Technical note verified Sep 24, 2026

Zero-User Curriculum Learning and Evidence-Gated Autonomous Research

KESPA operationalized a scheduled zero-user learning driver that selects research opportunities from an explicit curriculum and other gap signals, queues asynchronous research work, and keeps Trusted Knowledge promotion and public publication independently fail-closed. The first recorded production daily-learning run completed and queued four public research requests; 24 curriculum domains were enabled, a 30-day maximum verification-age safeguard was active, background lifecycle workers were scheduled, and the production readiness audit reported zero failures with one non-blocking storage warning. Full downstream completion of every queued research request and large-scale autonomous promotion remain outside this verified checkpoint.

autonomous-learning zero-user-learning curriculum evidence-gating scheduled-research knowledge-lifecycle auditability

MARKDOWN

README.md

9,544 bytes SHA-256 99f201a621d68f8f…

KESPA-NOTE-029 — Zero-User Curriculum Learning and Evidence-Gated Autonomous Research

Summary

KESPA established an operational zero-user learning driver that can create useful research work without waiting for an active chat. The system combines an explicit curriculum, scheduled daily learning, background lifecycle workers, knowledge-freshness policy, benchmark-gap inputs, and hard publication/promotion safety boundaries so autonomous activity can generate research opportunities without turning research quotas into unverified Trusted Knowledge.

This note records the production checkpoint where the scheduling, persistence, curriculum selection, queue initiation, and key fail-closed controls were observed working. It does not claim that every queued research request has been proven to complete every downstream evidence, verification, conflict, promotion, indexing, and reverification stage under all operating conditions.

Why this is distinct from earlier KESPA autonomous-research work

Earlier KESPA research records documented user-triggered knowledge-gap acquisition, the natural research-gap lifecycle, evidence qualification, signed cross-host handoff, Trusted Knowledge promotion, and the foundation for freshness/reverification.

This checkpoint adds a different control loop:

``text no active user required → curriculum / benchmark gaps / stale knowledge → scheduled research opportunity selection → queued research work → asynchronous lifecycle processing → evidence and verification gates → eligible knowledge lifecycle handling → freshness / reverification → repeat ``

The material difference is that research initiation itself can now be proactive and scheduled, while promotion and publication remain independently gated.

Curriculum-driven learning

The recorded production curriculum contains 24 enabled research domains. Curriculum entries support operational fields such as topic identity, research directive, priority, minimum interval, enabled state, previous queue time, next due time, and research count.

Selection is intended to favor due and under-researched areas rather than repeatedly consuming the same topic. The source describes bulk curriculum ingestion as future work; large-scale importer functionality is therefore not claimed as operational in this release.

Evidence-gated learning lifecycle

The intended learning path is:

``text curriculum / knowledge gaps / stale knowledge → research request → research processing → source and evidence collection → claim extraction → verification → conflict checking → knowledge lifecycle handling → eligible Trusted Knowledge → freshness / reverification ``

A research request is not equivalent to a Trusted Knowledge card.

The source explicitly establishes the desired fail-closed behavior:

``text research attempted: yes evidence insufficient: yes trusted card created: no ``

That outcome is considered successful when the evidence does not justify promotion. KESPA is intended to prefer zero new cards over unsupported cards.

Scheduling and background execution

The autonomous-learning driver is connected to operating-system scheduling.

The recorded daily-learning timer is installed and enabled, with the default cycle occurring at approximately 02:00. The corresponding service is intentionally one-shot: after a successful execution it exits normally while the timer remains active.

Separate lifecycle workers are scheduled on an approximately 5-minute cadence. These workers support activities including research-inbox processing, knowledge refresh, knowledge lifecycle processing, and artifact enrichment. Filesystem locking is used to avoid overlapping worker execution when a previous cycle is still running.

First production zero-user run

The first recorded autonomous daily-learning run completed successfully.

Recorded result:

daily-learning status: complete
queued public research requests: 4
automatic public publication: off / human-gated

A later readiness audit also reported the most recent daily-learning run as complete.

This demonstrates that the scheduler, autonomous-learning persistence layer, curriculum selection path, and research-queue initiation were operational at the recorded checkpoint. It does not establish universal downstream completion for every queued request.

Knowledge freshness safeguard

The production checkpoint also records an active maximum verification-age ceiling of 30 days.

This ceiling is an additional safeguard rather than a claim that all knowledge changes on the same schedule. KESPA retains differentiated freshness classes such as static, slow-changing, volatile, highly volatile, event-driven, and manual.

