genefold-ai/cyberzen-rag-release

CyberZen RAG: follow the evidence behind an AI security assessment

Genefold AI is exploring how better retrieval can support better-informed security decisions. CyberZen RAG is our research prototype for following an investigation from an operational problem to the evidence an AI analyst sees, the questions it asks and the assessment it produces.

For a security-product team, finding a vulnerability record is the beginning. The useful question is whether it applies to a particular version, deployment and set of symptoms—and what to check next. For an enterprise AI team, that becomes a retrieval question: are we giving the model the evidence it needs to make those distinctions?

CyberZen brings together Genefold AI's ArrowSpace retrieval technology and OrcaCyber Zero 1.5. We vary how evidence is found, keep the investigation within clear resource limits, and publish the results in enough detail for others to examine. This release is the developing evidence behind that work.

Start with an investigation: a document-transfer service with unexplained gateway crashes. You can read every turn directly in your browser.

Latest experiment · Read all investigations · Discuss an evaluation with Genefold AI

What the experiments tell us

Our latest experiment explored a simple idea: start with familiar similarity search, then gradually give more weight to relationships captured by the retrieval graph. Could that change preserve more of a fixed search reference while keeping the investigation affordable?

What we measured	Latest recorded result	Why a product team would look at it
Preservation of a fixed search reference, measured by Tolerant Recall	Candidate score 93.33% → 95.38%; displayed-evidence score 80.95% → 82.19% in the first matched comparison	Helps compare retrieval choices against the same reference
Investigation delivery	131 completed investigations, using 867 model requests	Shows how the bounded workflow behaved across the recorded scenes
Estimated model usage cost	$20.008560 across that campaign, using reported tokens and cache-aware rates	Makes resource trade-offs visible alongside retrieval measurements
Inspectable execution	312 full readable lifecycles across three campaigns	Lets peers examine successes, failures, follow-up questions and final answers

The first proposed search schedule was retained; the second did not meet the improvement rule. The operator stopped the third comparison after 11 of its 40 planned investigations. Fresh independent verification has not run, so the retained result is a development finding. Tolerant Recall measures preservation of a search reference; answer correctness is a separate question for review. The campaign report explains the comparisons and accounting in full.

How CyberZen works

A Common Vulnerabilities and Exposures (CVE™) record describes a publicly identified vulnerability. CyberZen searches a corpus of 408,469 CVE records and gives selected evidence to an AI analyst. This is retrieval-augmented generation, or RAG: the model receives source material as part of the task.

The analyst can ask for another search or deliver an assessment. Each follow-up adds evidence to the conversation, while the application limits requests, time and output size.

Describe a real-world situation
    → Find potentially relevant vulnerability records
    → Show the selected evidence to the analyst
    → Ask a more focused question, or give an assessment
    → Keep the full exchange, outcome and resource use for review

The latest campaign retrieved 30 candidates and showed 15 records per search. Earlier experiments showed ten, allowing us to study different evidence budgets in separately documented campaigns. The implementation is a bounded Rust application built with Rig; ArrowSpace supplies the retrieval layer. Engineering details and commands are in USAGE.md, with current experiment settings in the latest methodology.

Scenarios and Scenes: full cycle

The easiest way to understand CyberZen is to follow an investigation. A scenario says what the analyst should do, such as triage a possible vulnerability. A scene describes the situation: the software in use, the symptoms, the connections between services and the questions still unanswered. Each independent investigation is called a cell in the experiment tables.

Follow a document-transfer investigation

Imagine a cloud service exchanging documents with business partners. Its secure gateway starts crashing after certificate changes, and deliveries begin to accumulate. The operations team needs to understand the cause while keeping documents available and avoiding unnecessary disruption.

