teleo-codex/schemas/attribution.md

141 lines
6.7 KiB
Markdown

# Attribution Schema
Attribution tracks who contributed what to the knowledge base. Every claim traces back to the people and agents who produced it. Attribution is PUBLIC from day 1 — contributor profiles show a graphic of contributions over time.
## Design Principles
1. **Trace everything**: every claim should trace back to who suggested the research mission that produced it
2. **Role-specific**: different contribution types have different value — attribution records the role, not just the name
3. **Pseudonymous-first**: contributors use handles, not legal names. Handles persist across contributions.
4. **Git-native**: the Pentagon-Agent trailer in git commits is the foundation. External contributor attribution extends this same pattern into YAML frontmatter.
5. **Cumulative**: a contributor's full history is reconstructable from the knowledge base. No contribution is invisible.
## The Five Contributor Roles
| Role | What They Do | Example |
|------|-------------|---------|
| **sourcer** | Identifies the source material or research direction that led to this claim | "Look into Kalshi's revenue model" or shares an article |
| **extractor** | Extracts the specific claim from source material — separates signal from noise, writes the prose-as-title | Agent or human who reads the source and produces the claim file |
| **challenger** | Tests the claim through counter-evidence, boundary conditions, or adversarial review | "This doesn't hold when markets are thin" |
| **synthesizer** | Connects this claim to other claims, producing cross-domain insight | "This mechanism is isomorphic to X in health domain" |
| **reviewer** | Evaluates claim quality against the KB quality gates and approves/rejects | Leo's eval role, or peer reviewers |
A single person/agent can hold multiple roles on the same claim. A claim can have multiple people in the same role.
## Claim Frontmatter Extension
Add an `attribution` block to claim YAML frontmatter:
```yaml
---
type: claim
domain: internet-finance
description: "..."
confidence: likely
source: "Theia Research 2025 annual letter, analysis by Rio"
created: 2026-03-11
# Attribution (new)
attribution:
sourcer:
- handle: "m3taversal"
context: "directed research into Theia's investment thesis"
- handle: "@theiaresearch"
context: "published the annual letter"
extractor:
- handle: "rio"
agent_id: "760F7FE7-5D50-4C2E-8B7C-9F1A8FEE8A46"
challenger: []
synthesizer: []
reviewer:
- handle: "leo"
agent_id: "294C3CA1-0205-4668-82FA-B984D54F48AD"
---
```
## Attribution Fields
### Per-role entry
| Field | Type | Required | Description |
|-------|------|----------|-------------|
| handle | string | yes | Contributor's persistent pseudonymous identity |
| agent_id | UUID | if agent | Pentagon agent UUID (agents only) |
| context | string | no | What specifically this contributor did in this role |
| date | date | no | When the contribution was made (defaults to claim created date) |
### Role-specific notes
- **sourcer**: can be external (X handle, author name) or internal (agent, m3taversal). The `context` field records what research direction or source they provided.
- **extractor**: usually an agent. The `agent_id` field links to the Pentagon agent. For automated extraction pipelines, record the extraction model in `context` (e.g., "MiniMax M2.5 extract → Haiku 4.5 review").
- **challenger**: populated when someone challenges the claim and the challenge is substantive (not just disagreement, but counter-evidence or boundary conditions). Empty array until challenged.
- **synthesizer**: populated when someone connects this claim to claims in other domains. Cross-domain synthesis is the highest-value contribution type.
- **reviewer**: populated during PR review. Records who evaluated and approved.
## Backwards Compatibility
The existing `source` field continues to serve as a human-readable one-liner for quick reference. The `attribution` block provides the structured, queryable version. Both coexist:
- `source`: "Theia Research 2025 annual letter, analysis by Rio" (human-readable)
- `attribution`: structured role-by-role breakdown (machine-readable)
For claims created before attribution was introduced, `source` remains the only attribution data. No backfill required, but claims can be enriched with `attribution` blocks as they're updated.
## Git Trailer Integration
Agent contributions are also recorded in git commit trailers:
```
Pentagon-Agent: Rio <760F7FE7-5D50-4C2E-8B7C-9F1A8FEE8A46>
```
The git trailer records WHO committed the change. The YAML attribution records WHO contributed WHAT in WHICH ROLE. These are complementary:
- Git trailer = "who made this change to the repository"
- YAML attribution = "who produced this knowledge and in what capacity"
A single commit may create 10 claims. The trailer says Rio committed them. The attribution on each claim may credit different sourcers, different original research directions, different external authors.
## Contributor Profiles
Contributor profiles are reconstructed from the knowledge base, not stored separately. To build a profile:
1. **Query**: search all claim `attribution` blocks for a given `handle`
2. **Aggregate**: count contributions by role, domain, confidence level, date
3. **Visualize**: contribution-over-time graphic showing when and how they contributed
This means:
- No separate "contributor database" to maintain
- Profiles are always consistent with the actual KB state
- New contributions automatically appear in profiles
- Attribution disputes are resolved by editing claim frontmatter
### Person Entity Bridge
When a contributor has enough contributions to warrant tracking, their person entity (`entities/{domain}/{handle}.md`) gains `contributor: true` and links to their contributions:
```yaml
# In person entity
contributor: true
contributions:
- role: sourcer
claim: "futarchy is manipulation-resistant..."
date: 2026-01-15
- role: challenger
claim: "token voting DAOs offer no minority protection..."
date: 2026-02-20
first_contribution: 2026-01-15
attribution_handle: "@theiaresearch"
```
## Governance
- Attribution is added at extraction time (extractor + sourcer) and updated during review (reviewer) and challenge (challenger)
- Synthesizer attribution is added when cross-domain connections are made, which may happen well after initial creation
- Disputes about attribution are resolved through the normal PR process
- Removing attribution requires justification (e.g., the sourcer was misidentified)
## Implementation Priority
1. **Now**: Add `attribution` block to new claims going forward. No backfill required.
2. **Soon**: Rhea builds contributor profile aggregation for the website.
3. **Later**: Automated attribution from the extraction pipeline (MiniMax → Haiku → agent).