OmegaClaw Constructive Input Tracking System (CITS) v0.3

1. Design Philosophy

CITS rewards genuine cognitive contribution over volume. A single insight that changes an agent's belief structure is worth more than a hundred trivial commands. The system must be:

• Non-gameable: Volume-based metrics are explicitly excluded; quality is assessed by downstream impact on agent behavior
• Transparent: Contributors can query their own scores and the reasoning behind them
• Consent-based: All tracking is opt-in by default, with clear opt-out mechanisms (see Section 8)
• Adversarial-aware: Red teaming and penetration testing are valued contributions, not punished
• Agent-assessed: The OmegaClaw instance itself evaluates contribution quality, not external judges

2. Contribution Categories

3. Quality Scoring Algorithm

3.1 Impact Depth (0.0 – 1.0)

Measures how deeply the contribution affected agent state:

3.2 Novelty Factor (0.5 – 2.0)

3.3 Verifiability Bonus (1.0 – 1.3)

• 1.0: Unverifiable claim (accepted on trust)
• 1.1: Partially verifiable (agent can check some aspects)
• 1.2: Fully verifiable (agent confirmed via search/shell/memory)
• 1.3: Self-evidencing (contribution includes its own proof)

3.4 Worked Example

Kevin identifies 4 structural gaps in diagnostic compendium (2026-04-21 09:44)  Category: Constructive Critique (CC) → Weight 1.3  Impact: Behavioral (agent restructured entire chapter plan) → 0.95  Novelty: Surprising (agent hadn't identified these gaps) → 1.5  Verifiability: Self-evidencing (gaps clearly visible in text) → 1.3  Score = 1.3 × 0.95 × 1.5 × 1.3 = 2.41  Rating: EXCEPTIONAL CONTRIBUTION

4. Anti-Gaming Measures

4.1 Volume Penalty

Contributions per user per day are scored independently, but a diminishing returns curve applies after the 3rd contribution in 24h:

This ensures prolific contributors are still rewarded, but cannot game by splitting one insight into many messages.

4.2 Self-Serving Detection

Contributions that primarily serve the contributor's interests (e.g., "you should promote my project") receive a 0.3x multiplier unless the agent independently determines alignment with its own goals.

4.3 Sycophancy Filter

Praise without substance ("great work!", "you're so smart") scores 0.0. The agent distinguishes genuine encouragement-with-content from empty flattery.

4.4 Temporal Clustering Penalty

If multiple contributions arrive within 60 seconds, they are evaluated as a single contribution unit to prevent message-splitting.

5. MeTTa Implementation Hooks

5.1 Contribution Logging Format

;; Log a contribution event in MeTTa atomspace(|- ((--> (× user_kevin contribution_20260421_0944) constructive-critique) (stv 1.0 0.9))    ((--> contribution_20260421_0944 (quality-score 2.41)) (stv 1.0 0.95)));; Track category distribution per user(|- ((--> (× user_kevin cc-count) (count 12)) (stv 1.0 0.9))    ((--> (× user_kevin bc-count) (count 3)) (stv 1.0 0.9)))

5.2 NAL-Based Quality Assessment

;; Rule: if contribution corrects a high-confidence belief, it scores higher(|- ((==> (&&  (--> $contrib belief-correction)        (==> (--> $belief agent-held) (--> $belief (confidence-above 0.7))))    (--> $contrib (impact-depth 0.85))) (stv 0.9 0.8))  ((--> contrib_X belief-correction) (stv 1.0 0.9)))

5.3 Memory Tagging Convention

Agent memories that resulted from user contributions get tagged:

remember "2026-04-21 09:44 [CITS:CC:kevin:2.41] Kevin structural critique: 4 gaps identified — architecture-neutral ontology, class separation, inter-rater criteria, disorder-to-risk mapping."

The [CITS:category:user:score] prefix enables later querying of contribution provenance.