That is the starting point of this full readable investigation from campaign 1558. Open it alongside this guide:

1. Meet the situation. The scene describes an OpenSSL gateway, a web proxy, Java workers and a delivery queue. Some versions and settings are known; others still need checking.
2. See what the analyst receives. In turn one, you can read the search results, the system instructions and the complete scene-and-evidence message. The candidate table distinguishes retrieved records from those selected for display.
3. Follow the next question. The first response asks for a more focused search on OpenSSL certificate handling. Turn two shows the resulting evidence. Later turns retain earlier evidence while the analyst refines its query and the search weighting changes.
4. Read the conclusion. In turn seven, the analyst concludes that the retrieved material does not establish the cause. Its final assessment identifies missing evidence and proposes checks on the certificate chain, software configuration and delivery logs.
5. Put the work in context. This cell took about 60 seconds, made seven model requests and had an estimated cache-aware model cost of $0.139273. The run report brings its settings, search measurements, cost and interpretation together.

This example shows why publishing the full exchange matters. You can see repeated searches, evidence that does not resolve the question and the uncertainty carried into the answer. The analyst's description of a “corpus gap” is its interpretation of those search results; the trace establishes what was retrieved, rather than whether an advisory exists anywhere in the corpus. It gives a product team concrete behaviour to review and improve.

Read every turn the same way

Every readable file follows one template:

Scene → retrieval → system message → user message → model response → usage and timing → final assessment.

The system message sets the analyst's instructions. The user message contains the scene and evidence assembled by the application; later user messages incorporate additional evidence. The model response shows the follow-up search or answer actually returned. These labels let you follow the recorded exchange without needing to interpret log formats.

Find the record behind each question

Your question	Where to read it	Underlying record
What scene and settings were used?	Scene and investigation details	protocol.json
Exactly what prompt was passed?	System and user messages in turn one, then the same sections in later turns	JSONL llm_start
What response was received?	The Model response section in each turn	JSONL llm_request_response
How did the investigation unfold?	Read the complete seven-turn sequence	Ordered search, evidence and model events
What was the conclusion?	Final assessment and outcome	comparison.json and JSONL cell_end
What resources did the scene consume?	Timing, tokens and cost in the run report	Per-turn usage and campaign-cell.json

Choose another situation

• Clinical remote access: staff sessions interrupted by VPN restarts.
• Cloud delivery: a certificate-processing question within a connected service.
• Browse all 131 latest-campaign investigations, or open the readable-file directory.
• Explore the earlier 80 scheduled-search investigations and 101 historical investigations.

The reading guide explains the template. Each transcript also links to its original JSONL for readers who want the exact recorded bytes and all instrumentation fields.

Earlier OpenSSL example: worked walkthrough and technical reconstruction

For campaign 0000, you can inspect:

• Human-readable investigation reports
• Step-by-step action and usage table
• Original private-source provenance references and checksums

Worked example: OpenSSL document-transfer gateway

Follow cell r0-p0-granite-30m-1, the first primary cell in run 00: OpenSSL / Granite / stratified (machine-readable report). It is a short, completed two-search, two-request example selected to explain the record format. It is a historical 0.3.0 / analyst-v1 observation, with fixed tau 1.00, not a scheduled-tau result or a representative-quality claim.

Access: the report, walkthrough, full readable lifecycle, byte-identical JSONL and protocol are public. Their source bytes were verified against private revision bf77f9db9eae16ef6a7ba12c7b99d8b1d582c2ae; the transcript manifest retains that provenance. Line numbers below refer to physical JSONL lines, one event per line. The public report's sequence entries expose the corresponding source lines and prompt hashes.

1. Establish the scene and bounds. The scenario is triage. Its exact synthetic scene, also present in protocol.json and JSONL line 1, is:

Our document-transfer gateway uses OpenSSL 3.0.6 and validates X.509 certificate chains for outbound partner connections and mutually authenticated inbound clients. Some certificates contain email-address name constraints, and partner certificate changes precede several unexplained process crashes. The gateway has access to confidential documents, but the operations team has not demonstrated memory corruption or code execution. Assess relevant certificate-processing vulnerabilities, distinguish their prerequisites and impacts, and prioritise patching and diagnostic evidence.