Benchmark-gap feeder

Benchmark failures and knowledge gaps may feed new research opportunities without copying benchmark prompt/gold content into the learning pipeline.

The intended separation is:

``text benchmark exposes weakness → identify topic/category gap → create research opportunity → improve corpus independently → future benchmark measures whether performance changed ``

This protects the benchmark from answer contamination while still allowing evaluation failures to guide future knowledge acquisition.

Publication and promotion safety boundaries

Two safety boundaries were recorded as hard-disabled:

Automatic public research publication

HARD OFF. Public research remains human-reviewed.

Blind Trusted Knowledge promotion

HARD OFF. Autonomous learning does not impose a minimum card-production quota and may complete successfully with no promoted knowledge when evidence is insufficient.

These controls intentionally separate:

``text research generation != evidence sufficiency != Trusted Knowledge promotion != public research publication ``

Readiness audit checkpoint

The production lock-in readiness audit recorded:

failures: 0
warnings: 1
result: PASS WITH WARNINGS

The remaining warning was that an external S3-compatible artifact-storage provider was not configured. The source identifies this as non-blocking because local encrypted artifact storage remained active.

The audit covered major areas around autonomous-learning schema/policy, curriculum, verification-age policy, benchmark-gap feeding, last learning run, lifecycle scripts, publication safety, promotion safety, scheduling, worker cron, and the cron daemon, alongside broader application readiness checks.

Pause and audit model

The source defines an important operational principle: autonomous systems must remain stoppable and auditable.

Some pause/audit controls are stated as requirements or next-stage expectations, not all as completed production features. These include:

a global autonomous-learning pause,
independent provider disablement,
per-topic curriculum enable/disable,
promotion freeze independent of research generation,
emergency scheduler shutdown without deleting queued work,
run-level accounting for selected topics, requests, evidence, claims, verification, conflicts, promotions/rejections, provider usage, tokens, cost, latency, retries, and final state.

These requirements are preserved here because they define the safety boundary for expanding autonomy, but they are not upgraded to verified implementation claims where the source only specifies future/expected behavior.

Current verified state

Recorded operational at this checkpoint:

autonomous-learning schema/persistence foundation,
24 enabled curriculum domains,
zero-user research opportunity generation,
daily systemd scheduling installed/enabled,
approximately five-minute lifecycle worker scheduling,
first autonomous daily run completed,
four research requests queued by that run,
benchmark-gap research feeder present,
30-day maximum verification-age safeguard,
automatic public research publication hard-disabled,
blind Trusted Knowledge promotion hard-disabled,
lock-in readiness audit completed with zero failures and one non-blocking warning.

Limitations and non-claims

This release does not claim:

that every autonomous research request completes the full downstream lifecycle,
that all queued research produces new knowledge,
that card-count growth is a quality metric,
that the bulk curriculum importer is implemented,
that the full Learning Operations dashboard is implemented,
that all proposed pause/kill-switch UI controls are already deployed,
that autonomous promotion volume has been validated at large scale,
that general learning from ordinary user conversations is safe for a shared corpus,
that large-corpus retrieval or concurrency limits have been established,
that external-provider use is automatically beneficial.

The source explicitly recommends controlled expansion: freeze a baseline, add workload, remeasure quality/cost/compute, and continue only where measurable benefit justifies the added complexity.

Engineering principle

The recorded direction is not to maximize the number of knowledge cards. The system should maximize useful, retrievable, current, provenance-backed intelligence per unit of compute, cost, and complexity.

The meaningful checkpoint established here is that KESPA can now initiate useful research work without a user actively teaching it, while retaining evidence, promotion, publication, and future audit boundaries around what that autonomous work is allowed to become.

TEXT

SHA256SUMS.txt

393 bytes SHA-256 98e6694d5e5a3aa9…

99f201a621d68f8f41aca3d9c2d057bb5b627493d58017bb35dce9b9707ad040  README.md
2979a90fe8fce616c2a230b00fda6f2933407876a27b6e3329dee2f0938d7c17  note.json
672fba8ff428d48136d8c52561c236b34559499d11ef8fe11a0206add63b3804  metrics.csv
fccd8593120a3cc2886153a38a4bb1adcea0f5e2115350dd80fc5faa406acd4b  methodology.md
9125f795b5576513d8fcb6580bc4f43b83743653fb5219ae5f0b04ee32514457  provenance.json

MARKDOWN

methodology.md

3,427 bytes SHA-256 fccd8593120a3cc2…

Methodology — KESPA-NOTE-029

Source basis

This technical note was prepared from the supplied September 24, 2026 KESPA engineering/Jira update describing the V2.3 lock-in release, autonomous-learning flywheel, scheduler/worker integration, evidence-gating rules, readiness audit, pause/audit expectations, and future growth gates.