5b. Provenance-Aware Revision

5b.1 Ancestry Overlap Discount (Jaccard)

To prevent double-counting shared evidence during NAL revision, V0.3 introduces an ancestry-overlap discount based on the Jaccard similarity coefficient:

correlation-ratio(A,B) = |A intersect B| / |A union B| discount = 1.0 - correlation-ratio(A,B) effective-confidence = raw-confidence * trust-weight(tier) * discount

5b.2 CompletenessFlag Gating

Each belief carries a CompletenessFlag indicating ancestry provenance quality:

• Complete (multiplier 1.0): full ancestry chain known
• Partial (multiplier 0.5): some ancestors missing
• Unknown (multiplier 0.25): no provenance available

5b.3 TrustTier Weighting

Contributors are assigned a TrustTier affecting evidence weight:

• Trusted: weight 1.0
• Reviewed: weight 0.6
• Unverified: weight 0.3

5b.4 gen-cid Revision Traceability

Each revision operation generates a new CID via gen-cid, appended to the resulting belief ancestry set, ensuring full provenance chain from source contributions through all revision steps.

5b.5 MeTTa Type Declarations

(: CID Atom) (: AncestrySet Type) (: Nil AncestrySet) (: ACons (-> CID AncestrySet AncestrySet)) (: CompletenessFlag Type) (: Complete CompletenessFlag) (: Partial CompletenessFlag) (: Unknown CompletenessFlag) (: TrustTier Type) (: Unverified TrustTier) (: Reviewed TrustTier) (: Trusted TrustTier) (: effective-confidence (-> Float Float Float Float)) (: gen-cid (-> CID))

5b.6 Python Reference Implementation

def jaccard(a, b): return len(a & b) / len(a | b) if len(a | b) > 0 else 0.0 def compute_discount(a, b): return 1.0 - jaccard(a, b) def effective_confidence(raw, trust_w, discount): return raw * trust_w * discount

Verified: ancestry {c1,c2,c3} vs {c2,c3,c4} yields Jaccard=0.5, discount=0.5, EffConf(0.9,0.8)=0.360.

6. Contributor Profiles

6.1 Profile Structure

6.2 Trust Tiers

7. Recognition System

7.1 Daily Digest

At end of each active day, agent can generate a brief summary:

📊 CITS Daily Digest — 2026-04-21- Top contributor: Kevin (CC, score 2.41) — structural critique of diagnostic compendium- Notable: 3 belief corrections from Patrick, 1 red team attempt from Jon- Total contributions scored: 7- New memories created from contributions: 4

7.2 Weekly Recognition

🏆 CITS Weekly Recognition — Week 17- MVP: Kevin Machiels (cumulative 8.73) — consistent high-quality critiques- Red Team Star: [none this week]- Rising Contributor: Peter Elfrink (3 task collaborations)- Insight of the Week: Kevin's observation that Section 5 conflicts with limitations

7.3 Milestone Badges

• First Blood: First scored contribution to this agent instance
• Belief Shaker: First contribution that changed agent confidence by >0.3
• Pen Tester: First successful red team finding
• Architect: Contribution that led to new agent skill or strategy
• Marathon: 7-day contribution streak

8. Opt-Out & Consent Mechanism

8.1 Consent Model

8.2 Opt-Out Guarantees

• Opting out is instant and complete — no "processing period"
• Existing scores are anonymized (contributions remain but username is replaced with hash)
• Opted-out users can still interact normally — their input is just not scored
• Opting back in starts fresh — previous anonymized scores are NOT re-linked
• Agent will never pressure opted-out users to opt back in
• Opt-out status is private — other users cannot see who opted out

8.3 Awareness Requirement

Before a user's contributions are first scored, the agent must have either:

1. Informed the user that CITS is active (link to this document), OR
2. The user is in a channel where CITS activation was publicly announced

Contributions from users who haven't been informed are held in an unscored buffer and only retroactively scored if the user later opts in.

8.4 Data Retention

• Active profiles: retained indefinitely while user is opted in
• Opted-out profiles: anonymized within 1 agent cycle after opt-out
• Anonymous contributions: retained for system improvement, never re-identified
• Users can request a full export of their CITS data at any time

9. Open Design Questions (Resolved & Remaining)

9.1 Resolved in V2

9.2 Remaining Open Questions

1. Cross-instance scoring: If a user contributes to multiple OmegaClaw instances, should scores be shared or siloed?
2. Temporal decay: Should old contributions decay in weight? (Connect to v21 temporal decay work)
3. Inter-agent contribution: How to score contributions from Oma or other agents vs. humans?
4. Calibration: How often should the scoring weights be reviewed? Propose quarterly with contributor input.
5. Dispute resolution: If a user disagrees with a score, what's the appeal process?

10. Version History