← Research archive
Technical note verified Sep 15, 2026

Autonomous Natural Research Gap Lifecycle

KESPA operationalized a natural knowledge-gap research path that detects low-confidence public-knowledge gaps during normal use, queues isolated research jobs, gathers multi-source evidence, verifies claims, imports approved releases, promotes trusted knowledge, indexes it locally, and records provenance through NexLedger without modifying the legacy/manual lifecycle.

autonomous-research knowledge-gap trusted-knowledge groq nexledger
MARKDOWN

README.md

3,682 bytes SHA-256 c0112f748220aa98…

KESPA-NOTE-010 — Autonomous Natural Research Gap Lifecycle

Date: 2026-09-15 Status: Verified — end-to-end production path proven Project: KESPA AI / NexLabs Studios

Purpose

KESPA's manual trusted-knowledge lifecycle could already take supplied training material through evidence, verification, approval, promotion, and indexing.

The next step was harder:

Can KESPA detect when it does not know enough during normal use, research the missing public-general
knowledge, verify it, and add it to trusted knowledge without silently weakening the existing trust boundary?

The natural-research lifecycle was built as an isolated path rather than by modifying the older manual learning pipeline.

Trigger policy

A normal response can become a research-gap candidate when the recorded request satisfies conservative conditions such as:

  • Brain response path;
  • non-cache request;
  • RAG actually searched;
  • learning-style goal;
  • trust after response at or below 0.72;
  • best RAG score at or below 0.45, or no useful result.

The detector excludes:

  • private/internal requests;
  • memory-related requests;
  • high-stakes requests;
  • ordinary routing-only misses.

Eligible jobs are classified as:

public_general_knowledge

Isolated research flow

The implemented path is:

normal question -> gap detector -> signed queue -> research worker -> evidence -> claims -> release -> signed handoff -> verification -> approval -> trusted knowledge -> index -> NexLedger

The research worker does not modify, wrap, or piggyback on the legacy/manual lifecycle scripts.

Evidence policy

The natural-research worker uses a stricter rule than ordinary search-assisted answering:

  • search snippets are never factual evidence;
  • each promoted claim requires support from at least 2 independent source organizations;
  • high-stakes, sensitive, and time-sensitive requests are refused by this worker;
  • claims are verified before release;
  • the imported release is verified again before promotion.

End-to-end proof: Cuckoo filter

The first full natural-gap proof used the question:

What is a Cuckoo filter?

Recorded lifecycle:

  • request: research-gap-20a905c55f6ad819c57f6920
  • release: forge-release-07990ffc7fc9d09f9be2dd3b
  • release SHA-256:

    • 07990ffc7fc9d09f9be2dd3b5594285e5ee0789f7430e24fb1a812cbbf996ed3

  • evidence sources: 3
  • independent source organizations: 3
  • candidate claims: 4
  • verified claims: 4
  • Groq requests: 2
  • Groq tokens: 11,034
  • imported candidate: #82
  • promoted trusted knowledge: #74
  • NexLedger asset: FORGEAI#TK123V1

Final result:

PASS

Why this matters

This turns "learning from use" into a controlled research lifecycle rather than a model-memory shortcut.

The system does not persist a weak answer merely because a user asked a question.

Instead:

low-confidence gap -> explicit research -> evidence -> verification -> trusted promotion

That preserves the core KESPA rule:

Candidate intelligence is not trusted knowledge.

Boundary

This path was intentionally isolated from the older manual lifecycle.

The proof did not require changing the legacy/shared learning scripts, and search snippets were never accepted as evidence.

The authoritative record remains trusted knowledge; the semantic index remains rebuildable.

Limitation

One successful Cuckoo-filter run proves the machinery works end to end.

It does not establish a mature autonomous-research accuracy rate. That requires many more runs with measured false acceptance, false rejection, source quality, research yield, latency, token cost, and energy.

TEXT

SHA256SUMS.txt

393 bytes SHA-256 48abd7fdf45cc120… 
c0112f748220aa98412e2e55e1f15970f103b02c529795dcc188159178ede6a2  README.md
7f38c092f723fad85037fc98a3f5d15731fc2c78cd95387248a39ce7418a9590  note.json
dc3e6b8e1856817f4a5ed392b6eb836f7a5c8fb1ae5b1261aa4cd7a2714c349f  metrics.csv
b56ad2dacd55be2ef9842b4d7d57f56df2247b8440d42a640643116833da3bd3  methodology.md
098b005499956619c3b03e8418a3b880c48e7442ec3e6cd3463fe2fe5c4698bb  provenance.json
MARKDOWN

methodology.md

1,522 bytes SHA-256 b56ad2dacd55be2e…

Methodology

Gap detection

Persisted assistant responses are scanned after normal use.

A conservative policy identifies public-general-knowledge requests where retrieval was attempted but the resulting trust/retrieval signals indicate KESPA may not know enough.

Queue and handoff

Eligible requests are written into an isolated signed queue. A separate home-side poller retrieves jobs and acknowledges delivery.

The research result is emitted as a release package and handed back through a signed HMAC-authenticated receiver with timestamp, nonce, replay protection, and release/hash validation.

Research

The isolated worker gathers public evidence from multiple source organizations.

Search-result snippets are discovery aids only and are never considered evidence.

Claims require at least two independent source organizations and are refused entirely for high-stakes, sensitive, or time-sensitive research classes.

Import and promotion

The app-side isolated lifecycle validates the release package, imports a research candidate, performs local verification, creates system-policy approval on PASS, promotes trusted knowledge, and indexes the promoted record.

