/ › expansion › 05-simulator-agi-horizon

Stream 05 — Simulator + AGI Horizon

Make every quantitative claim in the blueprint a distribution with provenance, and every AGI-horizon claim a triangulated synthesis with falsifiable triggers. v0 ships a working Monte Carlo on the Paraguay-Itaipú Anchor 0 Fase 1 financial structure plus a 13-source posterior over G/M/B scenario probabilities.

Why this is load-bearing

Bankability armor. A multilateral CFO or sovereign-fund analyst should be able to reproduce every load-bearing number in the blueprint from a manifest hash. v0 closes the gap from "single-point estimates with [ASSUMPTION] tags" to "P10/P50/P90 distributions with reproducible runs".
AGI horizon triangulation. Chapter 11's single-source [ASSUMPTION] on G/M/B scenario probabilities becomes a posterior over 13 sources (Aschenbrenner SA, Davidson/Shulman, Cotra, Metaculus, Manifold, AI Impacts, Epoch, Anthropic RSP, OpenAI PF, GDM FSF, METR, Apollo, MIRI). Result shifts G 35→31% and B 20→24%; M unchanged at 45%.
Stream substrate. Atlas (V2 fan, V8 Sankey, V9 country choropleth), Pulse (gate countdowns, KPI-attainment indicators), Onboarding (persona-tailored decade narratives) all consume this stream's deterministic JSON / parquet outputs.
Falsifiable triggers. METR Time Horizon ≥ 48 h at 50%, Epoch SWE-Bench-Verified ≥ 85%, and 6 other empirical thresholds replace narrative re-scenarization with measurable events.

Decisions taken

#	Decision	Rationale (short)
D1	Engine in Python (NumPy + SciPy)	Codebase is Python; vectorization meets perf targets; portable to Pyodide
D2	Storage = parquet + SQLite	Parquet for run detail; SQLite for queryable metadata, no infra
D3	Run hash = SHA256(manifest ‖ commit ‖ engine_semver)	Audit-grade reproducibility
D4	Default outputs are P10/P50/P90 + bands	No point-estimate publication without uncertainty
D5	LLM narratives always downstream of a numerical run	Hallucination containment via tool-use scaffold + numeric audit
D6	AGI G/M/B = posterior over 13 priors	Replaces single-anchor `[ASSUMPTION]` with documented Bayesian aggregation
D7	Validator + red-team gate before audience release	Existing project guardrail; binding
D8	Depth-first on Anchor 0 Fase 1 first	Mirrors blueprint's depth-first strategic posture
D9	No paid data sources for v0	Truth-infra: every cell auditable from open sources
D10	Engine semver + changelog	Reproducibility across blueprint revisions

Performance — Anchor 0 prototype, observed

Workload	Target	Observed
1K Monte Carlo, single core	< 2 s	0.08 s
10K Monte Carlo + PRCC tornado on 3 KPIs	< 60 s	0.46 s
Reproducibility (same manifest → same run hash)	byte-identical	verified, two consecutive runs

Anchor 0 Fase 1 — observed distributions (10K runs, manifest v0)

KPI	P10	P50	P90	Mean
`capex_total` (USD M)	2,161	2,727	3,553	2,805
`capital_committed` (USD M)	1,770	2,236	2,931	2,305
`K.financing_gap` (USD M)	363	484	657	500
`K.dscr.anchor0` (year-5)	2.06	4.26	7.25	4.51
`K.npv.anchor0` (USD M @ 8% real)	−1,857	−835	+56	−874
`energy_cost` (USD/MWh)	36.0	40.8	46.3	41.0
`p_anexoC_by_cod`	0.37	0.41	0.45	0.41
`K.kpi_attainment_2032` (P meet)	—	—	—	2.5%

Sensitivity — top 5 drivers per KPI (PRCC, 10K runs)

KPI	Top driver	#2	#3	#4	#5
`K.financing_gap`	`REG.friction_multiplier` (+0.93)	`FM.private_match.realization` (−0.90)	`CX.HPC.greenfield` (+0.79)	`CX.cable.landing` (+0.29)	`FM.multilateral.share` (−0.24)
`K.dscr.anchor0`	`HS.offtake.tariff` (+0.95)	`REG.friction_multiplier` (−0.90)	`CX.HPC.greenfield` (−0.72)	`FM.private_match.realization` (−0.33)	`FM.bono_sur.share` (−0.31)
`K.npv.anchor0`	`HS.offtake.tariff` (+0.93)	`REG.friction_multiplier` (−0.92)	`CX.HPC.greenfield` (−0.75)	`CX.cable.landing` (−0.26)	`CX.energy.subestation` (−0.15)

AGI horizon — 13-source weighted posterior

Each source provides a probability distribution over AGI year. Mapped to G (post-2032), M (2029–2032), B (pre-2029) buckets and aggregated with weights normalized over credibility × recency × empirical anchor.

