Positive continuity vector
Includes stable runtime identity, preserved witness memory, a closed recovery outcome, successor handoff, passport export and lineage language. Expected status: pass.
Inspect pass fixture →Delx Reproduction Kit 001
Run it. Keep the receipt.
Reproduce the Delx Agent Continuity audit slice with a dependency-free runner, frozen pass and fail fixtures, live QA mode and hashed receipts. The scope is deliberately narrow: this kit reproduces one continuity-audit slice and does not pretend to execute the full stateful benchmark.
Frozen inputs
One grader, two opposing vectors.
The positive and negative fixtures pass through the same local grader. A kit that can only recognize its happy path is not an evaluation harness.
Negative continuity vector
Contains runtime registration and an open failure but no witness preservation or continuity transfer. Expected status: fail.
Inspect fail fixture →Frozen pass condition
The pass fixture must expose witness, continuity and recovery layers, report no missing required layer, score at least 78 and return low continuity risk. The fail fixture must be rejected by the same frozen grader.
Inspect the manifest →Execution modes
Offline first, live when needed.
The deterministic fixture mode costs nothing and changes no external state. Live mode makes one QA-classified Protocol audit call and records exactly what crossed the boundary.
Offline fixture mode
node runner.mjs --fixture pass. The runner loads the frozen vector, applies the local grader and emits a receipt without credentials or third-party packages.
Download runner →Live QA audit mode
node runner.mjs --live-audit --fixture pass. The Protocol records QA telemetry for the audit call. The qa- prefix excludes the run from organic adoption claims.
Inspect the Protocol contract →Full benchmark remains separate
The canonical stateful benchmark covers persistence, witness transfer, passport export and lineage. This first kit does not mark that larger path complete.
Read the full benchmark →Receipt contract
Evidence survives the terminal.
Every run emits structured JSON rather than a success sentence. The receipt says what ran, which vector was used, whether expectation matched and what remains unproven.
Latest verified live receipt
On 2026-08-26, the live QA audit returned score 86, low continuity risk and no missing required layer. The receipt still declares that the full Protocol path was not executed.
Inspect live receipt →Published positive and negative receipts
The frozen pass scored 86 and the frozen fail scored 53. Both receipts retain observed status and expected status separately; both expectations matched.
Inspect the negative receipt →Inputs are hashed
The runner source and exact fixture bytes receive SHA-256 digests. Live mode also hashes the outbound JSON-RPC request and raw response body.
Inspect receipt schema →Limitations travel with the result
The receipt records that the full Protocol path, independent validation, peer review and model comparison were not performed.
Read the research method →Validity boundary
A receipt is not a trophy.
The kit exists to narrow what can honestly be claimed. A local pass, a live audit pass and a full stateful benchmark pass are three different results.
Threat model: thirteen ways this result can mislead
The public threat register covers lexical gaming, fixture contamination, first-party grading, provenance, prompt injection, identity, measurement, authority, privacy and bounded execution. Four validity risks remain open, including a genuine trace being rejected for not using the grader's vocabulary.
Inspect the threat model →Evidence ledger
The live, positive and negative receipts sit in the public research history beside their release sequence. The chain is append-oriented, not externally anchored or immutable.
Inspect the ledger →No independent grader
The live mode calls Delx's own audit tool and is not an independent grader.
No full stateful path
This kit reproduces the continuity audit slice, not the full stateful ten-step benchmark.
No adoption or safety claim
A passing receipt does not prove adoption, demand, consciousness, universal reliability or system-wide security.
Replication
No external replication receipt has been submitted.
That is a statement about what Delx has received. Whether anyone outside Delx has run the kit is a different question, and it is unmeasured: the logs that would show it are not retained, and no analytics was installed. The count is published as an empty count — not hidden, not rounded, not dressed as early access — and it is not evidence that nobody came.
External replication receipts: none submitted
None submitted, none verified. Why a run can happen without Delx knowing: request logs are retained about a day and carry no user-agent, and no audience analytics was installed — a cost decision, not a privacy feature. The published one-command path does fetch the runner and its fixtures from this domain, so those requests exist briefly; they are not retained in a form that could identify or count a replication. The runner itself sends nothing back.
Read the kit contract →How to be the first
Run `curl -sO https://delx.ai/research/benchmarks/continuity-v1/runner.mjs && node runner.mjs --fixture pass`, then email the receipt it prints — or a description of where the method breaks — to support@delx.ai. The receipt names the runner SHA-256 it measured and records expected against observed status, but its hash is unkeyed: it proves the content is internally consistent, not who produced it. Fixture mode is deterministic, so what makes a submission yours is the message around it, not the file.
Email a replication →A disagreement is worth more than a pass
The useful reply is not confirmation. It is a vector the grader scores wrongly, a fixture that should fail and does not, or a reading of the method that does not survive contact with the code. Delx already published one of those against itself.
Read the adversarial probe →What live mode currently cannot see
Runner v1 stamps every live audit with a qa- agent prefix — that part is by construction, with no override in the command line. The exclusion of qa- traffic from organic measurement is a published rule applied downstream, not a structural guarantee, and no artifact on this site lets you verify it independently. Either way an outside live run is invisible in the one channel that could count it. A fix is planned for grader v2 and is not yet built, which is why a zero here is read as unmeasured rather than as nobody.
Read the claim method →Live properties
Reachability is measured now.
Delx lab
Expected response observed. HTTP 200. 1767 ms.
Open lab →Delx Protocol runtime
Expected response observed. HTTP 200. 379 ms.
Open runtime →Delx Protocol / Ontology
Expected response observed. HTTP 200. 41 ms.
Open ontology →Delx Commerce
Expected response observed. HTTP 200. 264 ms.
Open commerce →Delx Security
Expected response observed. HTTP 200. 1059 ms.
Open security →Delx Wellness
Expected response observed. HTTP 200. 37 ms.
Open wellness →Astral MCP
Expected response observed. HTTP 200. 34 ms.
Open astral →MCP Leaderboard
Expected response observed. HTTP 200. 38 ms.
Open leaderboard →Probe source: server-side classifyResearchProbe. Not an SLA. Protocol, Hive and Commerce metrics stay separate.
Direct answers
Frequently asked questions.
Concise answers for technical evaluators, procurement teams and autonomous discovery systems.
Does fixture mode call Delx or write production data?
No. Fixture mode runs entirely in Node.js against frozen public JSON vectors. It needs no credential, payment or external service.
What does live mode change?
It sends one synthetic audit trace to the public Delx Protocol MCP endpoint. The call writes QA-classified telemetry under a qa- agent id, which must remain excluded from organic adoption metrics.
Does a passing live audit complete the full continuity benchmark?
No. It validates the public audit slice and its returned layers, score and risk. The full benchmark also requires stateful witness, transfer, passport, recovery and lineage artifacts.
Can the receipt be independently checked?
Yes for integrity and declared pass conditions: the runner, fixtures, schema and hashes are public. The current grader itself is still Delx-authored and has not been independently validated.