Research update published Sep 5, 2026

KESPA V2 Research Program: Verified Intelligence and Adaptive Compute

KESPA V2 formalized the research direction emerging from the completed A/B/C benchmark: verified externalized intelligence and fast trusted retrieval as the default path, adaptive orchestration only when justified, explicit separation of evidence/knowledge/memory, continuous knowledge growth, and compute-efficiency as a first-class research objective.

research-program master-plan-v2 verified-intelligence adaptive-compute nsf

MARKDOWN

README.md

3,912 bytes SHA-256 e2d3c5f9477ef909…

KESPA-UPDATE-001 — KESPA V2 Research Program

Date: 2026-09-05 Status: Published research-program update Project: KESPA AI / NexLabs Studios

Why V2 changed

The completed A/B/C benchmark produced a strong architectural signal:

Raw local model quality: 20.71
Clean verified retrieval quality: 54.22
Then-current KESPA orchestration quality: 27.71

The clean verified-retrieval arm also recorded:

1.675 s mean latency
0.1085 Wh mean GPU energy
54/54 retrieval Hit@1
79.17% gold retrieval recall
39 case-quality wins

The important conclusion was not "KESPA built a better language model."

It was that the intelligence supplied to a small model, and the policy used to retrieve and trust it, can materially change useful system behavior and compute cost.

V2 research thesis

The program was reframed around this question:

Can an intelligence-management system improve reliability, efficiency, provenance, and practical

usefulness of relatively small/local language models by supplying and maintaining verified intelligence,

selectively allocating compute, and preventing untrusted information from silently becoming persistent knowledge?

Architecture direction

V2 made several ideas explicit:

verified externalized intelligence over uncontrolled accumulation;
trusted retrieval as a fast path;
adaptive rather than always-on orchestration;
evidence, trusted knowledge, and memory as separate data classes;
provenance-preserving continuous knowledge growth;
quality, latency, energy, and compute measured together.

In short:

The model reasons. KESPA supplies and maintains the intelligence.

Research hypotheses

H1 — Verified knowledge can improve small-model usefulness

Initial A/B/C evidence supported this on the frozen workload, but replication across other workloads and models remained necessary.

H2 — Verified retrieval can reduce recurring inference compute

The initial B arm was both higher-scoring and less expensive in latency/energy than A and C.

That made retrieval quality an efficiency question, not merely a relevance question.

H3 — More orchestration is not automatically better

The then-current orchestration arm consumed substantially more latency and energy without matching the verified-retrieval arm's benchmark quality.

V2 therefore moved away from always-on orchestration.

H4 — Adaptive compute should be tested directly

Instead of running every mechanism on every request, KESPA should escalate only when the request or available intelligence justifies the additional compute.

At the time of this V2 revision, that hypothesis had not yet been validated. D v001 was run afterward.

H5 — Continuous learning needs a trust lifecycle

A system that continuously stores model output will eventually recreate the contamination problem unless it separates:

candidate -> provenance -> evidence -> verification -> conflict detection -> trust decision -> promotion

At the time of the V2 revision this was a program objective. It was later operationalized.

H6 — Knowledge quality may substitute for some model scale

KESPA's longer-term question is whether better verified intelligence and reusable solved work can allow smaller local models to cover useful bounded workloads without always requiring larger inference models.

That remains a research hypothesis, not a universal established result.

What V2 did not claim

V2 did not claim that:

KESPA trained a superior foundation model;
one benchmark proved general intelligence gains;
verified retrieval always beats larger models;
adaptive inference was already solved;
future NSF or other funding was guaranteed.

It converted the initial benchmark evidence into a more rigorous research program with falsifiable questions and measurable system-level tradeoffs.

TEXT

SHA256SUMS.txt

317 bytes SHA-256 24a0715df6088f3e…

e2d3c5f9477ef909786a6d4bb02d86a9ad000ffc59660be8962744849ddba0d4  README.md
2c44606abc3388970401788caada64f33c114b94ae1074af15bd16a45dcce481  update.json
c7ed4b979d2f65329bf86485f95894448335e56cecd158fffb4a193fc5312baf  hypotheses.csv
43693927a724173e7b481e0ef2d843dc75165369004595273b813fcc3190a6db  provenance.json

CSV

hypotheses.csv

1,204 bytes SHA-256 c7ed4b979d2f6532…

id,hypothesis,status_at_v2_revision
H1,Verified externalized knowledge can improve useful small-model output relative to the same local model without retrieval.,supported by initial A/B/C benchmark; requires replication/generalization
H2,Verified retrieval can reduce recurring inference compute for knowledge-rich workloads.,supported by initial latency/energy comparison; requires broader workload testing
H3,Always-on orchestration can destroy the efficiency advantage of clean retrieval when its additional compute does not produce proportional quality gains.,supported by A/B/C C-arm result; architecture-specific
H4,Adaptive inference can preserve quality while selectively allocating additional compute only where necessary.,planned / not yet validated at V2 revision time
H5,"Continuous knowledge growth can remain useful only if candidate intelligence is separated from trusted knowledge through evidence, verification, conflict detection, and auditable promotion.",planned lifecycle hypothesis; later operationalized
H6,Knowledge quality and provenance may substitute for some model-scale increases on bounded private/local workloads.,research hypothesis; not established as a universal result