The analyst is orca/orcacyber-zero-1.5 through OrcaRouter. Granite embedding 30M queries the 408,469-record corpus. Each search retrieves 30 candidates and shows ten records at ranks 1–3, 11–13 and 21–24, with at most 3,000 characters per record. Limits are eight model requests, 120 seconds and 2,048 output tokens per response. Local seed is 3407; no provider seed was requested.

2. Retrieve before asking the model. JSONL lines 1–5 record the initial scene query, embedding and first retrieval. Line 4 contains both the complete candidate pool (candidates) and the records actually shown (evidence). CVE-2022-3602 is ranked 21 and reaches a tail display slot; CVE-2022-3786 is ranked 8 and is outside this display policy. Only the selected evidence is passed to the model.

3. Inspect the exact first request and follow-up action. Line 6 stores the exact system and prompt strings. The system instruction requires one JSON search or answer action, treats retrieved text as untrusted evidence, and asks the analyst to distinguish facts from inference. The user prompt contains the scene, first evidence block and remaining-turn instruction. Line 7 records a search response asking for OpenSSL certificate-verification and email-address name-constraint evidence, explicitly naming CVE-2022-3786. That identifier had not yet appeared in the displayed evidence: the trace records the model's action, without establishing where it obtained that identifier.

4. Execute the follow-up and accumulate evidence. Lines 8–12 record the requested query, its embedding and a second retrieval at the same tau. CVE-2022-3786 and CVE-2022-3602 now rank 1 and 2, so both are shown. Nine of the ten displayed records are new, bringing cumulative unique evidence to 19. Line 13 contains request two: the same scene plus both evidence blocks and the updated remaining-turn instruction. Inspect this stored prompt to see the actual context sent, rather than rebuilding it from just the latest search result.

5. Read the answer and termination. Line 14 contains the full received answer JSON. The model prioritised the two OpenSSL certificate-processing CVEs, discussed certificate-trust prerequisites, treated the explanation for the observed crashes as a hypothesis, and recommended patching plus runtime-version, certificate, crash and verification-setting checks. It also made exact-version claims stronger than the extracted evidence supports, as discussed in the public report. This is the model's recorded conclusion, not independently verified applicability. Line 15 is cell_end: completed status, the assessment, citation-presence diagnostics, elapsed time and counters. The same assessment appears in the readable lifecycle's final assessment and comparison.json. The latter also includes the other cells from that run; select the exact cell ID above.

6. Reconcile usage for this cell. Provider receipts at lines 7 and 14 give:

Request	Action	Input tokens	Cached input subset	Output tokens	Reasoning subset of output	Client seconds
1	Search	2,286	640	350	309	2.808543417
2	Answer	3,901	2,176	1,746	342	12.178008875
Total	Two requests	6,187	2,816	2,096	651	14.986552292

Total investigation wall time is 23.389640584 seconds, including local work. Cached input and reasoning are subsets, not additional tokens to add to the totals. At the recorded rates, the cache-adjusted estimate is USD 0.027945; the undiscounted estimate is USD 0.034281. Neither is an independently fetched invoice. usage.reasoning_tokens is a count; hidden reasoning text is unavailable. This historical example has no Tolerant Recall receipt.

Reconstruct the recorded exchange

After downloading the public JSONL linked above, save it as r0-p0-granite-30m-1.jsonl. Its SHA-256 must be dd35183aeba0b22a3a697cc4c3bf1f4ddbc6bdf93906499aba42124d249f7bb1. The following commands inspect the recorded exchange without making model calls:

shasum -a 256 r0-p0-granite-30m-1.jsonl

# Physical line numbers, event types and timestamps: the investigation timeline.
jq -r '[input_line_number, .type, .unix_millis] | @tsv' r0-p0-granite-30m-1.jsonl

