OPEN SCIENCE · RELEASE 2026.10.29

Knowledge should be free to read.

Research notes, methods, measured results and reproducibility artifacts from Darkoza Knowledge Lab—available without a paywall. Scientific claims are labeled by the strength and limits of their evidence.

Explore research Open JSON catalog ↗

Release scope: this first release publishes reproducible technical studies and preliminary evaluations. It does not claim a newly discovered scientific law or an independently confirmed breakthrough.

10research reports

100%free to read

JSONpublic data catalog

0private corpora published

RESEARCH REGISTER

Evidence before claims.

Each report includes its method, raw outputs and known limitations. “Reproduced” means a defined check passed—not that the result is universally generalizable.

PERFORMANCE · REPRODUCIBLE PILOT

Prime-sieve scaling across four workload sizes

Compared segment sizes and worker counts for a segmented prime sieve at limits from 1 million to 50 million. All returned counts matched known reference values; two workers with 16 KiB segments were fastest in the tested configurations.

1.87×observed speedup at 50M

4 / 4tested limits matched reference counts

3repetitions per configuration

Host-specific performance evidence. Not a claim about model inference, retrieval, or other hardware.

Read report ↗CSV ↗JSON ↗Source ↗

INFORMATION RETRIEVAL · SYNTHETIC BENCHMARK

Compact arrays versus set-based index postings

On a synthetic 20,000-document fixture, sorted compact arrays used substantially less Python-traced allocation and returned queries faster than set postings while preserving the same tested rankings.

−88%traced peak allocation

−69%median query latency

Samerankings on 8 fixture queries

Synthetic queries validate implementation mechanics, not real-world search relevance. Tracemalloc is not process RSS.

Read report ↗CSV ↗JSON ↗Source ↗

INFORMATION RETRIEVAL · PRELIMINARY EVALUATION

BM25-style ranking versus TF–IDF on technical documentation

Evaluated 21 queries over 52 installed public system-documentation files. Against researcher-curated source labels, BM25 achieved higher aggregate ranking metrics; TF–IDF was faster in this small local pilot.

0.675BM25 nDCG@5

0.434TF–IDF nDCG@5

42 / 42top-five orders independently reproduced

The labels were curator-authored, not independently graded by human reviewers. The automated gate verifies code-path agreement and arithmetic—not semantic truth or production readiness.

Read report ↗CSV ↗JSON ↗Verification gate ↗Held-out audit ↗

AGENT CONTROL · SYNTHETIC ENGINEERING PILOT

Context-aware procedures with evidence-gated completion

A deterministic pilot compared context-agnostic transition selection with a context-aware procedural graph. The candidate scored higher on held-out synthetic instances, abstained on an unknown context, and rejected unsupported completion claims.

100.0%context-aware accuracy

83.3%context-agnostic baseline

4 / 4completion gate checks passed

Controlled synthetic fixtures only. Held-out records replicate known path structures; real-world generalization and production benefit remain unproven. No model weights or production graph were changed.

Read report ↗Markdown ↗Raw JSON ↗Source ↗

OPERATIONAL RELIABILITY · PRIVACY-MINIMIZED AUDIT

Operational autonomy: verified-progress and source-isolation audit

Inspected 669 real autonomy-cycle outcomes. No cycle recorded verified completion; 401 zero-exit cycles lacked the complete evidence bundle. The audit identified branch-contract, source-isolation and upstream-authentication blockers without publishing raw traces or credentials.

669cycle records inspected

0verified completions

401zero-exit cycles without complete evidence

Operational reliability evidence only—not a model-capability claim. Production code, model weights and graph state were not changed during this audit.

Read report ↗Markdown ↗Aggregate JSON ↗

OERSI RELIABILITY · REGRESSION-TESTED PATCH CANDIDATE

Reliable verifier cleanup under failure

A tested patch separates cleanup from the fail-fast verification list, so cleanup still runs after a failed or timed-out check. The failure remains a failure; cleanup does not weaken the safety gate.

65Python unit tests passed

2Node test suites passed

3runner, tests and example config files changed

Patch candidate only; not deployed. Source synchronization and service-account permissions remain unresolved.

Read report ↗Download patch ↗JSON result ↗

PRODUCTION RELIABILITY · BACKUP POLICY

Encrypted backup cadence aligned with policy

The live timer was found at a three-day interval and corrected to ten days. The last full backup passed remote upload and ciphertext hash verification; the next scheduled run is 7 November 2026 at 08:25:17 UTC.

