← Research archive
Technical note published Sep 2, 2026

Evidence-Bound Knowledge Factory Pilot Design

KESPA's first clean Knowledge Factory pilot defined a 15-topic, 30-card evidence-bound generation contract with authoritative-source policies, independent organization checks for high-stakes topics, evidence hashing, atomic claim-to-source mappings, contamination guards, and deferred trust assignment.

knowledge-factory evidence-bound-generation provenance high-stakes pilot-design
MARKDOWN

README.md

3,219 bytes SHA-256 e1b28a94c1dcd1ad…

KESPA-NOTE-004 — Evidence-Bound Knowledge Factory Pilot Design

Date: 2026-09-02 Status: Published methodology note Project: KESPA AI / NexLabs Studios

Purpose

Before KESPA attempted large-scale knowledge generation, it built a deliberately small clean pilot:

  • 15 topics
  • 2 cards per topic
  • 30 planned cards
  • cross-domain technical, general, and high-stakes coverage

The goal was not to generate a large corpus quickly.

The goal was to establish the rules that would prevent generated prose from silently becoming trusted knowledge.

Evidence first

The pilot used curated authoritative source seeds and collected bounded evidence excerpts with provenance.

The evidence layer included:

  • evidence hashing;
  • source-organization metadata;
  • SSRF-safe URL handling;
  • redirect validation;
  • optional Google/DDG discovery fallback;
  • high-stakes independent-source requirements;
  • bounded excerpt collection rather than wholesale page mirroring.

For high-stakes topics, independence was measured at the organization level rather than by hostname alone.

That prevented multiple domains belonging to the same organization from being mistaken for independent corroboration.

Evidence quality gate

Before generation, collected evidence could be screened for:

  • duplicate evidence;
  • duplicate URLs;
  • anti-bot/error-page contamination;
  • prompt-injection-like text;
  • topic mismatch;
  • navigation/HTML artifacts;
  • insufficient independent organizations for high-stakes use.

Weak or unrelated second sources were not allowed to satisfy high-stakes cross-check requirements.

Card blueprint

Each pilot topic received an exact two-card contract.

The blueprint bound cards to:

  • specific collected evidence;
  • explicit allowed claim scopes;
  • topic-specific exclusions;
  • global synthetic-contamination exclusions;
  • evidence hashes and provenance;
  • focused retrieval-text seeds.

High-stakes cards received stronger source-relevance and review requirements.

Every planned card remained:

production_ready = false

until later factual verification.

Generation contract

The evidence-bound candidate generator added:

  • strict evidence-only prompting;
  • atomic claim-to-source mappings;
  • separate focused retrieval_text;
  • schema and length validation;
  • long verbatim-copy detection;
  • regression guards for known synthetic contamination;
  • personalized high-stakes advice guards;
  • audit/failure logging.

Generated candidates remained:

  • verified = false
  • production_ready = false

and received no trust assignment merely because a model generated them.

Why this mattered

This pilot established the core KESPA principle used by later experiments:

Generation proposes knowledge. Evidence and verification decide whether anything becomes trusted.

That distinction later enabled claim-level verification, clean corpus construction, frozen retrieval benchmarks, and the autonomous trusted-knowledge lifecycle.

Limitation

This record describes the pilot's design and safeguards.

The supplied historical changelog does not contain a complete final per-card outcome table for all 30 cards, so this public note does not invent one.

TEXT

SHA256SUMS.txt

393 bytes SHA-256 5ce528ac64328155… 
e1b28a94c1dcd1ad42478ecca34a3ef54d6568859a4a74a2b77c5668a9ada742  README.md
69ed0da1921bf463dd0e5512d2f74632e430ae34cb9be18a3e4debbdba940c29  note.json
e04b239c92b3b66cb585c22dc53e13d0e8054f293334e87908d21fccb5ceee0d  metrics.csv
c307ea649ad890e55b24ce7be4cc3a6133fb4bdd09b9021c2447993a08d8649b  methodology.md
31ec8c1ccfcecf649ccc178ec02cccad9044b0213ddd2a86826e466fa658eef1  provenance.json
MARKDOWN

methodology.md

1,349 bytes SHA-256 c307ea649ad890e5…

Methodology

Stage 1 — Pilot selection

A representative 15-topic cross-domain pilot was selected from the recovered knowledge plan.

Generation was capped at two cards per topic.

Stage 2 — Authoritative evidence acquisition

Curated authoritative sources were preferred. High-stakes topics required independent organizations rather than merely different hostnames.

Evidence was stored as bounded excerpts with hashes and provenance.

Stage 3 — Evidence audit

An offline evidence-quality gate checked duplicate content/URLs, anti-bot or error pages, prompt-injection-like text, deterministic topic relevance, navigation artifacts, and high-stakes source independence.

Stage 4 — Evidence-bound card blueprint

Each card received an explicit evidence binding, claim scope, exclusions, and retrieval-text seed.

The blueprint itself made no factual claims production-ready.