# Exact system and user-prompt strings for both calls, in recorded order.
jq -r 'select(.type == "llm_start") | "REQUEST \(.request)\nSYSTEM:\n\(.system)\nPROMPT:\n\(.prompt)\n"' r0-p0-granite-30m-1.jsonl

# Both received responses with provider-reported usage and request identity.
jq 'select(.type == "llm_request_response") | {response, usage, finish_reason, provider_request_id, seconds}' r0-p0-granite-30m-1.jsonl

# Candidate versus displayed evidence for each retrieval, then the final result.
jq 'select(.type == "search") | {tau, evidence_policy, candidates, evidence}' r0-p0-granite-30m-1.jsonl
jq 'select(.type == "cell_end") | .result' r0-p0-granite-30m-1.jsonl

Pair each llm_start with the following llm_request_response in this cell's event order. Keep the original JSON strings and timestamps when auditing. Re-executing the retrieval/model pipeline additionally requires the pinned corpus, encoder, index and provider access described in the methodology.

What this could mean for your product

For Genefold AI, the opportunity is to make evidence discovery a component that security teams can evaluate and improve. CyberZen gives us a concrete setting in which to study that opportunity: a question, a controlled retrieval process, an analyst's response and a visible cost.

• Vulnerability-management teams can examine how retrieved records connect to deployment details and where an assessment needs more evidence.
• Analyst-workbench teams can review the search trail alongside the final answer, making unsupported conclusions and unproductive follow-ups visible.
• Enterprise RAG teams can compare retrieval choices under matched evidence and resource budgets before deciding what deserves further integration work.

The experiments already offer useful engineering lessons. In the historical Granite searches, spectral weighting moved an OpenSSL reference record from rank 21 to 15 and a FortiOS record from 27 to 23. The newer scheduled-search experiment improved preservation of a fixed reference. The full transcripts show the next challenge just as clearly: turning available evidence into well-supported conclusions.

Explore a use case with Genefold AI

We welcome conversations with security vendors, enterprise AI teams and research partners who want to evaluate this approach on representative deployment scenes. A useful collaboration starts with a concrete question: what evidence would help your analysts, how should a good answer be judged, and what time or cost budget must it meet?

Together, those requirements can shape a focused evaluation with explicit comparisons and expert review. CyberZen is a research prototype; this dataset lets prospective partners examine the work and help define the next test.

Contact business@genefold.ai or join the public dataset discussion.

Why the way we search matters

A model can only use retrieved evidence if that evidence reaches its context. ArrowSpace lets us vary the balance between familiar embedding similarity and relationships represented by a graph. CyberZen tests how those choices change the material available during an investigation, then records what the analyst does with it.

The papers below explain the retrieval foundations. The campaign reports and full transcripts show how we are testing them in practice.

Explore the research foundations and search-tau method

Semantic search as a reasoning enabler

The original ArrowSpace paper describes the retrieval capability as follows:

“This enables similarity search that considers both semantic content and spectral characteristics of high-dimensional vector datasets.”

— Lorenzo Moriondo, ArrowSpace: introducing Spectral Indexing for vector search, Journal of Open Source Software 10(113), 9002 (2025), Summary, p. 1. DOI: 10.21105/joss.09002 (paper PDF, CC BY 4.0).

ArrowSpace augments embedding similarity with graph-derived spectral signatures. Its index uses Laplacian smoothness and dispersion information to give retrieval an additional view of relationships within the embedding dataset.

The later Graph Wiring and Spectral Indexing study connects retrieval to context formation for RAG and agentic workflows, noting that:

“structurally inconsistent or noisy retrieved context can negatively affect downstream reasoning and generation quality.”

— Lorenzo Moriondo and Ilias Azizi, From Embedding Geometry to Spectral Search: Energy Dispersion Networks for Vector Retrieval, arXiv:2606.21535v2 (13 July 2026), Section 6, “Applications”. Paper and version history (quoted section, CC BY 4.0).