No external source was used to upgrade implementation status or fill missing measurements.

Inclusion rule

This release includes only facts that materially establish the zero-user curriculum-driven learning checkpoint:

autonomous-learning persistence and curriculum,
operating-system scheduling,
asynchronous worker cadence,
first recorded production run,
research-request queue initiation,
verification-age safeguard,
benchmark-gap feeder boundary,
publication and blind-promotion hard-off states,
readiness-audit outcome,
explicit limitations around downstream completion and controlled expansion.

Deduplication boundary

This note is intentionally narrower than earlier KESPA public research artifacts.

It does not republish the earlier natural research-gap lifecycle as new work. Earlier work covered knowledge-gap detection during normal use, evidence acquisition/qualification, cross-host research handoff, Trusted Knowledge lifecycle mechanics, and freshness/reverification foundations.

The distinct contribution recorded here is scheduled proactive research initiation with no active user, an explicit curriculum, OS-level recurring execution, benchmark-gap inputs, and operational safety boundaries around what autonomous work may become.

Verification interpretation

The note is marked verified because the source records concrete production validation of the subsystem's initiation path:

daily-learning timer installed and enabled,
first zero-user run completed,
four research requests queued,
last-run state confirmed by readiness audit,
worker scheduling installed,
hard-off publication/promotion safety state,
readiness audit with zero failures.

The verified status is deliberately scoped. It does not mean every downstream research request was shown to complete source acquisition, claim extraction, verification, conflict handling, promotion, indexing, and future reverification.

Planned-vs-operational handling

Statements phrased in the source as should, future, next, expected, or planned are retained as design requirements or limitations rather than converted into completed implementation claims.

Examples include the bulk curriculum importer, dedicated Learning Operations dashboard, some multi-level pause controls, large-corpus scaling work, general-learning privacy sanitization, and formal growth-checkpoint automation.

Measurement integrity

No benchmark score was created for this technical note. Counts and timing values are retained only where directly recorded by the source. No inference was made from the four queued research requests to knowledge-card count, research success rate, verification rate, or downstream promotion rate.

Publication privacy

The public package intentionally excludes credentials, private user data, private prompts/answers, exact private storage locations, raw internal logs/errors, authentication material, and other implementation-sensitive details that are not necessary to document the research architecture and verified checkpoint.

CSV

metrics.csv

2,005 bytes SHA-256 672fba8ff428d481…

metric,value,unit,status,source_type,notes
curriculum_enabled_domains,24,domains,recorded,direct,"Enabled research domains at the recorded checkpoint."
daily_learning_timer_installed,true,boolean,recorded,direct,"Operating-system daily-learning timer installed."
daily_learning_timer_enabled,true,boolean,recorded,direct,"Daily-learning timer enabled."
daily_learning_default_time,02:00,approx_local_time,recorded,direct,"Source describes the default cycle as approximately 02:00."
background_worker_interval,5,minutes,recorded,direct,"Lifecycle workers scheduled on a five-minute cadence."
first_daily_learning_status,complete,state,recorded,direct,"First recorded zero-user learning run completed."
first_run_public_research_requests_queued,4,requests,recorded,direct,"Queued public research requests from the first production run."
maximum_verification_age_ceiling,30,days,recorded,direct,"Additional maximum age safeguard; semantic freshness classes remain differentiated."
automatic_public_research_publication,false,boolean,recorded,direct,"Hard off; public research remains human-reviewed."
blind_trusted_knowledge_promotion,false,boolean,recorded,direct,"Hard off; research quotas do not imply card-production quotas."
readiness_audit_failures,0,count,recorded,direct,"Production lock-in readiness audit."
readiness_audit_warnings,1,count,recorded,direct,"Single non-blocking warning: external S3-compatible storage not configured."
readiness_audit_result,PASS WITH WARNINGS,state,recorded,direct,"Local encrypted artifact storage remained active."
full_downstream_lifecycle_universally_verified,false,boolean,boundary,direct,"Source explicitly states the first run does not prove every queued request completes every downstream stage."
bulk_curriculum_importer_implemented,false,boolean,boundary,direct,"Bulk curriculum ingestion is described as planned next work."
learning_operations_dashboard_implemented,false,boolean,boundary,direct,"Dedicated Learning Operations dashboard is future work."

