blueDelta

Manifest diff and candidate minimization

Compares two versioned lab run manifests, excludes only the candidates ruled out by an explicit dependency rule, and — given a test catalog — recommends the cheapest experiment that would separate what's left.

Module — in active build

Request a pilot conversation Read the manual

Runs locally · CLI or web UI · Ed25519-signed receipts · no cloud account required

Who it's for

Lab and research-integrity reviewers.

What it is not

Never claims a surviving candidate caused the result, never runs an experiment or controls an instrument, and never infers from data the manifests don't contain.

Screens

blueDelta answers one narrow question: between a baseline run and a run that diverged, which recorded differences are still consistent with the divergence once the dependency rules have excluded the rest — and, given a test catalog, which experiment would separate the survivors most cheaply? The four screens below show the library of saved comparisons, one finished run where more than one candidate is still standing, the field-by-field evidence behind it, and the signed receipt that closes it out.

blueDelta Library screen: the suite shell's sidebar (Case intake with Library, New analysis, and Import, Evidence map, Reviewer, Case file) beside a table listing three saved comparisons — DELTA-0003 Draft, DELTA-0002 Active, and DELTA-0001 Active — with a New analysis button. — Synthetic demo data
Library — every saved comparison, filterable by Draft, Active and Completed, searchable by ID, title or status. Start from files takes a baseline and a divergent run. The footer on every screen: blueDelta reports recorded differences and candidate consistency — it does not assert cause.

Open a comparison and the Workbench shows where its minimization stopped.

blueDelta Workbench run-detail screen: a Multiple-candidates minimization verdict, 24 evidence objects, both baseline and divergent manifests at 100% completeness, and a recommended cheapest separating test at $150.00. — Synthetic demo data
Workbench — run finished, 24 evidence objects, both manifests 100% complete against what the rules require. Minimization state MULTIPLE-CANDIDATES: more than one recorded difference survives, so blueDelta names no cause. Recommendation SELECTED: T-FULL-SEPARATOR is the cheapest catalogued test that would tell the survivors apart. The claim boundary sits above the numbers, not in a footnote.

Evidence is one click away, and shows why the workbench called it MULTIPLE-CANDIDATES; when the rules leave a single survivor the state reads SINGLE-CANDIDATE, and when the manifests cannot support the comparison it reads CANNOT-CERTIFY.

blueDelta Evidence screen for a run: a MULTIPLE-CANDIDATES badge, three count cards reading 19 field deltas, 3 surviving candidates, 2 excluded candidates, a filterable table of field-level deltas each row showing the object, a verdict chip (UNCHANGED, CHANGED, or SURVIVING), the recorded left and right values, and an Inspect action, beside an empty evidence inspector panel. — Synthetic demo data
Evidence — every one of the 19 field deltas between the baseline and divergent manifests, typed UNCHANGED, CHANGED, or SURVIVING. Three surviving candidates, including a calibration date change on INST-01, are what the dependency rules could not exclude; Inspect opens the exact recorded left and right values behind any row.

Receipt is the same run's signed output.

blueDelta Receipt screen: a claim-boundary box restating that a surviving candidate is consistent with the divergence and is never asserted as its cause, a signed envelope listing schema version, product, contract hash, payload hash, output hash, input hashes, nonce, parent receipt and signer public key, a Verify receipt button, a Run tamper demonstration button, and a copyable command to verify the receipt fully offline with the standalone verifier. — Synthetic demo data
Receipt — the signed envelope for this run, with a boxed reminder above the numbers that a surviving candidate is consistent with the observed divergence, not its proven cause. Run tamper demonstration mutates a local copy to show a flipped signature byte gets caught; the offline command at the bottom performs the identical check with the application stopped.

Status, measured

SQLite persistence, API, and CLI landed (Milestone 6). See the status page for the measured current state.

A signed receipt proves what blueDelta ran and recorded; it does not prove an underlying real-world claim is true. blueDelta does not approve, deny, recommend, pay, retract, or alter anything a named human is responsible for deciding.