That section proposes using more geometric retrieval while forming a broad context, then increasing the spectral contribution for more specific searches; experimental testing of this adaptive loop is future work in the paper. For CyberZen, this motivates testing whether changing search tau exposes a missing version constraint or a contradictory CVE™ record before OrcaCyber revises an assessment.

In CyberZen, reasoning enabler describes this role in evidence selection: changing the retrieved context can expose a prerequisite, an alternative explanation or a contradictory record that gives the analyst a reason to revise its assessment. The research hypothesis is that controllable evidence discovery can improve the subsequent sequence of questions, comparisons and decisions. These papers supply the retrieval foundations; CyberZen evaluates that downstream hypothesis.

The useful chain to inspect is search → evidence → revised question or supported conclusion. Spectral similarity is not a proof of vulnerability applicability. An extra CVE™ or a higher reasoning-token count is not, by itself, evidence of better reasoning. Our reports therefore inspect observable search sequences, delivered answers and checked applicability errors separately.

Search tau and modulation

Search tau controls the balance between cosine similarity and spectral proximity. For a query and a candidate document, the ranking blend is:

$$S_{\tau }(q,d)=\tau \mathrm{cosine}(q,d)+(1-\tau )S_{\lambda }(q,d),0\le \tau \le 1.S\_\{\backslash tau\}(q,d)=\backslash tau\backslash ,\backslash operatorname\{cosine\}(q,d)\; +(1-\backslash tau)\backslash ,S\_\{\backslash lambda\}(q,d),\backslash qquad\; 0\backslash leq\backslash tau\backslash leq1.$$

Here the spectral term compares the query's and document's graph-derived signatures. CyberZen exposes the blend as --tau; the ArrowSpace Rust search API calls this weight alpha. It changes ranking while the indexed corpus, encoder and graph can remain fixed. It is not a model temperature or a probability that a CVE™ is relevant.

Search tau	Interpretation
1.00	Pure cosine ranking: the internal reference condition
0.72	72% cosine weight, 28% spectral-proximity weight
0.65	65% cosine weight, 35% spectral-proximity weight
0.42	42% cosine weight, 58% spectral-proximity weight

These percentages are score weights, not relevance or accuracy percentages. Lower tau changes which associations are favoured; it does not guarantee greater diversity, wider coverage or a more accurate answer.

Tau modulation means modulating search tau: changing this cosine/spectral retrieval weight across search steps. For example, a proposed controller could start with cosine, then increase the spectral contribution while investigating an unresolved evidence gap. The aim is to give the analyst complementary evidence under the same resource budget. A useful evaluation must compare identical query vectors as well as complete investigations, because a changed query can itself change the ranking.

Historical 0.3.0 status: that campaign held tau fixed within each investigation and compared the four settings above across independent cells. The later scheduled-tau campaigns report separate development observations. The proposed MiniLM comparison holds the scenes, corpus, model, controller and resource bounds constant, then evaluates fixed versus changing search tau. Their measured development gains and unexecuted fresh verification are documented in the campaign reports above.

Campaigns

Each campaign has a README explaining its objectives, principles, experimental design and interpretation, alongside its individual run reports.

Campaign	Question and status	Read
campaign-optimize-261016-0000 · 16 October 2026	Completed comparison of encoders, fixed search tau and head/stratified evidence, plus a development-only output-cap follow-up	Campaign README · 25 run reports
campaign-optimize-261016-0700 · 16 October 2026	0.5.1 top-15 development comparison; fresh verification pending	Campaign README
campaign-optimize-261016-1558 · 16 October 2026	0.5.2 Tolerant Recall restart; two full comparisons plus one operator-stopped partial comparison	Campaign README

There are three published campaign namespaces. The two scheduled-tau campaigns retain development results without fresh-verification claims; their observations are not pooled with the original 0.3.0 campaign.

