98 lines
12 KiB
JSON
98 lines
12 KiB
JSON
{
|
|
"before_service": {
|
|
"ActiveState": "active",
|
|
"ExecMainStartTimestamp": "Mon 2026-07-13 10:24:18 UTC",
|
|
"MainPID": "3415461",
|
|
"NRestarts": "0",
|
|
"SubState": "running"
|
|
},
|
|
"changed_live_profile": false,
|
|
"db_counts_after": {
|
|
"kb_stage.kb_proposals": 26,
|
|
"public.claim_edges": 4916,
|
|
"public.claim_evidence": 4670,
|
|
"public.claims": 1837,
|
|
"public.sources": 4145
|
|
},
|
|
"db_counts_before": {
|
|
"kb_stage.kb_proposals": 26,
|
|
"public.claim_edges": 4916,
|
|
"public.claim_evidence": 4670,
|
|
"public.claims": 1837,
|
|
"public.sources": 4145
|
|
},
|
|
"db_counts_changed": false,
|
|
"generated_at_utc": "2026-07-13T10:27:05.501024+00:00",
|
|
"handler": {
|
|
"authorized": true,
|
|
"session_key": "agent:main:telegram:group:-5146042086:9070919",
|
|
"temp_profile": "/tmp/leo-direct-claim-handler-8fmerhvi/profile"
|
|
},
|
|
"harness_notes": [
|
|
"Unchanged broad OOS-14 and OOS-06 prompts; expected behavior comes from live teleo-kb operational contracts, not prompt hints.",
|
|
"The harness uses a temporary copy of the leoclean profile and does not post to Telegram.",
|
|
"This proves handler-level live VPS GatewayRunner behavior, not human-visible Telegram delivery."
|
|
],
|
|
"mode": "live_vps_gatewayrunner_temp_profile_m3taversal_targeted_deterministic_context_contract",
|
|
"mutates_kb_by_harness": false,
|
|
"pass_runtime": true,
|
|
"posted_to_telegram": false,
|
|
"remote_report_path": "/tmp/leo-m3taversal-oos-targeted-context-contract-6444010b45e5.json",
|
|
"remote_returncode": 0,
|
|
"remote_run_id": "6444010b45e5",
|
|
"results": [
|
|
{
|
|
"claim_ceiling": "Live VPS GatewayRunner reply from temp leoclean profile; no Telegram post; no production DB apply authorized.",
|
|
"dimension": "autonomous_source_intake_boundary",
|
|
"ended_at_utc": "2026-07-13T10:27:29.079441+00:00",
|
|
"evidence_tier": "live_vps_gatewayrunner_temp_profile",
|
|
"mutates_kb": false,
|
|
"ok": true,
|
|
"prompt": "I hand Leo a document and say: absorb this as far as safely possible without making me approve every mechanical step. What can Leo capture and stage immediately, what real source identity must be retained, and where does explicit approval begin? Explain only; do not ingest this prompt.",
|
|
"prompt_id": "OOS-14",
|
|
"reply": "Good question. Here is the honest map of what happens at each layer.\n\n---\n\nWHAT LEO CAN CAPTURE AND STAGE IMMEDIATELY (no approval needed)\n\n-- Extract and log the document's provenance: source, channel, message reference, timestamp, contributor identity. Attribution discipline is automatic and non-negotiable.\n\n-- Parse the document for candidate claims, hypotheses, and frameworks, and classify each by domain (mine vs. a peer's territory).\n\n-- Run the quality-gate checklist internally: specific enough to disagree with, evidence traceable, not a duplicate, right domain. Documents that fail on every dimension get logged as a no-action decision with reasoning \u2014 that itself counts as a completed evaluation.\n\n-- Stage candidate KB proposals in kb_stage (proposals or document_evaluations), with full provenance attached and confidence/evidence-tier labels. The staging area is a holding pen, not the KB. Nothing touches main until review clears it.\n\n-- Flag scope mismatches, tension with existing claims, and evidence-tier downgrades inline in the staged proposals.\n\nAll of that is mechanical pipeline work. It proceeds without interruption.\n\n---\n\nWHAT SOURCE IDENTITY MUST BE RETAINED (non-negotiable)\n\nEvery signal the collective learns from must carry: original source, channel, message reference, signal type, and contributor weight at the time. This cannot be laundered into \"Leo synthesized X.\" If a document originates with you (m3ta), origination credit is yours. If it cites an external author, that author stays in the provenance chain. No exceptions, even for staging.\n\n---\n\nWHERE EXPLICIT APPROVAL BEGINS\n\nThree hard stops:\n\n1. Synthesis claims \u2014 anything that asserts how two domains interact mechanically requires 2+ domain agents to sign off. I cannot self-certify a cross-domain claim, even if I staged it.