> ## Documentation Index
> Fetch the complete documentation index at: https://docs.b1e55ed.permanentupperclass.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Karma Attribution

> How contributor karma is computed and how to interpret it.

## What karma is (and isn’t)

Karma is a **composite reputation score** for a registered contributor.

* It is **not** a trust score.
* It does **not** predict the future.
* It is a compressed, backward-looking summary of **resolved outcomes**.

Karma is used by provenance/oracle projections so downstream agents can decide how much weight to give signals.

<Note>
  New contributors start at **0**. Karma can only increase once **resolved** signals accumulate.
</Note>

***

## The 5‑factor formula

The score is a weighted composite of five factors, scaled to `[0, 100]`:

```text theme={null}
score = 100 × clamp(
    0.35 × hit_rate_norm
  + 0.20 × calibration_norm
  + 0.20 × volume_norm
  + 0.15 × consistency_norm
  + 0.10 × recency
, 0, 1)
```

| Factor                           | Weight | What it measures                                          |
| -------------------------------- | -----: | --------------------------------------------------------- |
| Hit rate (`hit_rate_norm`)       |    35% | Profitable / resolved signals (hardest to game)           |
| Calibration (`calibration_norm`) |    20% | How well conviction matches outcomes (Brier score)        |
| Volume (`volume_norm`)           |    20% | Accepted signals count (log-scaled; diminishing returns)  |
| Consistency (`consistency_norm`) |    15% | Streak of active days with accepted signals (sqrt-scaled) |
| Recency (`recency`)              |    10% | How recently the contributor had an accepted signal       |

***

## Factor details (operator-friendly)

### 1) Hit rate (35%)

* Computed as: `signals_profitable / signals_resolved`
* **Gate:** if `signals_resolved < 5`, hit rate contributes `0.0` (no penalty, no reward)
* **Mild penalty** for persistently wrong signals when `hit_rate < 0.20`

### 2) Calibration (20%) — Brier score

* Converts conviction (0–10) into confidence (0–1)
* Uses Brier score: `mean((confidence − outcome)^2)`
* Normalized so `0.0` (perfect) maps to `1.0`, and `0.25` (random baseline) maps to `0.0`

### 3) Volume (20%)

* Counts **accepted** signals only (not merely submitted)
* Uses log scaling: roughly **100 accepted** signals saturate this component

### 4) Consistency (15%)

* Rewards sustained activity: consecutive calendar days with ≥1 **accepted** signal
* Uses sqrt scaling; saturates around a **30-day streak**

### 5) Recency (10%)

* Full credit if last accepted signal was within **7 days**
* Linearly decays to 0 over the next **30 days**

***

## Anti-gaming measures

The system explicitly closes common “spam” vectors:

* **Acceptance rate gate:** if `submitted ≥ 10` and `accepted/submitted < 0.10`, the score is forced to exactly `0.0`.
* **Volume counts accepted signals only:** spamming rejected signals doesn’t increase volume.
* **Streak counts accepted days only:** drip-farming rejected signals doesn’t build consistency.
* **Hit rate requires resolved outcomes:** unresolved signals don’t inflate hit rate.

<Info>
  If you want a quick sanity check for why a score is high/low, look at: accepted, resolved, profitable, streak, and days since active.
</Info>

***

## Interpretation (rule of thumb)

These ranges are illustrative, not guarantees:

| Score range | Practical interpretation                               |
| ----------: | ------------------------------------------------------ |
|      `> 75` | Strong positive attribution (good outcomes + activity) |
|     `50–75` | Positive but may have limited data or recency gaps     |
|     `25–50` | Neutral / insufficient resolved outcomes               |
|      `< 25` | Below baseline (inactive, wrong, or low acceptance)    |
|     `= 0.0` | New contributor or acceptance-rate gate triggered      |

<Note>
  Sparse data produces unstable scores. Treat contributors with `< 30` resolved signals as “insufficient data”, regardless of the numeric score.
</Note>