← Research archive
Technical note published Sep 1, 2026

Source-Organization Independence and High-Stakes Evidence Policy

KESPA hardened its pilot evidence policy by moving independence from hostname counting to source-organization counting, adding explicit source_organization provenance, strengthening high-stakes cross-check requirements, replacing weak/failed seeds with authoritative sources, and preserving evidence-excerpt, hashing, redirect, and SSRF safety rails. Syntax and offline cross-check policy tests passed; no LLM, Chroma, or Brain changes were involved.

evidence-policy source-independence high-stakes provenance knowledge-factory
MARKDOWN

README.md

4,149 bytes SHA-256 3e470c77dc3c431f…

KESPA-NOTE-022 — Source-Organization Independence and High-Stakes Evidence Policy

Date: 2026-09-01 Status: Published historical technical record Project: KESPA AI / NexLabs Studios

Purpose

One of the first Knowledge Factory safety corrections was deceptively simple:

Two different domains do not necessarily mean two independent sources.

KESPA initially had evidence mechanics that could reason about source/domain counts.

That was not strong enough for high-stakes corroboration.

The policy was hardened so evidence independence is measured by source organization, not merely hostname.

Independence correction

The pilot evidence system changed from:

hostname/domain count

to:

source-organization count

Evidence records gained explicit:

source_organization

provenance.

Evidence reports also gained organization counts.

This prevents one organization operating several domains from accidentally satisfying an independence requirement multiple times.

A specific correction also prevented SEC-related domains from being falsely counted as separate independent organizations.

High-stakes policy

Cognitive Behavioral Therapy (CBT) was promoted into the high-stakes verification policy.

The evidence path added independent cross-checking rather than allowing a single authoritative-looking source to satisfy the full requirement.

The same architecture supports stricter handling across health, legal, finance, security, and other higher-risk knowledge domains.

Pilot source hardening

Several pilot source sets were corrected.

CBT

Failed seeds were replaced with sources from:

  • NIMH
  • APA
  • VA

Cooking / food safety

Failed source URLs were replaced with:

  • FoodSafety.gov
  • current USDA FSIS material

Investing

The U.S. Department of Labor was added as an independent evidence source.

Copyright

Official U.S. Code Title 17 was added as independent evidence.

Additional evidence rails

The same implementation period also added or preserved:

  • evidence-excerpt-only collection;
  • explicit source-usage review;
  • evidence hashing;
  • provenance metadata;
  • SSRF-safe public URL handling;
  • redirect validation;
  • high-stakes multi-domain evidence gates;
  • curated authoritative source seeds;
  • optional Google/DDG fallback discovery.

OpenStax was removed from the automated AI-ingestion source policy during this hardening pass.

Unknown/new source families were intended to default toward review rather than implicit trust.

Why organization identity matters

Consider three pages:

``text docs.example.org research.example.org support.example.com ``

Counting hostnames would produce:

3 sources

But if all three are controlled by the same organization, they are not three independent corroborators.

KESPA's corrected model asks:

How many independent source organizations support this?

not merely:

How many URLs/domains did we collect?

That is a stronger basis for evidence policy.

Separation from generation

This work changed source qualification and evidence provenance.

It did not:

  • call an LLM;
  • generate knowledge cards;
  • write to Chroma;
  • change brain_api.py.

That separation allowed evidence-policy mechanics to be tested without confusing source quality with model behavior.

Validation

The historical implementation record reports:

  • syntax validation: PASS
  • offline cross-check policy tests: PASS
  • LLM calls: 0
  • Chroma changes: 0
  • Brain API changes: 0

Research significance

This correction is part of a broader KESPA principle:

Corroboration should represent genuinely independent evidence, not superficial URL diversity.

That matters particularly for automated knowledge systems because a naïve source counter can manufacture confidence from multiple pages that all trace back to the same institution.

Limitation

This note records policy mechanics and offline validation.

It does not claim a measured reduction in false factual acceptance across a statistically representative high-stakes dataset.

That question requires separate empirical evaluation.

TEXT

SHA256SUMS.txt

393 bytes SHA-256 1ffbc665ea3fe32a… 
3e470c77dc3c431f52315b9096de25d3c68fa56bc1b26f3148440f8cc1039038  README.md
29877de4628b9b43b25a13436b0cbf3be165a41bbe6b03bfbc4563fe810cd62f  note.json
57152f11ccb2b176db7f763a82cbeb120c5cdcd4ad301ab4543ed44ed2255a3b  metrics.csv
ca911b7ea09c1765da801795e2225fc4e26f0c96b4a31df53e357d512166b1a3  methodology.md
b8429fa3d64c6ed0a98bec8230d2b1c76aa024be9648f2ad6498d042810d868e  provenance.json
MARKDOWN

methodology.md

1,027 bytes SHA-256 ca911b7ea09c1765…

Methodology

Evidence-source identity

Each evidence item is associated with a source organization.

Independence gates count distinct organizations rather than blindly counting domains or URLs.

High-stakes corroboration

Higher-risk topics require independent cross-checking under stricter evidence policy.

The CBT pilot was explicitly moved into this class during the September 1 hardening.

Curated source correction

Failed or unsuitable pilot seeds were replaced with more authoritative source families for CBT, food safety, investing, and copyright.

Network and provenance rails

Evidence acquisition uses public-URL/redirect safety checks, bounded evidence excerpts, hashing, and provenance metadata.

