Experiment published Sep 14, 2026

Autonomous Trust Policy v0.1 to v0.2 Backlog Reclassification

KESPA relaxed one auto-publication condition from two supporting evidence items to one while retaining verification, confidence, consensus, and zero-conflict requirements. Applying the policy change to the existing backlog made 29 previously held candidates eligible without rerunning evidence collection or local verification, raising projected backlog autonomy from 36.7% to approximately 95.9%.

autonomous-learning trust-policy auto-publish human-review policy-tuning

MARKDOWN

README.md

2,854 bytes SHA-256 fd269ccf61e5f7a5…

KESPA-EXP-006 — Autonomous Trust Policy v0.1 → v0.2 Backlog Reclassification

Date: 2026-09-14 Status: Published policy experiment Project: KESPA AI / NexLabs Studios

Research question

Can KESPA reduce unnecessary human-review load by changing only the supporting-evidence threshold while preserving its other verification, consensus, confidence, and conflict gates?

Policy v0.1

The original autonomous publication policy required:

verification: PASS
confidence: >= 0.90
AI consensus: yes
qualifying evidence: >= 2
supporting evidence: >= 2
conflicting evidence: 0
verifier conflict: false

Successful autonomous approvals were recorded as:

approver_type = system_policy
policy_version = forge-auto-publish-policy-v0.1.0

Observed product problem

The backlog was producing more human-review cases than desired for routine public knowledge.

The research concern was not simply "make everything green."

The goal was to determine whether otherwise strong candidates were being held because the publication policy required two independently classified supporting evidence items even when verification, consensus, confidence, and conflict checks were already favorable.

Policy v0.2

The policy change was deliberately narrow:

minimum supporting evidence: 2 -> 1

The source record states that all other trust, conflict, and consensus requirements were retained.

System-generated publish jobs also used real system_policy approval metadata aligned to:

forge-auto-publish-policy-v0.2.0

Backlog reclassification result

Applying the revised policy to the already processed backlog made:

29 previously held candidates newly eligible
with no evidence recollection
and no local reverification

Recorded autonomy estimate:

v0.1 backlog autonomy: 36.7%
v0.2 projected backlog autonomy: ~95.9%
absolute change: +59.2 percentage points

Why this is useful experimentally

This is cleaner than regenerating the candidates under a new pipeline.

The upstream candidate evidence and local verification remained the same. The changed variable was the publication-policy threshold for supporting evidence.

That makes the backlog useful as a policy-reclassification study.

Important limitation

~95.9% is a projected/expected backlog autonomy figure, not a completed prospective production success rate.

A higher automation rate is not enough to validate the policy.

The next scientifically useful comparison should measure whether v0.2 changes:

false acceptance;
false rejection;
human-review rate;
conflict detection;
evidence quality;
verification accuracy;
latency/compute/energy.

The objective remains maximizing safe autonomous knowledge acquisition, not maximizing auto-publication by itself.

TEXT

SHA256SUMS.txt

399 bytes SHA-256 504780048415f86a…

fd269ccf61e5f7a5053424606007bc51299879ee06fe40406379ee397a0607ac  README.md
c83a763d4b551bc9e73c92303c7d164afa5822ed47a8dde86752d00baf98f464  experiment.json
1bf6fc2c57b9c7e50242a8184981946319fabd4e693483f0113fcb7a3423d204  results.csv
02eda1b81e7ec05786a3ebd9897d7cfe4b74af5cf6ca31801e3e006c84fe0566  methodology.md
d7b3eca2e462646464c907903f8b34c27446fd157be461b7d46abc3cc67ae968  provenance.json