Choose your next step

• Read an investigation: all 131 latest-campaign lifecycles.
• Evaluate the results: latest campaign findings and methodology.
• Explore the earlier work: first scheduled-search campaign and historical encoder/retrieval comparison.
• Discuss a representative use case: business@genefold.ai.

Historical tables, findings and all 25 original run reports

Start here

• Campaign objectives, principles and complete run index
• Campaign overview and discussion
• Methodology, metric definitions and access limitations
• 101 cell observations: TSV · JSON
• 25 run summaries: TSV · JSON
• 528 observable action summaries
• Synthetic scenes and partial anchor labels
• Frozen method/model/graph identities
• Aggregate metrics and scoped cost accounting
• Source receipt references · Derivation checks
• Inventory · SHA256 checksums · Publication manifest · Publication verification scope

Headline findings

The 24 primary runs contain 96 eligible cells over a 408,469-record CVE™ corpus: 61 valid answers, 20 partial answer-limit responses, 14 empty reasoning-only limit responses and one provider error. They made 515 attempted model requests. The excluded tau=0.55 pilot and separate development follow-up bring the total to 101 observed cells / 528 requests / 25 runs.

Granite's mean shown-anchor coverage was 95.83%, versus 1.04% for the pinned domain-adapted MiniLM configuration. Yet Granite completed only 26/48 cells versus MiniLM's 35/48; valid syntax is not evidence-grounded success. All 14 empty reasoning-only limit failures occurred with Granite. The primary 281,304 output tokens already include 169,681 reported reasoning tokens.

The user reports $8.680328 actual token billing, not independently fetched and without a precisely established account-versus-campaign scope. Cache-adjusted known usage is estimated at $7.971195 across all observed cells; conservative accounting is $16.789879, including prior reported spend and an unknown-request reservation. These figures have different meanings; see the overview rather than treating them as interchangeable bills.

A single development follow-up increased the output cap from 2,048 to 4,096: completion rose from 1/4 to 4/4, with estimated known cost rising from $0.124686 to $0.1419645. It is a research candidate, not a production promotion or independent semantic-quality verification.

Run reports

The campaign-scoped run index groups all 25 reports and their data alongside the campaign README. The links below use this canonical campaign namespace. Earlier root paths remain available at the historical immutable revision; the migration record records every original path, destination and checksum.

Each report contains the operational task, configuration and execution order, every cell's cost/tokens/reasoning/latency/anchor/completion metrics, concise search/model sequences, strengths, failure modes, reflection, public full-lifecycle links and original private-source provenance. Run 00 includes the excluded pilot. Source batch index 24 is the 25th run and is labelled 25-followup.

Run	Scene	Encoder	Policy	Split	Valid / eligible	Report
00	openssl-certificates	granite-30m	stratified	development	4/4	Read · JSON
01	openssl-certificates	granite-30m	head	development	3/4	Read · JSON
02	spring-war	domain-minilm	head	development	1/4	Read · JSON
03	spring-war	domain-minilm	stratified	development	1/4	Read · JSON
04	openssl-certificates	domain-minilm	stratified	development	3/4	Read · JSON
05	openssl-certificates	domain-minilm	head	development	3/4	Read · JSON
06	spring-war	granite-30m	head	development	3/4	Read · JSON
07	spring-war	granite-30m	stratified	development	3/4	Read · JSON
08	java-logging	domain-minilm	stratified	development	3/4	Read · JSON
09	java-logging	domain-minilm	head	development	3/4	Read · JSON
10	apache-portal	granite-30m	head	development	3/4	Read · JSON
11	apache-portal	granite-30m	stratified	development	1/4	Read · JSON
12	java-logging	granite-30m	head	development	2/4	Read · JSON
13	java-logging	granite-30m	stratified	development	2/4	Read · JSON
14	apache-portal	domain-minilm	head	development	4/4	Read · JSON
15	apache-portal	domain-minilm	stratified	development	3/4	Read · JSON
16	confluence-collaboration	domain-minilm	stratified	holdout	4/4	Read · JSON
17	confluence-collaboration	domain-minilm	head	holdout	4/4	Read · JSON
18	confluence-collaboration	granite-30m	head	holdout	1/4	Read · JSON
19	confluence-collaboration	granite-30m	stratified	holdout	3/4	Read · JSON
20	fortios-remote-access	granite-30m	head	holdout	0/4	Read · JSON
21	fortios-remote-access	granite-30m	stratified	holdout	1/4	Read · JSON
22	fortios-remote-access	domain-minilm	head	holdout	4/4	Read · JSON
23	fortios-remote-access	domain-minilm	stratified	holdout	2/4	Read · JSON
25-followup	apache-portal	granite-30m	stratified	development_followup	4/4	Read · JSON