Source	P(G)	P(M)	P(B)	Weight
Aschenbrenner SA + LP	0.10	0.40	0.50	0.13
Davidson / Shulman (Forethought)	0.30	0.45	0.25	0.16
Cotra (bio-anchors)	0.55	0.30	0.15	0.11
Metaculus (median nov 2033)	0.40	0.45	0.15	0.15
Manifold	0.35	0.45	0.20	0.11
AI Impacts (researcher panel)	0.65	0.25	0.10	0.14
Epoch AI capability	0.30	0.50	0.20	0.17
Anthropic RSP v2.0	0.25	0.50	0.25	0.14
OpenAI PF	0.20	0.50	0.30	0.10
GDM FSF	0.30	0.50	0.20	0.11
METR Time Horizon 1.1	0.25	0.50	0.25	0.20
Apollo Research evals	0.20	0.55	0.25	0.13
MIRI shutdown posture	0.15	0.40	0.45	0.06
Posterior (normalized)	0.31	0.45	0.24	—

Comparison to chapter 11 default [ASSUMPTION]: G shifts 35 → 31%, B shifts 20 → 24%, M unchanged. Proposed blueprint update in proposed-blueprint-update.md.

Empirical re-scenarization triggers (replace `[ASSUMPTION]` in chapter 11)

Trigger	Source	Threshold
METR Time Horizon @50%	METR monthly	≥48 h amber, ≥168 h red — recalibrate
Epoch SWE-Bench-Verified	Epoch quarterly	≥85% amber, ≥95% red — recalibrate
Hyperscaler annual capex	SEC 10-Q aggregation	> USD 1 trn — recalibrate
Cluster individual size	Hyperscaler announcements	> 50 GW — recalibrate
USG / EU export weights	Federal register / OJEU	any restriction by jurisdiction — recalibrate
Frontier-lab nationalization	News + filings	any — recalibrate
Open weights = SOTA gap	Benchmarks	< 6 months — recalibrate
New Aschenbrenner / Shulman public update	situational-awareness.ai / podcasts	any — recalibrate

Sector cusp-curve — where each sector sits on the AGI cusp

Sector	Pre-cusp through	Crossing in	Post-cusp from	LATAM leverage
Energy (renewable + grid)	2026	2026–2030	2030+	High — pillar 12 substrate
Governance (admin AI)	2027	2028–2032	2032+	High — SAT, justicia, identidad
Agriculture (precision + biotech)	2028	2029–2032	2032+	High — Cerrado, Pampas, Andean
EV	2027	2027–2030	2030+	High — litio + assembly
Biotech (drug discovery)	2027	2027–2031	2031+	Moderate — clinical-trial corridor
Healthcare delivery	2028	2029–2033	2033+	Moderate — telesalud + bilingual
Robotics (humanoid)	2028	2029–2032	2032+	Moderate-low — minerals + Mexico assembly
Education	2028	2029–2033	2033+	Low — mostly globally fungible
Defense	2026	2026–2029	2029+	Low — non-alignment IS leverage
Finance (lending / payments)	2026	2026–2029	2029+	Moderate — pyme corridor

Handoff — what consumer streams get

Stream	Reads from	Browser-runnable?
Stream 3 — Atlas	`{run_hash}.parquet` (GW × scenario × year), `.sankey.json` (V8), `.country.json` (V9)	pre-rendered
Stream 4 — Pulse	`{run_hash}.json` summary + 6 trigger-threshold YAML rules	pre-rendered indicators; client-side interactivity for single-run
Stream 6 — Onboarding	LLM-as-simulator narratives (validator + red-team-gated) + per-persona V1/V3/V4/V9 charts	welcome narrative pre-rendered; "what-if" client-side via Pyodide (5–8 knobs)
Stream 1 / 9 — Audience surfaces	Tier-by-view matrix (citizen / minister / capital allocator / multilateral / researcher / NGO)	matrix in `99-handoff.md` §4

Documents

state/expansion/05-simulator-agi-horizon/00-charter.md — mission, scope, decisions, acceptance criteria
01-parameter-taxonomy.md — 28 parameters across AGI scenario / capex / financing / adoption / sector / gate / political-economy; correlation blocks
02-engine-spec.md — Monte Carlo + Iman–Conover correlation + reproducibility hashing; module layout; CLI surface
03-output-views.md — 10 view types feeding Atlas / Pulse / Onboarding / Audience surfaces; provenance footer mandate
04-llm-as-simulator.md — narrative engine with tool-use schema (read_run, compare_runs, cite, build_counterfactual_manifest), validator + red-team gate, numeric audit
05-runtime-prototype/ — working Python engine (simulator.py), Anchor 0 manifest (manifest_anchor0_v0.json), run_example.md with 10K-run output and PRCC tornado
06-agi-horizon-multi-source.md — 13-source synthesis; weighted posterior over G/M/B; per-source LATAM implications
07-decade-positioning.md — 2026 → 2035 × G/M/B × five dimensions (capabilities, compute, labor, governance, LATAM); KPI 2030 + 2035 outlooks per scenario
08-cusp-curve-by-sector.md — where each of 10 sectors sits on the AGI-cusp curve; LATAM leverage map
99-handoff.md — explicit contracts for Streams 3 / 4 / 6 / 1 / 9; trigger YAML for Stream 4
proposed-blueprint-update.md — proposed edits to chapters 04, 05, 11, 13, 14 + new chapter 16 (sector cusp); humans review and merge separately

Reproduce locally

cd state/expansion/05-simulator-agi-horizon/05-runtime-prototype
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python simulator.py validate --manifest manifest_anchor0_v0.json
python simulator.py run --manifest manifest_anchor0_v0.json --n-runs 10000 --out outputs/

# Output JSON lands at outputs/db114bb3…json with the
# distributions in the table above. Re-running yields the
# same run_hash (timestamps excluded).