Source usage is reviewed through explicit policy rather than inferred solely from successful HTTP retrieval.

Validation

The recorded change passed syntax validation and offline cross-check policy tests.

No LLM generation, Chroma mutation, or Brain API change was part of this policy validation.

CSV

metrics.csv

636 bytes SHA-256 57152f11ccb2b176… 
metric,value,unit_or_status
independence_basis,source_organization,policy
source_organization_provenance,YES,control
organization_counts_in_reports,YES,control
high_stakes_cross_checking,YES,policy
cbt_promoted_to_high_stakes,YES,policy
evidence_excerpt_only,YES,policy
evidence_hashing,YES,provenance
ssrf_safe_public_url_validation,YES,security
redirect_validation,YES,security
openstax_automated_ingestion_policy,REMOVED,policy
sec_multi_domain_false_independence,BLOCKED,policy
syntax_validation,PASS,validation
offline_cross_check_policy_tests,PASS,validation
llm_calls,0,calls
chroma_changes,0,changes
brain_api_changes,0,changes
JSON

note.json

3,166 bytes SHA-256 29877de4628b9b43… 
{
    "schema": "kespa.public_technical_note.v1",
    "id": "KESPA-NOTE-022",
    "title": "Source-Organization Independence and High-Stakes Evidence Policy",
    "date": "2026-09-01",
    "status": "published",
    "purpose": "Document the evidence-policy hardening that changed KESPA's definition of independent corroboration from hostname diversity to source-organization diversity and strengthened source handling for high-stakes pilot topics.",
    "policy_changes": {
        "independence_unit_before": "hostname/domain count",
        "independence_unit_after": "source organization",
        "source_organization_provenance_added": true,
        "organization_counts_added_to_evidence_reports": true,
        "same_organization_multi_domain_false_independence_blocked": true,
        "high_stakes_cross_checking_strengthened": true,
        "evidence_excerpt_only_collection": true,
        "explicit_source_usage_review_required": true,
        "evidence_hashing_and_provenance": true,
        "ssrf_safe_public_url_validation": true,
        "redirect_validation": true
    },
    "pilot_source_hardening": [
        {
            "topic": "Cognitive Behavioral Therapy (CBT)",
            "changes": [
                "promoted to high-stakes verification policy",
                "added independent cross-checking",
                "replaced failed seeds with NIMH, APA, and VA sources"
            ]
        },
        {
            "topic": "Cooking / food safety",
            "changes": [
                "replaced failed source URLs with FoodSafety.gov and current USDA FSIS sources"
            ]
        },
        {
            "topic": "Investing",
            "changes": [
                "added U.S. Department of Labor as an independent evidence source"
            ]
        },
        {
            "topic": "Copyright",
            "changes": [
                "added official U.S. Code Title 17 as independent evidence"
            ]
        }
    ],
    "source_policy_correction": {
        "openstax_removed_from_automated_ai_ingestion_source_policy": true,
        "sec_domains_prevented_from_false_independence": true,
        "unknown_source_default": "REVIEW_REQUIRED"
    },
    "validation": {
        "syntax_validation": "PASS",
        "offline_cross_check_policy_tests": "PASS",
        "llm_calls": 0,
        "chroma_changes": 0,
        "brain_api_changes": 0
    },
    "interpretation": [
        "Two hostnames are not independent evidence if they belong to the same underlying organization.",
        "Source-organization provenance makes corroboration auditable and reduces false confidence from domain-count inflation.",
        "Higher-risk domains require stricter source diversity and human/policy review rather than weaker evidence thresholds.",
        "Evidence collection is intentionally separated from model generation so source qualification can be audited independently."
    ],
    "limitations": [
        "The September 1 record validates policy mechanics and offline cross-check behavior, not a broad empirical accuracy rate for high-stakes knowledge.",
        "Source-organization identity still depends on correct source cataloging and provenance metadata.",
        "The policy was designed for the pilot and can require later source-family maintenance as authoritative sites change.",
        "This note does not claim that institutional authority alone proves factual correctness."
    ]
}
JSON

provenance.json

1,563 bytes SHA-256 b8429fa3d64c6ed0… 
{
    "schema": "kespa.public_provenance.v1",
    "research_id": "KESPA-NOTE-022",
    "source_basis": "Historical KESPA Knowledge Factory evidence-policy changelog records dated 2026-09-01.",
    "source_observations": [
        "CBT was promoted to high-stakes verification and given independent cross-check requirements.",
        "Pilot CBT sources were replaced with NIMH, APA, and VA; cooking sources were replaced with FoodSafety.gov and USDA FSIS; DOL was added for investing and U.S. Code Title 17 for copyright.",
        "Evidence independence changed from hostname count to source-organization count.",
        "source_organization provenance and organization counts were added to evidence records/reports.",
        "SEC domains were prevented from falsely counting as independent organizations.",
        "Evidence-excerpt-only collection, evidence hashing/provenance, SSRF-safe public URL validation, and redirect validation were part of the evidence policy.",
        "Syntax validation and offline cross-check policy tests passed; no LLM calls, Chroma changes, or brain_api.py changes were involved."
    ],
    "status_note": "This entry is marked published rather than verified because the source record demonstrates implementation and offline policy-test success, but does not present a separate user-environment empirical outcome measuring high-stakes factual accuracy.",
    "publication_note": "This package describes source-policy mechanics and validation results. It does not republish third-party evidence text, private configuration, credentials, or proprietary source code."
}