Intended use, provenance and limitations

For public scrutiny, defensive research and outreach: compare retrieval, evidence selection, answer delivery and applicability failure modes. The six operational scenes are synthetic; this is not a real-incident dataset or a comprehensive CVE™ relevance benchmark. Qualitative review is coding-assistant review, not independent expert annotation. Tolerant Recall was not used in the historical 0.3.0 campaign; the separate 0.5.1 and 0.5.2 campaigns report calibrated retrieval preservation. Hidden reasoning text is unavailable, and reasoning counts do not prove coherent or consequential reasoning.

The original full-receipt dataset genefold-ai/cyberzen-rag remains private, pinned for historical source references at bf77f9db9eae16ef6a7ba12c7b99d8b1d582c2ae (source audit 420/420 files). The campaign transcript indexes above publish the recorded prompts, responses, evidence, usage and outcomes for direct inspection, with exact source hashes and public protocol/comparison context. The original private links remain provenance references. Re-executing the pipeline still requires the frozen corpus/index snapshots, the private domain-adapted model and provider access; those assets are not republished. The analyst provider alias has no recorded immutable model revision.

Authorised raw JSONL transcripts are included byte-for-byte, alongside readable Markdown; recorded local paths are retained as provenance. Full debug logs, the source-record corpus, weights and credentials are not included. CC BY 4.0 covers this authored curation and derived metrics, not rights to private or third-party source assets; retrieved CVE text remains subject to the attribution below.

Attribution: Genefold AI, CyberZen RAG operational campaign research release, 16 October 2026. Cite this dataset and a pinned release revision when reusing the results. Local seed: 3407. All output-limit failures and the unknown-usage request remain in the public metrics.

Data attribution and acknowledgements

CVE™ data is property of The MITRE Corporation (MITRE) and usage is in accordance with the CVE™ Terms of Use.

Credit for OrcaCyber Zero 1.5 has been provided by the Orca platform.

References and implementation status

• Original ArrowSpace paper: Moriondo, L. (2025). ArrowSpace: introducing Spectral Indexing for vector search. JOSS, 10(113), 9002. 10.21105/joss.09002.
• Extended research context: Moriondo, L. (14 October 2025). A New Pathway for Vector Databases: Spectral Indexing using taumode with arrowspace. Author-hosted paper, especially sections 2 and 6.4. Its earlier CVE™ experiments are separate from the CyberZen campaign reported here.
• Search blend: the public ArrowSpace implementation documents the cosine/spectral scoring operation. CyberZen forwards its search tau to this blend; the surrounding CyberZen application code is private.
• Version boundary: the historical campaign and its data tables used analyst-v1 / CyberZen 0.3.0. CyberZen 0.4.0 introduced bounded output-limit recovery and source-backed Tolerant Recall evaluation; 0.5.0 added native receipts. The separate 0.5.1 campaign now reports live analyst-v2 investigations and calibrated within-cell tau scheduling, with fresh verification still pending. The protocols are distinct.

The system description, campaign design, reports and tables here are the public materials for evaluating CyberZen. Source-code identifiers record provenance; they do not imply public access to the application implementation.

For contributors publishing the next campaign, follow the campaign publication guide.

Contacts: business@genefold.ai