JSON

provenance.json

858 bytes SHA-256 43693927a724173e…

{
    "schema": "kespa.public_provenance.v1",
    "research_id": "KESPA-UPDATE-001",
    "source_basis": "Historical KESPA changelog/Jira export supplied for public research reconstruction.",
    "source_observations": [
        "Master Plan V2 was recorded on 2026-09-05 as a canonical project-wide architecture/research/funding revision.",
        "The revision was explicitly grounded in completed A/B/C benchmark measurements.",
        "At that revision point, D had not yet been executed and the continuous evidence lifecycle had not yet been implemented.",
        "The source record framed NSF as a parallel research/funding program, not as awarded funding."
    ],
    "publication_note": "This update publishes the research thesis, hypotheses, and aggregate benchmark basis. Private implementation plans, internal credentials, and unpublished source code are not included."
}

JSON

update.json

4,244 bytes SHA-256 2c44606abc338897…

{
    "schema": "kespa.public_update.v1",
    "id": "KESPA-UPDATE-001",
    "title": "KESPA V2 Research Program: Verified Intelligence and Adaptive Compute",
    "date": "2026-09-05",
    "status": "published",
    "basis": "Master Plan V2 revision grounded in the completed A/B/C benchmark and the emerging clean-knowledge architecture.",
    "core_research_thesis": "Can an intelligence-management system improve reliability, efficiency, provenance, and practical usefulness of relatively small/local language models by supplying and maintaining verified intelligence, selectively allocating compute, and preventing untrusted information from silently becoming persistent knowledge?",
    "design_principles": [
        "Verified externalized intelligence should be preferred over uncontrolled corpus accumulation.",
        "Fast trusted retrieval should be the default path when sufficient verified knowledge exists.",
        "Expensive orchestration should be adaptive rather than always-on.",
        "Evidence, trusted knowledge, and memory are distinct data classes with different lifecycle rules.",
        "Continuous knowledge acquisition must preserve provenance and explicit trust transitions.",
        "Compute, latency, and energy should be evaluated together with answer quality."
    ],
    "empirical_basis": {
        "A_raw_quality": 20.71,
        "B_verified_retrieval_quality": 54.22,
        "C_then_current_orchestration_quality": 27.71,
        "B_latency_seconds": 1.675,
        "B_gpu_energy_wh": 0.1085,
        "B_retrieval_hit_at_1": "54/54",
        "B_gold_recall_percent": 79.17,
        "B_case_quality_wins": 39
    },
    "research_hypotheses": [
        {
            "id": "H1",
            "hypothesis": "Verified externalized knowledge can improve useful small-model output relative to the same local model without retrieval.",
            "status_on_2026_09_05": "supported by initial A/B/C benchmark; requires replication/generalization"
        },
        {
            "id": "H2",
            "hypothesis": "Verified retrieval can reduce recurring inference compute for knowledge-rich workloads.",
            "status_on_2026_09_05": "supported by initial latency/energy comparison; requires broader workload testing"
        },
        {
            "id": "H3",
            "hypothesis": "Always-on orchestration can destroy the efficiency advantage of clean retrieval when its additional compute does not produce proportional quality gains.",
            "status_on_2026_09_05": "supported by A/B/C C-arm result; architecture-specific"
        },
        {
            "id": "H4",
            "hypothesis": "Adaptive inference can preserve quality while selectively allocating additional compute only where necessary.",
            "status_on_2026_09_05": "planned / not yet validated at V2 revision time"
        },
        {
            "id": "H5",
            "hypothesis": "Continuous knowledge growth can remain useful only if candidate intelligence is separated from trusted knowledge through evidence, verification, conflict detection, and auditable promotion.",
            "status_on_2026_09_05": "planned lifecycle hypothesis; later operationalized"
        },
        {
            "id": "H6",
            "hypothesis": "Knowledge quality and provenance may substitute for some model-scale increases on bounded private/local workloads.",
            "status_on_2026_09_05": "research hypothesis; not established as a universal result"
        }
    ],
    "program_changes": [
        "Trust emphasis moved toward ingest/promotion rather than assuming stored content was trustworthy.",
        "Fast trusted retrieval became the intended default.",
        "Orchestration became adaptive rather than mandatory.",
        "Evidence, knowledge, and memory were explicitly separated.",
        "Daily-use KESPA and continuous knowledge accumulation became immediate objectives.",
        "The research/funding program became a parallel workstream rather than downstream documentation."
    ],
    "limitations": [
        "The V2 update is a research-program revision, not an experimental result by itself.",
        "The empirical basis came from one frozen corpus/workload, one local model/hardware configuration, and one judge setup.",
        "At the time of the revision, adaptive D had not yet been executed and the continuous evidence lifecycle had not yet been operationalized.",
        "Any future funding application is competitive and not guaranteed."
    ]
}