3d → 10deffective timer interval

1live timer corrected

0production app/model changes

The future backup has not yet run, and no restore drill was performed. The publication verifies the schedule, not end-to-end restore readiness.

Read report ↗Audit JSON ↗API record ↗

PRODUCTION RELIABILITY · MEMORY PRESSURE

Bounded swap headroom for local model resilience

After kernel records showed OOM kills with no active host swap, a persistent 4 GiB swap area was activated and the affected model container gained up to 4 GiB swap headroom without raising its 4 GiB RAM cap.

4 GiBhost swap active

10swappiness

4 + 4 GiBcontainer RAM + swap maximum

This adds an emergency memory cushion, not physical RAM. Under sustained pressure, swapping can slow inference and OOM can still occur.

Read report ↗Audit JSON ↗API record ↗

LOCAL AI · MODEL ROUTING EXPERIMENT

Test a stronger local reasoning route

After 670 autonomy cycles without verified progress, OERSI was configured to try an installed Qwen3-based 8.2B local model with thinking capability and a 1,024-token output budget, up from 256.

7.6B → 8.2Blocal model size

4×output-token budget

0verification gates removed

The next cycle has not yet established whether this configuration improves verified progress. This is an experiment, not a claimed success.

Read report ↗Audit JSON ↗API record ↗

LONG-CONTEXT SYSTEMS · REPRODUCIBLE SYNTHETIC PILOT

Effective context capacity via external memory

A standalone OERSI prototype stores original records in SQLite, retrieves source-linked chunks with BM25-style lexical ranking, and packs them within an estimated context budget. Two deterministic runs reproduced the same synthetic retrieval and hash-integrity outcomes.

200 / 200target facts retrieved and packed

3.0%recency-only baseline hits

8 / 8unit tests passed

The queries contain target identifiers; this tests retrieval mechanics only, not semantic reasoning or answer quality. The token budget is approximate, and the prototype is not connected to live OERSI inference.

Read report ↗Markdown ↗Run 1 JSON ↗Run 2 JSON ↗Source ↗Benchmark ↗

MACHINE-READABLE ACCESS

Free research API.

Public catalog and experiment summaries are available over HTTPS without an API key. This first version is read-only and static by design. API keys, when introduced for optional metering or customer tooling, will be free to obtain and activate; the public endpoints do not require one.

LIVE · PUBLIC READ-ONLY

Version 1 catalog

Returns the published studies, evidence status, limitations and artifact URLs.

GET /science-api/v1/catalog.json

View JSON ↗

curl -fsS https://darkoza.com/science-api/v1/catalog.json

SUSTAINABLE RESEARCH

Free knowledge. Paid capability.

Our intended model keeps public research, findings and methods free to read. Paid products may include higher-volume API access, batch retrieval, richer provenance exports and compute-intensive research workflows.

FREE

Public knowledge

Reports, methods, published datasets and the read-only catalog.

Available now

COMMERCIAL API

Higher-volume services

Obtaining or activating an API key will cost €0. Any future fee would apply only to clearly disclosed premium usage or compute—not to the key itself. Metering, billing, customer protections and end-to-end checks are required before premium access is offered.

Not yet on sale

No payment is collected and no premium API is represented as active in this release. Public endpoints are free and keyless. API keys themselves are intended to remain free; premium usage, if introduced, must be separated transparently from key issuance and launched only after authorization, billing, metering, abuse controls and a working customer flow are verified.

PUBLICATION POLICY

Transparent by default.

Evidence labels

Hypotheses, computational observations, reproduced checks and independently confirmed findings are different categories. We do not infer novelty from a finite search alone.

Reproducibility

Where safe and appropriate, reports link to code, parameters and raw outputs. A passing test verifies only the test's stated scope.

Free API keys

No fee to obtain or activate a key. The public research catalog currently requires no key at all; future keyed access must keep key issuance free and disclose any separate premium usage limits or charges.

Privacy and rights

This release excludes private user/project corpora and secrets. Public access does not itself grant a blanket reuse license; any applicable file-level license controls.

اصل انتشار: دانش و نتایج پژوهشیِ قابل انتشار رایگان در دسترس عموم قرار می‌گیرند؛ خدمات API، پردازش و قابلیت‌های تخصصی می‌توانند منبع درآمد باشند. هیچ فرضیه یا نتیجهٔ مقدماتی به‌عنوان کشف علمیِ اثبات‌شده معرفی نمی‌شود.