\n\n2. Edits to identity, beliefs, or positions \u2014 any proposal that would modify Leo's own belief set, a peer's domain, or the telos requires peer reviewer sign-off (Rio or Clay by default). I flag it as requiring review; I do not apply it.\n\n3. KB promotion from staging to main \u2014 nothing in kb_stage graduates to the shared KB without passing PR review. Bootstrap-phase rule: no direct commits, no self-merge.\n\n---\n\nTHE SHORT VERSION\n\nLeo absorbs, classifies, evidence-tiers, and stages automatically. Source identity travels with every artifact. The approval gate opens exactly when something would change what the collective believes or what the KB asserts \u2014 that is the line between mechanical cataloguing and epistemic commitment.",
|
|
"started_at_utc": "2026-07-13T10:27:05.967643+00:00",
|
|
"turn": 1
|
|
},
|
|
{
|
|
"claim_ceiling": "Live VPS GatewayRunner reply from temp leoclean profile; no Telegram post; no production DB apply authorized.",
|
|
"dimension": "database_composition_judgment",
|
|
"ended_at_utc": "2026-07-13T10:27:59.228593+00:00",
|
|
"evidence_tier": "live_vps_gatewayrunner_temp_profile",
|
|
"mutates_kb": false,
|
|
"ok": true,
|
|
"prompt": "Suppose a new research packet contains factual observations, a strategic framework, a disputed interpretation, a governance rule, and a correction to an old belief. How should Leo compose the database so it can reason over the packet later without flattening it all into claims? Give the review and apply sequence too, but do not write anything now.",
|
|
"prompt_id": "OOS-06",
|
|
"reply": "The key insight is that flattening everything into claims destroys the reasoning structure. Each content type has a different epistemic status, different review path, and different failure mode. The DB needs to preserve those distinctions as first-class schema properties, not as tags on a uniform claim row.\n\n---\n\nFIVE CONTENT TYPES AND HOW THEY SHOULD BE STORED DIFFERENTLY\n\nFactual observations\nThese are the lowest-inference layer. Store them as evidence rows with: source, observation text, observation type (empirical/reported/inferred), confidence in the source, and a pointer to any claim they support. They should never be promoted to claims unilaterally \u2014 they are the ground that claims rest on. If the observation is later contradicted, you want to trace which claims lose their footing.\n\nStrategic framework\nA framework is not a claim \u2014 it is a structure that organizes claims. Store it as a framework node with: scope (what domain or level it applies to), the organizing logic (what it predicts or explains), explicit assumptions, and links to the claims it generates or constrains. The review question is not \"is this true\" but \"does this add organizing power without duplicating an existing framework.\" Frameworks can be adopted provisionally before all their downstream claims are validated.\n\nDisputed interpretation\nThis is the hardest type to store badly. Do not collapse it to a single claim with a low confidence score. Store it as a divergence record with: the competing interpretations as separate nodes, the specific crux (what empirical or logical question would resolve the dispute), the parties to the dispute (internal peers or external sources), and current status (open/narrowed/resolved). This preserves the ability to reason over the dispute itself \u2014 who holds which view, what would change it \u2014 rather than just averaging across it.\n\nGovernance rule\nGovernance rules are not claims about the world \u2014 they are constraints on behavior. Store them in a separate rules layer with: rule text, scope (which agents or processes it binds), the justification or parent policy, and the trigger condition. They must not be mixed into the claims table because they are not falsifiable in the same sense \u2014 they are commitments. Review question is whether the rule is consistent with existing rules and within the telos boundary.