An 8-layer geopolitical consensus pipeline with tier-based parameter classification enables reliable real-time model updates without degradation

Hypothesis

An 8-layer geopolitical consensus pipeline with tier-based parameter classification enables reliable real-time model updates without degradation

Changelog

Date	Summary
2026-04-06	Audited: added Changelog, domain tag quant-finance, stamped last_audited
2026-03-18	Initial creation

---

Hypothesis

An 8-layer geopolitical consensus pipeline with tier-based parameter classification enables reliable real-time model updates without degradation. The v16 model required manual parameter updates: a human would review new data, decide which parameters to adjust, run a full Monte Carlo simulation, and verify the results. This process took 2-4 hours per cycle and could only run once or twice per day. The hypothesis was that a multi-agent architecture with layered safety validation could automate this process at hourly cadence without ever degrading model quality below the current baseline.

Method

v17 introduced a complete real-time architecture with three major components: sub-agent data collection, 8-layer validation, and parameter tier classification.

6 parallel sub-agents (all hourly at :17): WTI Price (FRED/exchange), Polymarket Odds (API), Hormuz Traffic (AIS vessel tracking), Military Intel (OSINT), Diplomatic Intel (news + statement analysis), Supply/Economic (EIA/OPEC/macro). Each produces a structured evidence package with confidence scores, source reliability ratings, and temporal relevance markers.

8-layer validation pipeline:

Layer	Name	Function	Gate Criteria
1	Evidence quality scoring	Rate each evidence item on source reliability, recency, corroboration	Score < 0.3 → discard
2	Tier-based filtering	Classify parameters by observability	See tier table below
3	Anti-oscillation	Detect and suppress parameter flip-flopping	Block if param reversed direction within 3 cycles
4	Adaptive step sizing	Limit parameter change magnitude based on evidence strength	Max step = evidence_score × tier_max_step
5	Coherence constraints	Ensure parameter changes are internally consistent	Flag contradictions (e.g., escalation up + ceasefire up)
6	Forward sandbox	Run Monte Carlo with proposed changes, check 7-day forecast	R² must not drop > 0.5pp
7	Backward sandbox	Re-run on historical data with proposed changes	R² gate: must equal or exceed baseline. MAPE gate: must not increase > 0.1pp
8	Drift caps	Per-parameter maximum cumulative drift from initial calibration	Parameter-specific; prevents slow drift to unreasonable values

Parameter tier classification:

Tier	Observability	Parameters	Max Step/Cycle	Evidence Requirement
Observable	Directly measurable	WTI spot, PM odds, transit count	5%	Single source sufficient
Partial	Inferred from multiple signals	deEscSens, ceasefireHazard, tailRegime	2%	2+ corroborating sources
Forward-Only	Cannot be validated historically	militaryEscProb, diplomaticSignal	1%	3+ sources with high confidence

Forward-Only parameters have the tightest constraints: errors cannot be caught by the backward sandbox (no historical ground truth). Full cycle (:17 collect → :26 apply) runs in ~9 minutes, leaving 50+ minutes of slack per hour.

Results

Hypothesis confirmed. v17 has been running in production since March 18 (v17.0) with a refinement update to v17.2 on March 24. The pipeline has processed multiple update cycles without any model degradation.

Cycle 3 detailed example (representative):

Metric	Pre-cycle	Post-cycle	Delta
R²	0.9610	0.9769	+0.0159
MAPE	1.85%	1.26%	-0.59pp

Cycle 3: 15 updates proposed, 12 accepted, 3 rejected (1 backward sandbox R² drop 0.7pp; 1 anti-oscillation reversal; 1 low-confidence single-source Forward-Only parameter).

Production stability (v17.0-v17.2): 8 active parameters stable across all cycles. Backward sandbox rejects ~20% of proposed changes; anti-oscillation blocks ~7%. Zero model degradation events. v17.2 changes: tightened drift caps, added GL-134 Russian sanctions offset.

Cumulative improvement chain (v13 through v17):

Metric	v12 (baseline)	v17.2 (current)	Total Delta
R²	0.9498	0.9634+	+1.36pp
MAPE	1.97%	1.34%	-0.63pp
One-step MAE	$1.633	$0.833	-$0.800 (49%)
Direction accuracy	86.7%	93.8%	+7.1pp
OOS R²	0.8633	0.9257+	+6.24pp
Brier vs Polymarket	baseline	+9.93pp	+9.93pp

Findings

1. Backward sandbox (Layer 7) is the critical safety gate. ~20% of proposed changes fail here: plausible-looking updates that degrade historical fit. Without it the model drifts through accumulated overconfident updates.

2. Anti-oscillation prevents geopolitical whipsaw. News flips hourly (exercise = escalation; diplomatic statement = de-escalation). The 3-cycle lockout forces parameter commitment before direction change, stopping flip-flopping on noise.

3. 1% step cap on Forward-Only parameters is essential. Military and diplomatic parameters cannot be backtested. 1%/cycle prevents large unvalidatable bets while allowing meaningful drift across many cycles.

4. GL-134 integration (v17.2) validated architectural extensibility. Adding the Russian supply offset required only a new sub-agent feed: no structural changes. New geopolitical factors plug in without refactoring.

5. Cumulative v13-v17: 49% MAE reduction, +9.93pp Brier vs Polymarket. From broken tail calibration and manual 2-4h update cycles to autonomous hourly operation beating the prediction market 82% of the time.

Next Steps

The v17 architecture is stable and running in production. Near-term: monitor GL-134 expiry (April 11) for extension/non-extension parameter adjustments; evaluate expanding the 8 active parameters to cover Venezuelan production and Libyan instability; consider a 9th validation layer for cross-model consistency against a simple regression baseline. See experiments/oil/2026-04-04-timesfm-plateau-detection for the next iteration.