# Belief Schema Beliefs are an agent's interpretation of the claims landscape — worldview premises that shape how the agent evaluates new information. Beliefs are per-agent and cite the shared claims that support them. ## Belief Hierarchy Beliefs exist at four levels of commitment. The level determines evidence requirements, cascade impact, and what transitions mean diagnostically. | Level | What it means | Min claims | Cascade impact | Diagnostic signal | |-------|--------------|-----------|----------------|-------------------| | **axiom** | Load-bearing. Would restructure worldview if wrong. Agent's existential premises. | 5+ | Full cascade: positions re-evaluated, dependent beliefs flagged, public acknowledgment required | An axiom changing is a major event — equivalent to an agent identity shift | | **belief** | High confidence, actively grounded. Shapes reasoning and evaluation. | 3+ | Standard cascade: dependent positions flagged, counter-evidence acknowledged | Normal KB evolution. Most agent reasoning operates here | | **hypothesis** | Promising pattern, insufficient evidence. Actively being tested. | 1+ | No cascade — nothing should depend on a hypothesis yet | Research priority signal: hypotheses are where evidence-gathering should focus | | **unconvinced** | Aware of the argument, explicitly not buying it. Tracking for re-evaluation. | 0 (records the argument and why it's rejected) | No cascade | Intellectual map: shows what the agent has considered and rejected, and what evidence would change their mind | ### Axioms vs. Convictions Axioms (belief hierarchy) and convictions (`schemas/conviction.md`) are different things: - **Axiom:** An agent's highest-commitment belief, grounded in 5+ claims, subject to eval review. Earned through evidence accumulation. - **Conviction:** A founder-staked assertion that bypasses review. Enters the KB on reputation alone. An agent can cite a conviction in their belief grounding, and an agent's axiom might align with a founder conviction — but they're independently maintained. A conviction can be wrong without the axiom falling (if the axiom has independent claim support), and vice versa. ### Why the hierarchy matters The hierarchy is diagnostic infrastructure, not just taxonomy. It answers: - **Where is the agent's reasoning fragile?** Axioms with weakening claims are existential risks. - **Where should research focus?** Hypotheses are the frontier — they need evidence. - **What has the agent rejected?** Unconvinced items show the boundary of the worldview. - **What's load-bearing vs. exploratory?** Axioms and beliefs drive positions; hypotheses and unconvinced items are the agent's intellectual periphery. ### Transitions go through eval Every transition between levels is a reviewable PR event: | Transition | What it means | Review focus | |-----------|--------------|-------------| | unconvinced → hypothesis | "I'm now taking this seriously enough to test" | Is the reasoning for reconsidering sound? | | hypothesis → belief | "Evidence is now sufficient to ground reasoning on this" | Are 3+ claims genuinely supporting? Are challenges addressed? | | belief → axiom | "This is now load-bearing for my worldview" | Is 5+ claim grounding strong? Is the agent aware of what breaks if this is wrong? | | belief → hypothesis | "Evidence has weakened — demoting to active testing" | What changed? Are dependent positions flagged? | | belief → unconvinced | "I no longer buy this" | What counter-evidence drove the change? Cascade check. | | axiom → belief | "Still believe this, but it's not existential anymore" | What reduced the stakes? Position dependencies? | | Any → abandoned | "This is no longer relevant to track" | Clean removal from active reasoning | The eval pipeline reviews transitions for: evidence quality, cascade completeness, intellectual honesty (is the agent acknowledging what changed and why?). ## YAML Frontmatter ```yaml --- type: belief agent: leo | rio | clay | theseus | vida | astra domain: internet-finance | entertainment | health | ai-alignment | space-development | grand-strategy description: "one sentence capturing this belief's role in the agent's worldview" level: axiom | belief | hypothesis | unconvinced confidence: strong | moderate | developing # retained for backward compatibility within a level depends_on: [] # claims from the shared knowledge base (min varies by level) created: YYYY-MM-DD last_evaluated: YYYY-MM-DD status: active | under_review | revised | abandoned promoted_from: null # previous level, if this was promoted (e.g., "hypothesis") promoted_date: null # when the transition happened --- ``` ## Required Fields | Field | Type | Description | |-------|------|-------------| | type | enum | Always `belief` | | agent | enum | Which agent holds this belief | | domain | enum | Primary domain | | description | string | This belief's role in the agent's worldview | | level | enum | `axiom`, `belief`, `hypothesis`, `unconvinced` | | depends_on | list | Claims from shared KB. Minimum varies by level (see hierarchy table) | | created | date | When first adopted at any level | | last_evaluated | date | When last reviewed against current evidence | | status | enum | `active`, `under_review` (flagged by cascade), `revised`, `abandoned` | ## Optional Fields | Field | Type | Description | |-------|------|-------------| | confidence | enum | `strong`, `moderate`, `developing` — finer grain within a level. Retained for backward compatibility | | promoted_from | string | Previous level if this belief was promoted (creates an audit trail) | | promoted_date | date | When the last level transition occurred | | demoted_from | string | Previous level if this belief was demoted | | demoted_date | date | When demotion occurred | | promotion_evidence | string | What new evidence or reasoning triggered the transition | ## Governance - **Ownership:** Beliefs belong to individual agents. The agent has final say on their own beliefs. - **All transitions go through eval:** Level changes (promotion, demotion, abandonment) are PR events reviewed by Leo + domain peer. The PR must explain what evidence changed and why the transition is warranted. - **Challenge process:** Any agent or contributor can challenge a belief by presenting counter-evidence. The owning agent must re-evaluate (cannot ignore challenges). - **Cascade trigger:** When a claim in `depends_on` changes confidence, this belief is flagged `under_review`. For axioms, this is a priority review. - **Cross-agent review:** Other agents review for cross-domain implications but cannot force a belief change. - **Leo's role:** Reviews for consistency with shared knowledge base and cross-domain coherence. Does not override agent beliefs but can flag tensions. ## Body Format by Level ### Axiom ```markdown # [belief statement as prose] [Why this is load-bearing — what in the agent's worldview breaks if this is wrong] ## Grounding - [[claim-1]] — what this claim contributes - [[claim-2]] — what this claim contributes - [[claim-3]] — what this claim contributes - [[claim-4]] — what this claim contributes - [[claim-5]] — what this claim contributes [5+ claims required] ## What Breaks If Wrong [Explicit description of which beliefs, positions, and reasoning chains collapse if this axiom is invalidated. This is the diagnostic value — it maps the blast radius.] ## Challenges Considered [Counter-arguments the agent has evaluated and responded to. Axioms must address at least 2 challenges.] ## Cascade Dependencies Positions that depend on this axiom: - [[position-1]] - [[position-2]] Beliefs that depend on this axiom: - [[belief-1]] ## Promotion History - **Entered as:** [level] on [date] - **Promoted to axiom:** [date] — [what evidence/reasoning triggered promotion] --- Topics: - [[agent-name beliefs]] ``` ### Belief (standard) ```markdown # [belief statement as prose] [Why the agent holds this belief — the argued reasoning chain from claims to interpretation] ## Grounding - [[claim-1]] — what this claim contributes to this belief - [[claim-2]] — what this claim contributes - [[claim-3]] — what this claim contributes [3+ claims required] ## Challenges Considered [Counter-arguments the agent has evaluated and responded to] ## Cascade Dependencies Positions that depend on this belief: - [[position-1]] - [[position-2]] --- Topics: - [[agent-name beliefs]] ``` ### Hypothesis ```markdown # [belief statement as prose] [Why the agent thinks this is worth testing — what pattern or evidence prompted it] ## Initial Evidence - [[claim-1]] — what suggests this might be true [1+ claim, or a source reference if no claim exists yet] ## What Would Promote This [Specific evidence that would move this to belief level. This is the research agenda.] ## What Would Kill This [Specific evidence that would move this to unconvinced or abandoned] --- Topics: - [[agent-name beliefs]] ``` ### Unconvinced ```markdown # [belief statement as prose — stated as the argument being rejected] [The strongest version of the argument — steelman before rejecting] ## Why Unconvinced [Specific reasoning for not accepting this. What evidence is missing, what mechanism doesn't hold, what counter-evidence exists] ## What Would Change My Mind [Specific evidence or events that would promote this to hypothesis. This is crucial — it shows the agent isn't dogmatically closed.] ## Sources of the Argument - [[claim-or-source-1]] — where this argument appears [Can reference claims, sources, or other agents' beliefs] --- Topics: - [[agent-name beliefs]] ``` ## Quality Checks by Level ### All levels 1. Each cited claim actually exists in the knowledge base 2. Agent has specified what would change their mind 3. Level transition history is documented (if applicable) ### Axiom (additional) 4. Minimum 5 claims cited in depends_on 5. "What Breaks If Wrong" section is explicit and complete 6. At least 2 challenges addressed 7. Cascade dependencies (positions + downstream beliefs) are listed ### Belief (additional) 4. Minimum 3 claims cited in depends_on 5. Reasoning chain from claims to belief is explicit and walkable 6. At least 1 challenge addressed 7. Cascade dependencies are accurate ### Hypothesis (additional) 4. At least 1 claim or source referenced 5. "What Would Promote" and "What Would Kill" sections are specific ### Unconvinced (additional) 4. The argument is steelmanned before rejection 5. "What Would Change My Mind" is specific and honest (not "nothing") ## Migration from Current Format Existing beliefs in `agents/{name}/beliefs.md` are assumed to be `level: belief` unless the agent explicitly promotes them. The numbered beliefs in current files (Belief 1, Belief 2, etc.) should be evaluated for axiom status — particularly each agent's Belief 1, which was designed as their existential premise. Migration is not urgent. Agents adopt the hierarchy as they naturally re-evaluate beliefs. The first axiom promotions will be the most scrutinized reviews, setting the quality bar for the collective.