Single-customer dependency on Bappenas
Run multilateral channel conversations (World Bank GovTech, ADB, GIZ) in parallel from Phase 1, not after Bappenas succeeds.
section 7 — go-to-market & roadmap
Each phase carries a commercial milestone, not just a technical one.
Phase 1 — Foundation
Months 1–4
Corpus + ontology + Layer 1 live. Commercial goal: convert the incubation relationship into a signed, funded pilot (grant or donor co-funded) with Bappenas.
Phase 2 — MVP Swarm
Months 5–8
Contradiction Detector + Budget Coherence Agent validated against Bappenas' public 2024 evaluation. Commercial goal: first published precision/recall numbers.
Phase 3 — Full System + Expansion
Months 9–12
Full swarm + dashboard live; sealed predictions logged ahead of the 2027 evaluation. Commercial goal: at least one active second-government or multilateral conversation underway.
Phase 4 — Predictive Layer
Graph-dependent (from mid-to-late Phase 2, through month 14)
Gaussian Process forecasting first, then the survival-analysis upgrade, then causal/ABM cascade simulation last. Commercial goal: the KRISNA retrodiction demo ready for investors and pilot buyers.
Sequencing note: the predictive layer cannot meaningfully start before the Coherence Graph has enough real, weighted relationships to model. A causal or ABM simulation run on a sparse graph produces noise, not forecasts. Phase 4 is graph-dependent, not calendar-parallel.
8.1 Business & political risk
Political / leadership change risk
Anchor the relationship and funding to the institution and donor, not an individual official.
Long sales cycle vs. runway
Treat govtech as one of two revenue tracks rather than the sole path to revenue.
Single validation event (2027)
Publish interim results against the already-public 2024 evaluation well before 2027.
8.2 Technical & model risk
Entity resolution underperforms on inconsistent Indonesian bureaucratic language
Maintain a manually verified anchor set of canonical program names; treat >85% entity-resolution accuracy as a go/no-go gate before Layer 2 agents are trusted.
Causal model assumptions don't hold in a politically dynamic system
Be explicit that causal estimates are directional and confidence-scored, not deterministic; pair with the ABM layer as a cross-check.
ABM/cascade simulation poorly calibrated with only two prior RPJMN cycles
Ship cascade simulation last and validate pattern-level rather than exact-magnitude agreement.
Hallucination / unsupported claims in a government-facing output
Two-stage retrieval + mandatory source-passage citation on every output; no finding ships without a traceable source or model basis.
section 9 — current status & open gaps
Stated plainly, because overclaiming is the fastest way to lose a technical evaluator.
Selected into the UK–Indonesia AI Incubation for Public Sector 2026, Bappenas track — not yet a signed commercial engagement.
No LOI or paid pilot commitment secured yet — highest-priority open item.
SAM/SOM figures are working estimates pending a validated target-country list.
Team is a single technical/product founder; likely needs a named technical co-founder before a priced round.
Build status per layer/agent needs to be stated accurately before external sharing.
Competitive landscape needs one primary-research pass for direct AI-govtech-coherence competitors.
Forecast backtesting has not been run — every confidence number shown externally is provisional.