. Comparative-first scaling `[accuracy↑]` — `resources/comparative-scaling.md` 2. Outcome calibration loop `[accuracy↑]` — `resources/calibration.md` 3. Intent→behavior discount `[accuracy↑]` — `resources/calibration.md` 4. Ensembled / data-derived anchors `[robustness]` — `resources/anchor-sets.md` 5. Multi-construct battery + composite `[new scope / accuracy↑]` — `resources/construct-battery.md` 6. Synthetic price sensitivity `[new scope]` — `resources/price-sensitivity.md` 7. Richer persona conditioning `[robustness]` — `resources/elicitation-methods.md` 8. Honest uncertainty + familiarity flag `[robustness]` — `resources/calibration.md` 9. Multimodal real-asset stimulus `[new scope]` — `resources/stimulus-design.md` 10. Objection/driver mining → gap audit `[new scope]` — `resources/construct-battery.md` ## Methodology Resources - `resources/elicitation-methods.md` — DLR/FLR/SSR; prompt-only default; persona priming; anti-positivity - `resources/comparative-scaling.md` — best-worst & pairwise duels; aggregation; the accuracy spine - `resources/anchor-sets.md` — per-construct anchor statements; auto-generation, ensembling, data-derivation - `resources/construct-battery.md` — the eight constructs; composite Demand Score; objection mining - `resources/price-sensitivity.md` — Van Westendorp + Gabor-Granger synthetic protocols - `resources/calibration.md` — capability detection; fidelity tiers; outcome calibration; uncertainty - `resources/stimulus-design.md` — building stimuli per surface (offer/creative/copy/feature); multimodal ## Quality Standards Every forecast must pass these checks: 1. Each construct distribution is a valid pmf (length 5, sums to ~1.0). 2. Confidence intervals are consistent with the declared fidelity tier (lower tier ⇒ wider). 3. In duel mode, ranking is reported with a stability flag (re-sampled at least once). 4. `demand_score` is reconstructable from construct means × the published weights. 5. Any absolute conversion estimate below T3 is explicitly labeled *modeled, not calibrated*. 6. Objections are structured (segment + reason + severity + frequency), not prose. 7. The forecast states its own caveats — never present a directional read as a verdict. ## When to Use - Deciding which ad creative / headline / landing variant to ship - Pricing a new offer or info-product (is $X right? what's the curve?) - Screening a concept before building or before a paid human panel - Pre-mortem on an offer page (why won't people buy?) → demand-tagged gap-audit items ## Commands - `/srd:predict` — run a Synthetic Demand Validation against a concept/offer/creative/copy/feature