JSON

experiment.json

2,884 bytes SHA-256 c83a763d4b551bc9…

{
    "schema": "kespa.public_experiment.v1",
    "id": "KESPA-EXP-006",
    "title": "Autonomous Trust Policy v0.1 to v0.2 Backlog Reclassification",
    "date": "2026-09-14",
    "status": "published",
    "research_question": "Can KESPA reduce unnecessary human-review load by changing only the supporting-evidence threshold while preserving its other verification, consensus, confidence, and conflict gates?",
    "policy_v0_1": {
        "verification": "PASS required",
        "minimum_confidence": 0.9,
        "ai_consensus": "required",
        "minimum_qualifying_evidence": 2,
        "minimum_supporting_evidence": 2,
        "conflicting_evidence": 0,
        "verifier_conflict": false,
        "recorded_approver_type": "system_policy",
        "policy_version": "forge-auto-publish-policy-v0.1.0"
    },
    "policy_v0_2": {
        "supporting_evidence_change": "2 -> 1",
        "other_trust_conflict_consensus_requirements": "retained",
        "system_generated_publish_jobs": "real system_policy approval",
        "approval_audit_policy_version": "forge-auto-publish-policy-v0.2.0"
    },
    "backlog_reclassification": {
        "previously_held_candidates_newly_eligible": 29,
        "evidence_recollection_required": false,
        "local_reverification_required": false,
        "baseline_autonomy_percent": 36.7,
        "projected_autonomy_percent": 95.9,
        "absolute_percentage_point_change": 59.2
    },
    "interpretation": [
        "The dominant blocker in the backlog was not necessarily model disagreement or detected contradiction; the recorded policy-tuning rationale identified the supporting-evidence count as the condition being relaxed.",
        "The policy change was intentionally narrow: only the supporting-evidence minimum changed from two to one.",
        "Because evidence collection and local verification were not rerun for the 29 newly eligible candidates, the change isolates publication-policy behavior from upstream evidence-generation behavior."
    ],
    "research_value": {
        "comparison": "policy v0.1 versus policy v0.2",
        "future_measures": [
            "observed auto-publication rate",
            "human-review rate",
            "false acceptance rate",
            "false rejection rate",
            "verification accuracy",
            "evidence quality",
            "latency",
            "compute",
            "energy"
        ]
    },
    "limitations": [
        "The approximately 95.9% autonomy figure is explicitly recorded as expected/projected backlog autonomy, not a completed prospective production measurement.",
        "The supplied source record does not provide the full backlog denominator in a directly auditable table, so this package preserves the reported percentages rather than reverse-engineering candidate counts.",
        "Making a candidate policy-eligible does not prove the candidate is factually correct.",
        "This policy change should be evaluated prospectively for false acceptance and false rejection, not judged only by higher automation rate."
    ]
}

MARKDOWN

methodology.md

987 bytes SHA-256 02eda1b81e7ec057…

Methodology

Experimental structure

The historical backlog was first processed under forge-auto-publish-policy-v0.1.0.

The policy required two supporting evidence items in addition to verification PASS, minimum confidence, AI consensus, qualifying-evidence, and zero-conflict conditions.

The v0.2 policy changed only the supporting-evidence threshold from two to one.

Isolation of the changed variable

The source record explicitly states that 29 previously held candidates became eligible:

without rerunning evidence collection;
without rerunning local verification.

This allows the public record to treat the result as a policy reclassification rather than a new evidence-generation experiment.

Measurement boundary

The recorded 36.7% -> ~95.9% change is a projected backlog-autonomy change.

This package does not convert that projection into an observed production success claim.

A prospective follow-up should track both automation gains and trust errors.

JSON

provenance.json

1,102 bytes SHA-256 d7b3eca2e4626464…

{
    "schema": "kespa.public_provenance.v1",
    "research_id": "KESPA-EXP-006",
    "source_basis": "Historical KESPA changelog/Jira export supplied for public research reconstruction.",
    "source_observations": [
        "Auto-Publish Policy v0.2 reduced supporting-evidence requirement from 2 to 1.",
        "All other trust/conflict/consensus requirements were retained.",
        "29 previously held candidates became eligible without rerunning evidence collection or local verification.",
        "Expected backlog autonomy increased from 36.7% to approximately 95.9%.",
        "System-generated publish jobs used system_policy approval and v0.2 audit metadata."
    ],
    "verification_note": "The supplied changelog records the policy change and projected backlog effect but does not include the complete candidate-level backlog table. This package therefore preserves the reported aggregate figures and labels the 95.9% figure as projected.",
    "publication_note": "No encrypted candidate/evidence content, private database rows, internal implementation code, Discord payloads, or private approvals are republished."
}

CSV

results.csv

574 bytes SHA-256 1bf6fc2c57b9c7e5…

metric,policy_v0_1,policy_v0_2,interpretation
minimum_supporting_evidence,2,1,only explicitly relaxed evidence threshold
verification_pass_required,YES,YES,retained
minimum_confidence,0.90,0.90,retained
ai_consensus_required,YES,YES,retained
conflicting_evidence_allowed,0,0,retained
verifier_conflict_allowed,NO,NO,retained
newly_eligible_held_candidates,0,29,reclassified without evidence recollection or reverification
backlog_autonomy_percent,36.7,~95.9,v0.2 value recorded as projected/expected
absolute_percentage_point_change,,+59.2,derived from recorded percentages