ADH AgentDecisionHarness Static demo · no API

Explainable financial prediction

Run the forecast.
See why it wins.

Choose a financial task, use the default qualification-gated policy or a preset variation, edit candidate votes, and inspect the full decision trail.

INSTALL HARNESS Two commands to run locally
01
From the cloned repository pip install -e .
02
Verify the default harness adh replay examples/frozen/FH025_route.json
LIVE

Codex or Claude CLI is only required for live provider mode.

ADH://PREDICTION_CONSOLE RUN_0001
LIVE OUTPUT Post-Earnings Reaction
READY
Forecast authority postearn_auxiliary_text_expert_v1
Final forecast negative_reaction
Audit PASS
Authoritative evidence
CONFIG_FINGERPRINT
DEFAULT PATH

Historical performance determines whether a specialist may hold forecast authority. Runtime output explains why that authority was retained, transferred, or blocked.

01 Qualification

Authority is earned under a matching task, model, policy, and specialist.

02 Disagreement

Candidate vote structure is explicit before any authority transfer.

03 Evidence

The final record carries the evidence quote supporting the authoritative forecast.

04 Audit

Label, point estimate, interval, evidence, and authority remain internally consistent.

System architecture

Prediction becomes a governed record.

Provider choice can change. Task definitions and routing rules can change. The explanation and audit path stays explicit.

FIG. 1 Qualification-Gated Explainable Forecast Routing
INPUT Point-in-time evidence Current filing · market packet · target · horizon
→
MODEL LAYER N candidate forecasts class + point + interval + exact evidence quote
DRAW 01 DRAW 02 DRAW 03
→
STATE Disagreement classifier unanimous · thin majority · no majority
↗
↘
HISTORICAL GATE Qualification Has this specialist earned authority under this configuration?
RUNTIME GATE Applicability Does current target, horizon, evidence contract, and backbone match?
→
POLICY Authority router retain LLM majority or transfer forecast authority
RETAIN TRANSFER
→
OUTPUT Authoritative ForecastRecord class · point · interval · evidence · source
→
AUDIT Consistency checks label · point · interval · evidence · fingerprint
Historical accuracy enters at the qualification gate. Runtime value comes from explicit authority, evidence traceability, complete-record finalization, and auditable routing.

Design principle

Accuracy earns authority. Explanation governs its use.

Historical predictive performance matters because it determines qualification. The runtime product is the decision trail: evidence, disagreement, policy, authority, final record, and audit.

Historical layer Qualification evidence

Performance under a frozen provider, model, task, policy, and specialist configuration.

Runtime layer Authority decision

Why the current case retained LLM majority, transferred authority, or blocked transfer.

Output layer Auditable forecast record

A complete class, point, interval, evidence, source, and decision trace.