A separate existing attestation path records the final trusted-knowledge asset in NexLedger.

Validation

The full path was validated using the Cuckoo-filter question. Four candidate claims were generated and all four passed claim verification. The resulting trusted-knowledge record was indexed and a NexLedger asset was confirmed.

CSV

metrics.csv

635 bytes SHA-256 dc3e6b8e1856817f… 
metric,value,unit_or_status
trust_after_trigger_max,0.72,score
best_rag_trigger_max,0.45,score
minimum_independent_source_organizations_per_claim,2,organizations
cuckoo_filter_evidence_sources,3,sources
cuckoo_filter_source_organizations,3,organizations
cuckoo_filter_candidate_claims,4,claims
cuckoo_filter_verified_claims,4,claims
cuckoo_filter_groq_requests,2,requests
cuckoo_filter_groq_tokens,11034,tokens
imported_candidate_id,82,candidate
trusted_knowledge_id,74,record
nexledger_asset,FORGEAI#TK123V1,asset
end_to_end_result,PASS,status
legacy_shared_lifecycle_modified,NO,boundary
search_snippets_used_as_evidence,NO,boundary
JSON

note.json

3,070 bytes SHA-256 7f38c092f723fad8… 
{
    "schema": "kespa.public_technical_note.v1",
    "id": "KESPA-NOTE-010",
    "title": "Autonomous Natural Research Gap Lifecycle",
    "date": "2026-09-15",
    "status": "verified",
    "purpose": "Document the isolated natural-research lifecycle that converts low-confidence public-knowledge gaps observed during normal KESPA use into evidence-backed trusted knowledge without modifying the existing manual-learning lifecycle.",
    "trigger_policy": {
        "response_source": "brain",
        "cache_required": false,
        "rag_searched": true,
        "allowed_goal": "learn",
        "trust_after_max": 0.72,
        "best_rag_score_max": 0.45,
        "excludes": [
            "private",
            "strict_internal",
            "memory",
            "high_stakes",
            "ordinary routing-only gaps"
        ],
        "classification": "public_general_knowledge"
    },
    "isolated_flow": [
        "normal user question",
        "gap detector",
        "signed research request queue",
        "home research poller",
        "isolated natural research worker",
        "multi-source evidence collection",
        "claim verification",
        "research release package",
        "signed app handoff",
        "isolated import",
        "local verification",
        "system-policy approval",
        "trusted-knowledge promotion",
        "local semantic indexing",
        "NexLedger attestation"
    ],
    "research_worker_policy": {
        "shared_legacy_lifecycle_scripts_modified": false,
        "shared_legacy_lifecycle_scripts_invoked": false,
        "search_snippets_used_as_evidence": false,
        "minimum_independent_source_organizations_per_claim": 2,
        "high_stakes_requests_refused": true,
        "time_sensitive_requests_refused": true,
        "sensitive_requests_refused": true,
        "model": "openai/gpt-oss-120b via Groq"
    },
    "validated_end_to_end_example": {
        "question": "What is a Cuckoo filter?",
        "request_id": "research-gap-20a905c55f6ad819c57f6920",
        "release_id": "forge-release-07990ffc7fc9d09f9be2dd3b",
        "release_sha256": "07990ffc7fc9d09f9be2dd3b5594285e5ee0789f7430e24fb1a812cbbf996ed3",
        "evidence_sources": 3,
        "independent_source_organizations": 3,
        "candidate_claims": 4,
        "verified_claims": 4,
        "groq_requests": 2,
        "groq_tokens_total": 11034,
        "candidate_id": 82,
        "trusted_knowledge_id": 74,
        "nexledger_asset": "FORGEAI#TK123V1",
        "lifecycle_result": "PASS"
    },
    "architecture_boundaries": {
        "manual_learning_lifecycle_modified": false,
        "legacy_shared_scripts_reused": false,
        "search_snippets_as_evidence": false,
        "trusted_knowledge_authoritative": true,
        "retrieval_index_rebuildable": true
    },
    "limitations": [
        "The Cuckoo-filter run proves one full end-to-end path; it is not a statistical estimate of autonomous research accuracy.",
        "The current worker deliberately refuses high-stakes, sensitive, and time-sensitive research requests.",
        "Multi-source agreement does not itself prove correctness; claims still require source/evidence verification.",
        "The system's long-term false-acceptance, false-rejection, and research-yield rates require continued measurement."
    ]
}
JSON

provenance.json

1,188 bytes SHA-256 098b005499956619… 
{
    "schema": "kespa.public_provenance.v1",
    "research_id": "KESPA-NOTE-010",
    "source_basis": "Historical KESPA operational/Jira records for the isolated natural-research lifecycle and its Cuckoo-filter end-to-end proof.",
    "source_observations": [
        "A conservative natural-gap detector queues public-general-knowledge research jobs from low-confidence RAG-backed interactions.",
        "The research worker is isolated from the legacy/manual lifecycle and uses search snippets only for discovery, never as evidence.",
        "Claims require support from at least two independent source organizations.",
        "The Cuckoo-filter proof produced three evidence sources, four candidate claims, four verified claims, candidate #82, trusted knowledge #74, and NexLedger asset FORGEAI#TK123V1.",
        "The release ID was forge-release-07990ffc7fc9d09f9be2dd3b with recorded SHA-256 07990ffc7fc9d09f9be2dd3b5594285e5ee0789f7430e24fb1a812cbbf996ed3."
    ],
    "publication_note": "This public note publishes lifecycle architecture and aggregate proof metrics only. HMAC secrets, Discord webhooks, private API keys, internal source code, and raw research evidence bodies are intentionally omitted."
}