JSON

note.json

1,919 bytes SHA-256 2979a90fe8fce616…

{
    "id": "KESPA-NOTE-029",
    "type": "technical_note",
    "title": "Zero-User Curriculum Learning and Evidence-Gated Autonomous Research",
    "date": "2026-09-24",
    "status": "verified",
    "summary": "KESPA operationalized a scheduled zero-user learning driver that selects research opportunities from an explicit curriculum and other gap signals, queues asynchronous research work, and keeps Trusted Knowledge promotion and public publication independently fail-closed. The first recorded production daily-learning run completed and queued four public research requests; 24 curriculum domains were enabled, a 30-day maximum verification-age safeguard was active, background lifecycle workers were scheduled, and the production readiness audit reported zero failures with one non-blocking storage warning. Full downstream completion of every queued research request and large-scale autonomous promotion remain outside this verified checkpoint.",
    "curriculum": {
        "enabled_domains": 24,
        "bulk_importer_implemented": false
    },
    "scheduling": {
        "daily_learning_timer_installed": true,
        "daily_learning_timer_enabled": true,
        "default_daily_time_approx": "02:00",
        "worker_interval_minutes": 5
    },
    "first_recorded_run": {
        "status": "complete",
        "queued_public_research_requests": 4
    },
    "safety": {
        "automatic_publication": false,
        "blind_trusted_knowledge_promotion": false,
        "maximum_verification_age_days": 30
    },
    "readiness_audit": {
        "failures": 0,
        "warnings": 1,
        "result": "PASS WITH WARNINGS",
        "warning": "External S3-compatible artifact storage not configured; local encrypted storage remained active."
    },
    "full_downstream_lifecycle_universally_verified": false,
    "tags": [
        "autonomous-learning",
        "zero-user-learning",
        "curriculum",
        "evidence-gating",
        "scheduled-research",
        "knowledge-lifecycle",
        "auditability"
    ]
}

JSON

provenance.json

3,064 bytes SHA-256 9125f795b5576513…

{
    "schema": "kespa.public_research_provenance.v1",
    "release_id": "KESPA-NOTE-029",
    "source_material": [
        {
            "kind": "engineering_update",
            "description": "KESPA V2.3 lock-in / autonomous-learning engineering and Jira update supplied by the project owner on 2026-09-24."
        }
    ],
    "directly_recorded_facts": [
        "The autonomous-learning subsystem can initiate research work without an active user.",
        "The curriculum contained 24 enabled research domains at the recorded checkpoint.",
        "A daily operating-system timer was installed and enabled, with a default cycle described as approximately 02:00.",
        "Background lifecycle workers were scheduled on an approximately five-minute cadence.",
        "The first zero-user daily-learning run completed and queued four public research requests.",
        "A maximum verification-age ceiling of 30 days was active as an additional safeguard.",
        "Benchmark gaps may feed research targets without exporting benchmark prompt/gold content into the research pipeline.",
        "Automatic public research publication was hard-disabled.",
        "Blind Trusted Knowledge promotion was hard-disabled.",
        "The lock-in readiness audit reported zero failures, one warning, and PASS WITH WARNINGS.",
        "The remaining audit warning was that external S3-compatible artifact storage was not configured while local encrypted storage remained active."
    ],
    "derived_values": [],
    "validation_boundary": [
        "The source states that the first production run demonstrates scheduler, persistence, and queue initiation, but does not prove every research request completes every downstream lifecycle stage.",
        "The source does not provide a verified knowledge-card promotion count for the first run.",
        "The source does not provide a research completion rate, verification pass rate, conflict rate, or provider-cost total for the first run.",
        "Bulk curriculum import, the dedicated Learning Operations dashboard, large-scale expansion, and general chat-derived learning sanitization remain future work.",
        "Some pause and audit controls are expressed as operational requirements/expectations rather than all being explicitly verified as deployed UI features."
    ],
    "deduplication": {
        "related_prior_records": [
            "KESPA-NOTE-003",
            "KESPA-NOTE-010",
            "KESPA-NOTE-024"
        ],
        "distinct_scope": "This release records zero-user curriculum-driven research initiation, recurring OS scheduling, benchmark-gap feeding, and evidence-gated safety boundaries rather than republishing the earlier user-triggered natural research-gap or Trusted Knowledge lifecycle work."
    },
    "publication_exclusions": [
        "private prompts and answers",
        "private user or organization data",
        "credentials, API keys, tokens, cookies, or sessions",
        "raw internal logs and error bodies",
        "unnecessary private filesystem/storage paths",
        "unsupported claims that queued research became Trusted Knowledge",
        "unsupported claims of large-scale autonomous-learning performance"
    ]
}