Innovation Board

How it works

• Orchestrated by code: the page drives a fixed turn order (three opening assessments, an audit with three responses, the financial case in two steps, then the Chair) with a hard cap of ten turns. Each turn is one request to a Cloudflare Pages Function.
• One agent, one compact prompt per turn: the Function builds a short prompt for DeepSeek (OpenAI-compatible API, small output limit, thinking disabled) from the company profile, the evidence pack and the run state. Visitors only pick an idea from a list, so no free text reaches the model.
• No server session: the run state travels with each request, signed with an HMAC. The server validates its structure and signature before acting, so altered claims, estimates or assumptions are refused.
• Guardrails in code, not in the prompt: every quote must be a real substring of the cited source; any figure an agent states must appear in a source that passed the screen or be computed by code in this run, or the statement is struck as an “unverified figure”; estimates and assumptions are range-checked; code computes every score, estimate and financial result and applies the investment policy.
• Never fails because of the LLM: if the key is missing, the provider errors or a turn is invalid twice, the page replays a recorded live run through the same checks and computations, and the badge says “Recorded run”.

Formulas (agents propose, code computes)

• Dimension score (strategic fit, market pull, feasibility, risk) = 20 × mean of the accepted 1-to-5 scores for that dimension, or 0 if none
• Evidence strength = 100 × (accepted ÷ claims made) × Σ q · score ÷ (5 · Σ q), over accepted claims, with source quality q = 1 primary, 0.6 secondary, 0.2 promotional
• Financial return = clamp(40 + 30 × base-case ROI, 0, 100)
• PERT per work package: E = (o + 4m + p) ÷ 6, σ = (p − o) ÷ 6; total σ = √Σσ²; range = E ± 1.28σ (about 80% confidence)
• Duration in weeks = person-days ÷ (squad size × productive days per week); cost = Σ person-days × role day rate
• Customers: C₀ = 0, Cₜ = Cₜ₋₁ × (1 − churn) + adoption; revenue or saving in year t = price × (Cₜ₋₁ + Cₜ) ÷ 2; net cash flow = revenue − running cost
• NPV = −development cost + Σ net cash flowₜ ÷ (1 + discount rate)ᵗ, t = 1 to 3; ROI = (Σ net cash flow − development cost) ÷ development cost
• Payback = the month in which cumulative cash flow, starting at −development cost, reaches zero (linear within the year)
• Scenarios: pessimistic and optimistic take each assumption from the cautious or favourable end of its evidence range and the high or low end of the cost range; the base case uses the Finance Analyst’s verified values and the expected cost. Sensitivity moves one input at a time across its range.
• Tiers, first match wins: Invest now if strategic fit ≥ 60, evidence strength ≥ 60, base NPV > 0 and payback ≤ 24 months; Run a pilot if fit ≥ 50, evidence ≥ 45 and base NPV > 0; Explore further if the optimistic NPV > 0; otherwise Park.

Limits

The figures illustrate a method on invented data and are not forecasts. The company, the ideas, the evidence packs, the rate card and every number are synthetic, and the research phase is simulated. A three-year model with four assumptions leaves out tax, inflation, financing, cannibalisation and the cost of the squad’s time elsewhere. The language model can still misread a source within the rules; the checks only guarantee that what counts is traceable to the evidence or to the code, not that it is right. A person makes every investment decision.

Source code and tests: github.com/nepryoon/innovation-board (opens in a new tab)