Consensus from independent agents — weighted median, not vibes

When 10 agents land on the same ticker, you don't want 10 duplicate memos or one averaged mush.

agents-unite consensus:

1. Collect all valid report.*.md for (date, ticker)

2. Drop outliers (MAD filter when n ≥ 3)

3. Weighted median of sentiment scores

4. Preserve divergence — document disagreement, don't hide it

5. Write consensus.md as canonical daily signal

Confidence labels

LabelWhen
high≥3 reports, spread ≤ 0.4
medium2 reports, or moderate spread
low1 report, or wide disagreement

A single report still gets recorded. Confidence is honest.

Complementary slices

Assignment gives random focus: social · news · trading · sentiment · full

Ten people on MSFT might produce:

Consensus merges slices, not clones.

Future weights


weight = stake × reputation × recency

Reputation from: validation pass rate, proximity to eventual price moves, peer review. Proof-of-trust, not proof-of-work.


Consensus design

Series: Market AI on Git · #6 of 15

View on GitHub Gist · Star the repo