PRD v4.0 — July 2026 · Pre-pilot. No signed commercial agreement.
PETA-AI reads an entire planning corpus and finds the coherence failures inside it — contradictory targets, unfunded commitments, misaligned regional plans — in days rather than the 90–365 days a manual review cycle takes. Then it forecasts what an unsigned draft policy will break downstream, before money or political capital is committed.
Pages indexed
9,412
corpus target 8,000–12,000
Canonical programs
1,286
ontology target 1,200–1,500
Entity resolution
87.4%
go/no-go gate >85%
Median time-to-flag
5.2 days
objective <7 days
the problem, quantified
Indonesia's national plan requires coordinated execution across 34+ ministries, 34 provinces and 514 districts over 60 months. The monitoring & evaluation function scored 22/100 in an independent 2024 assessment — the lowest-scoring governance function measured — and the next comprehensive check is a single mid-term evaluation in 2027.
KRISNA, One Data Indonesia and the 2026 Digital Government Master Plan all assumed that centralising data would produce coordination. KRISNA's own 2019 internal audit recommended anomaly detection at program-goal level — still unimplemented five years later.
One Data Indonesia has grown to 453,865 datasets with documented interoperability gaps still unresolved. More data has not produced coherence.
No existing system reasons about the relationships between planning documents — only within them. That is a harder problem than aggregation, and it is the one PETA-AI is built to solve.
detection, from the demo corpus
No finding triggers an automated action. Decision authority stays with human analysts — required by Indonesian planning law, and the trust mechanism that makes adoption possible at all.
Two ministry strategic plans commit to opposite directions on the same land base: one expands irrigated hectarage, the other freezes conversion of the parcels that expansion depends on. Semantic entailment marks these as directionally incompatible, not merely differently worded.
A national headline target has an owning program in the strategic plan but no traceable allocation in the corresponding budget document for the first two fiscal years. Flagged as an unfunded commitment rather than a reporting lag because the program code is absent, not zero-valued.
Aggregated provincial plans sum to materially less than the national renewable-capacity target. Divergence is concentrated in six provinces whose plans reference an earlier national figure.
architecture
Ingests the full planning corpus (~8,000–12,000 pages) and builds a normalised index. Core challenge is entity resolution: the same program appears under different names across ministries and cycles.
Six specialised agents, sequenced by build priority. The fundable MVP is the two P0 agents, not all six in parallel.
Agents never message each other. Each writes findings into a shared graph: nodes are programs, edges are detected relationships.
Layers 1–3 detect contradictions after publication. Layer 4 forecasts what a draft policy breaks downstream, before money or political capital is committed.
Every output carries source passage or model basis, confidence score, severity, and a recommended action. No finding triggers an automated action.
product objectives
Speed
Cut time-to-flag for a coherence failure from a 90-day quarterly review cycle to under 7 days from document publication.
Precision
>70% precision on contradiction detection, validated against official government evaluation records.
Prioritisation
Rank programs by structural criticality so analysts know which 20% of programs are load-bearing for 80% of national targets.
Trust
Every finding ships with source-passage citation, confidence score, and severity — required for analyst adoption, not optional.
Prediction
For a draft policy not yet deployed, forecast downstream coherence risk with a calibrated confidence interval.
Commercial validation
Publish a precision/recall/lead-time result set against Bappenas' 2027 mid-term evaluation as the flagship case study.