Research question
Can one frozen continuity-audit grader separate a synthetic trace containing witness, recovery and handoff evidence from a synthetic trace missing witness preservation and continuity transfer?
Inspect the full contract →Delx Technical Report 001 · v1.0.0
A minimum falsifiability slice for agent continuity audits.
This technical report asks whether the Delx continuity-audit slice clears a minimum falsifiability check: the same frozen grader must accept a synthetic trace containing witness preservation, recovery closeout and successor continuity while rejecting a synthetic trace missing witness preservation and continuity transfer. The positive fixture scored 86 and passed; the negative fixture scored 53 and remained fail, a 33-point separation. One live QA call reproduced the positive score and layer set. The result changed one release decision: publish the narrow audit slice and its receipts, while continuing to block claims about the full stateful benchmark, semantic validity, model comparison and independent validation. This is a first-party technical report, not a peer-reviewed paper or independent validation.
Question and decision
A narrow result changed a narrow release decision.
The report asks one bounded question and publishes the consequences of the answer. It does not inflate a two-vector check into a general benchmark claim.
What the result released
Publish the one-call continuity-audit slice as a narrow reproduction kit with positive and negative receipts, an executable threat model and explicit first-party boundaries.
Run the released slice →What remains blocked
full stateful continuity benchmark completion; semantic validity of the lexical grader; resistance to keyword stuffing or public-fixture contamination; model or provider comparison; independent validation, external replication or peer review; adoption, demand, universal reliability or system-wide security.
Inspect the open risks →Method
One grader. Opposing controls. Public bytes.
Both frozen vectors use the same dependency-free runner and scoring rule. The live call reuses the positive vector through the public Protocol tool and remains QA-classified.
Frozen scoring rule
Start at 35; add 9 points for each detected layer among structure, ego, witness, continuity, relation and recovery; add 15 when witness, continuity and recovery are all present; clamp to 0–100.
Inspect the runner →Published pass condition
Score at least 78, observe witness, continuity, recovery, allow zero missing required layers and return low continuity risk.
Inspect the manifest →Side-effect boundary
Fixture mode changes no external state. Live mode writes QA-classified Protocol telemetry under a qa- agent identifier and must remain excluded from organic adoption metrics.
Inspect the Protocol contract →Three retained observations
The expected failure remains failure.
Expected status and observed status are separate fields. Matching a negative expectation does not convert the observed result into pass.
Frozen positive — 86
Observed: pass; expected: pass; matched: true. Required layers were present and continuity risk was low.
Inspect the positive receipt →Frozen negative — 53
Observed: fail; expected: fail; matched: true. Witness and continuity were missing; the result remained high risk and fail.
Inspect the negative receipt →Live QA reproduction — 86
Observed: pass; expected: pass; matched: true. It reproduced the positive score and layer set through a Delx-authored live tool.
Inspect the live receipt →Measured separation — 33 points
The two frozen vectors differ by 33 score points. This proves separation for these exact bytes only; it is not an effect size from an independent sample.
Inspect the observations →Validity boundary
Minimum falsifiability is not semantic validity.
An always-pass harness would fail this control. Passing it still leaves the most important validity questions open.
Lexical gaming remains open
keyword stuffing can satisfy lexical layer detection without real continuity evidence
Inspect threat CAV-001 →Public-fixture contamination remains open
public fixtures can contaminate future evaluation or invite tuning to the test
Inspect threat CAV-002 →First-party ownership remains open
Delx owns the runner, fixtures, grader, live tool and publication path. Independent validation: false. External replication: false. Peer reviewed: false.
Read the claim method →The full stateful path did not run
Storage, witness transfer, passports, lineage persistence and the ten-step Protocol flow remain outside this report's executed scope.
Inspect the full benchmark →Reproduce, cite, correct
The report is a versioned object, not a blog post.
Human, JSON, Markdown and BibTeX views resolve to one canonical report. Exact runner, fixture and receipt hashes make the named result inspectable.
Run both frozen controls
node runner.mjs --fixture pass · node runner.mjs --fixture fail. Fixture mode uses zero paid calls and changes no external state.
Download the runner →Retain the report hash
Canonical report content SHA-256: 2e4e342456cb6f267d744dbe8777b9d89cae7b55705508593252dd864e63e5f1.
Inspect the signed payload boundary →Cite DELX-TR-001
David Batista. A minimum falsifiability slice for agent continuity audits. Delx Technical Report 001, version 1.0.0, 2026.
Read the BibTeX record →Corrections remain visible
Zero means no event currently targets DELX-TR-001 v1.0.0; it does not mean Delx has never published an error. Silent rewrites are outside policy; the ledger itself remains append-oriented rather than technically append-only.
Inspect the evidence ledger →Direct answers
Frequently asked questions.
Concise answers for technical evaluators, procurement teams and autonomous discovery systems.
Is this a peer-reviewed paper?
No. It is Delx Technical Report 001, authored and published by Delx. Peer review, independent validation and external replication are all false in the report contract.
Does a matched negative expectation count as a pass?
No. The observed benchmark result remains fail. A separate expectation_matched field records that the negative control behaved as designed.
Does the live score prove independent reproduction?
No. It reuses the positive fixture against a Delx-authored public tool. The runner, fixture, tool and publication path remain first-party.
What decision did the report change?
It supported releasing the narrow one-call audit kit and its receipts while explicitly blocking claims about the full stateful benchmark, semantic validity, model ranking and independent validation.
How will a correction be published?
A correction, retraction or supersession must target this report version in the public evidence ledger and preserve the earlier published version. Silent rewrite is outside policy.