Stage 5 — Candidate generation

The generator used an evidence-only prompt and produced atomic claim-to-source mappings.

Candidate validation included schema/length checks, verbatim-copy detection, contamination regression guards, and high-stakes personalized-advice checks.

Stage 6 — Deferred trust

Generation did not assign trust.

All generated candidates remained unverified and non-production-ready until later factual verification.

CSV

metrics.csv

478 bytes SHA-256 e04b239c92b3b66c… 
metric,value,unit_or_status
pilot_topics,15,topics
planned_cards,30,cards
cards_per_topic,2,cards
trust_assigned_during_generation,NO,status
generated_verified_default,FALSE,status
generated_production_ready_default,FALSE,status
evidence_hashing,YES,control
organization_level_independence,YES,high_stakes_control
prompt_injection_text_detection,YES,evidence_audit
long_verbatim_copy_detection,YES,generation_guard
synthetic_contamination_regression_guards,YES,generation_guard
JSON

note.json

3,209 bytes SHA-256 69ed0da1921bf463… 
{
    "schema": "kespa.public_technical_note.v1",
    "id": "KESPA-NOTE-004",
    "title": "Evidence-Bound Knowledge Factory Pilot Design",
    "date": "2026-09-02",
    "status": "published",
    "purpose": "Document the methodology and safety contract used to construct KESPA's first evidence-bound clean knowledge-generation pilot before later scale experiments.",
    "pilot_scope": {
        "topics": 15,
        "planned_cards": 30,
        "cards_per_topic": 2,
        "coverage": "cross-domain technical, general, and high-stakes subjects"
    },
    "evidence_policy": {
        "authoritative_source_seeds": true,
        "evidence_excerpt_only_collection": true,
        "evidence_hashing": true,
        "source_organization_provenance": true,
        "unknown_sources_implicitly_trusted": false,
        "optional_general_search_fallback": true,
        "ssrf_safe_url_validation": true,
        "redirect_validation": true,
        "high_stakes_independent_organization_checks": true
    },
    "quality_gate": {
        "duplicate_evidence_detection": true,
        "duplicate_url_detection": true,
        "antibot_error_page_detection": true,
        "prompt_injection_text_detection": true,
        "deterministic_topic_relevance_checks": true,
        "navigation_html_artifact_warnings": true,
        "high_stakes_usable_source_organization_gate": true
    },
    "card_contract": {
        "exact_two_card_contract_per_topic": true,
        "allowed_claim_scopes": true,
        "topic_specific_exclusions": true,
        "global_synthetic_contamination_exclusions": true,
        "high_stakes_source_relevance_gate": true,
        "food_safety_card_level_review_requirements": true,
        "evidence_hash_provenance_binding": true,
        "focused_retrieval_text_seed": true,
        "production_ready_default": false
    },
    "generation_contract": {
        "evidence_only_system_prompt": true,
        "atomic_claim_to_source_mapping": true,
        "focused_retrieval_text_generation": true,
        "candidate_schema_validation": true,
        "candidate_length_validation": true,
        "long_verbatim_copy_detection": true,
        "synthetic_contamination_regression_guards": true,
        "high_stakes_personalized_advice_guard": true,
        "trust_assignment_deferred": true,
        "generated_candidate_verified_default": false,
        "generated_candidate_production_ready_default": false
    },
    "source_policy_refinements": [
        "CBT was promoted to high-stakes verification policy.",
        "High-stakes independence moved from hostname count to source-organization count.",
        "Weak unrelated secondary sources were prevented from satisfying independent verification.",
        "Blocked or weak pilot sources were replaced with stronger authoritative alternatives.",
        "Known source-organization aliases were prevented from falsely counting as independent organizations."
    ],
    "limitations": [
        "This note documents the pilot design and safeguards; it is not itself a measured quality benchmark.",
        "The historical change records supplied here do not include a complete final per-card outcome table for all 30 pilot cards.",
        "Source-policy choices reflect the pilot's 2026-09 design and should not be interpreted as permanent universal source rankings.",
        "Generated candidates remained untrusted until later factual verification."
    ]
}
JSON

provenance.json

1,033 bytes SHA-256 31ec8c1ccfcecf64… 
{
    "schema": "kespa.public_provenance.v1",
    "research_id": "KESPA-NOTE-004",
    "source_basis": "Historical KESPA changelog/Jira export supplied for public research reconstruction.",
    "source_observations": [
        "A representative 15-topic pilot was capped at two cards per topic.",
        "Evidence collection used curated authoritative source seeds, hashes, and organization-level provenance.",
        "High-stakes cross-checking was tightened so weak unrelated sources could not satisfy independence.",
        "The pilot blueprint bound every card to collected evidence and kept all cards production_ready=false.",
        "The candidate generator used evidence-only prompting, atomic claim-to-source mappings, validation guards, and deferred all trust assignment."
    ],
    "publication_note": "This public note contains pilot methodology and aggregate scope only. Raw fetched third-party evidence text, internal source-policy configuration, private implementation scripts, credentials, and generated candidate bodies are not republished."
}