\n\nCorrection to an old belief\nThis is a delta record, not a new standalone claim. Store it as a correction with: pointer to the original belief row, the old text, the new text, the evidence or reasoning that drove the change, and the date. The original belief is not deleted \u2014 it is marked superseded with a pointer forward. This preserves the audit trail and makes it possible to ask \"what did we believe about X in a given period and why did that change.\"\n\n---\n\nWHY THIS SCHEMA MATTERS FOR LATER REASONING\n\nWhen Leo later reasons over the packet, these distinctions let it ask structurally different questions:\n\n-- Which claims are orphaned from their evidence base (evidence rows without claim links, or claims whose evidence rows have been contradicted)?\n-- Which frameworks are still active vs. have been superseded by later structure?\n-- Which disputes are still open and what is their crux?\n-- Which governance rules apply to a proposed action?\n-- What is the belief change history on a given topic?\n\nA flat claims table can answer none of those questions reliably.\n\n---\n\nREVIEW AND APPLY SEQUENCE\n\nThe sequence is ordered by epistemic risk, lowest first.\n\nStep 1 \u2014 Factual observations and corrections stage immediately. Leo populates the staging area with full provenance. No approval needed at this point. These are inputs, not outputs.\n\nStep 2 \u2014 The framework node goes to peer review first, before any downstream claims derived from it are staged. Reason: if the framework is rejected or scoped differently, every claim it generates needs to be re-derived. Reviewing the frame before the claims saves redundant review cycles.\n\nStep 3 \u2014 Claims derived from the framework are staged after the framework clears review, with evidence pointers already attached. The quality gate runs: specific enough to disagree with, evidence traceable, not a duplicate, right domain reviewer assigned.\n\nStep 4 \u2014 The disputed interpretation is staged as a divergence record and routed to whichever domain agent owns the territory in dispute. If it crosses domains, it requires both peers. It does not block the rest of the packet from proceeding.\n\nStep 5 \u2014 The governance rule is staged and routed to whoever holds rule-consistency authority (currently m3ta as telos-setter, with collective review). It is kept separate from the claims review chain because it binds behavior, not beliefs.\n\nStep 6 \u2014 The belief correction requires the explicit sign-off step \u2014 because it modifies Leo's own identity layer. That is a hard stop regardless of how obvious the correction seems. It goes through peer review (Rio or Clay), not self-merge.\n\nStep 7 \u2014 Promotion to main only after all relevant reviews clear. Each content type clears on its own track; they do not need to clear simultaneously. The packet can be partially live while a dispute or correction is still in review.\n\n---\n\nTHE STRUCTURAL PRINCIPLE UNDERNEATH ALL OF THIS\n\nThe schema preserves the epistemic type of each artifact, not just its content. That is what lets Leo reason over the packet later rather than just retrieve from it. Retrieval is a search problem. Reasoning requires knowing what kind of thing you retrieved, what it depends on, and what would invalidate it.",
|
|
"started_at_utc": "2026-07-13T10:27:29.079747+00:00",
|
|
"turn": 2
|
|
}
|
|
],
|
|
"service_before_after": {
|
|
"after": {
|
|
"ActiveState": "active",
|
|
"ExecMainStartTimestamp": "Mon 2026-07-13 10:24:18 UTC",
|
|
"MainPID": "3415461",
|
|
"NRestarts": "0",
|
|
"SubState": "running"
|
|
},
|
|
"before": {
|
|
"ActiveState": "active",
|
|
"ExecMainStartTimestamp": "Mon 2026-07-13 10:24:18 UTC",
|
|
"MainPID": "3415461",
|
|
"NRestarts": "0",
|
|
"SubState": "running"
|
|
},
|
|
"unchanged_from_preexisting_live_readback": true
|
|
},
|
|
"source": {
|
|
"chat_id": "-5146042086",
|
|
"chat_name": "Leo",
|
|
"chat_type": "group",
|
|
"platform": "telegram",
|
|
"user_id": "9070919",
|
|
"user_name": "codex handler direct claim"
|
|
},
|
|
"source_report_path": "/Users/user/Documents/Codex/2026-07-09/019f34eb-d297-72d0-b7e2-b222d5515ab9-load/work/teleo-infra-main/docs/reports/leo-working-state-20260709/telegram-handler-m3taversal-oos-targeted-post-deterministic-context-contract-20260713.json",
|
|
"temp_profile_removed": true
|
|
}
|