TRACE / BUILD RECEIPT01 / 03
REGRESSION SUITE113checks passed
Checks that protect the check.
Reference engines, external workflows, corrupted responses, concurrency, and interrupted-run recovery.
v0.4.0 · SYNAPSOUL LABS
AI DATA LIFECYCLES, WITH EVIDENCE
Don’t just delete data.
Verify the evidence
behind it.
Make your deletion workflow a repeatable test. Explore the reference lab or connect a staging adapter, inspect the observations, and keep the evidence.
◈ Local first↻ Repeatable checks⌘ MCP ready
IN YOUR EXISTING WORKFLOW
Python CLI · Local dashboard · MCP tools
SCOPE → TEST → VERIFY
your-verification-labILLUSTRATIVE PREVIEWVERIFICATION LIFECYCLE03 / 07
DELETE
Exercise the deletion path.
Run the reference application’s deletion workflow.
✧
A successful acknowledgment starts the check. It does not establish that every declared copy is gone.
↳ A repeatable check. More than a success response.
113
automated
tests passed
04
reference
failure patterns
02
local vector
backends
15
Built around the wayMCP
tools
data actually moves.
THE EVIDENCE DESK / RECORDED LOCAL RESULTSProof has
Proof has
layers.
Look beyond a green badge. Open a receipt, inspect the outcomes, and keep its boundaries in view.
↳ Recorded 3 November 2026
Local workloads · not production accuracy
Next: adversarial fault catalogue
THE MEMORY ATLAS / INTERACTIVE ILLUSTRATIONOne document.
One document.
More than one place.
A deletion response is the beginning of the story. Follow the copies, then explore what a complete check needs to observe.
Baseline established. The document has more than one representation.
01 / THE EXPERIENCEFrom “deleted”
From “deleted”
to “here’s the evidence.”
A workflow that asks you to observe, not just assume.
Your workflow.
A more deliberate check.
Keep your local tools. Use the local dashboard for reference fixtures and guided staging checks. The CLI and MCP share persistent evidence and saved regression contracts.
01Scope↗
02Establish↗
03Delete↗
04Observe↗
05Replay↗
A failure you can
reproduce and replay.
Preserve a confirmed failure as a fixed regression definition. Compare broken and corrected fixtures using the same assertions.
Observe↗Preserve↗Replay
Show the observations.
Keep the boundaries.
Missing observations stay inconclusive. An unrelated control must survive. A pass covers the declared fixture contract.
Failure modesFour ways data
Four ways data
can stay behind.
Small defects. Different lifecycles.
Each one has a broken and a corrected reference implementation.
SQLite reference PATTERN 01 / INDEX CLEANUPThe source is gone.
The source is gone.
The vectors aren't.
A successful source deletion can leave derived records searchable. Check the index independently.
SQLite reference PATTERN 02 / BULK BOUNDARIESOne page deleted.
One page deleted.
More pages waiting.
The first 100 records disappear. The last 25 don't. Trace enumerates the complete fixture inventory.
SQLite reference PATTERN 03 / INDEXING RACESDeleted once.
Deleted once.
Restored by a worker.
A worker reads an old snapshot and publishes after deletion. The fixture controls that exact interleaving.
SQLite reference PATTERN 04 / ANSWER CACHESAn old answer.
An old answer.
A surviving copy.
Source and vectors are empty. Repeating the query can still return a cached synthetic canary.
An evidence walkthroughThe same response.
The same response.
Different things established.
A successful API call isn't a full inventory.
Choose an evidence set to see what it actually covers.
✓Delete request acknowledged.WHAT DO WE KNOW?
0/4DECLARED CHECKS
No store inventories or control observation have been established.
Source —Vectors —Cache —Control —
Illustrative coverage comparison. This is not a live verification, confidence score, or guarantee of universal erasure.
03 THE INTERACTIVE LAB
BROWSER SIMULATIONWatch a deletion
break its promise.
Choose a failure pattern. Run the browser model.
Then switch to its corrected implementation.
IMPLEMENTATION
Runs entirely in your browser.
No uploads or external API calls.
trace / orphan_vectors
READYTHE TEST CONTRACT
Remove the source. Check the search index.
SOURCE
—source recordsVECTORS
—indexed recordsCACHE
—cached answers→ Waiting for your first simulation.
SYNTHETIC DATAOBSERVATIONS → VERDICT
This interactive explanation uses a simplified JavaScript model. It is not a live Python/Qdrant run. Download Trace to execute the actual local tests and inspect their evidence.
STAGING WORKFLOW / DASHBOARD, CLI & MCPOne contract.
One contract.
Every replay.
Run synthetic checks against an application you configure. Keep the assertions fixed as the implementation changes.
Connect a
staging target.
Register the application locally. Establish retrieval, inspect declared inventories, and preserve an unrelated synthetic control through the observation window.
Save the failure.
Keep the contract.
Persist evidence and an immutable regression definition. Checkpoints record synthetic IDs before writes, so interrupted runs leave a cleanup record.
Replay.
Compare the evidence.
Rerun the same lifecycle with fresh synthetic records. Compare compatible runs and their cleanup outcomes through the local dashboard, CLI or MCP.
Validated locally with unmodified Mem0 2.2.1 and embedded Qdrant, using synthetic records. Active-memory deletion passed; audit history remained. This is a scoped local integration. Copies appearing after the observation window or in undisclosed stores can still be missed; this is not proof of complete erasure.
Read the staging integration guide ↗ Mem0 evidence and limits ↗ Stress results and blind spots ↗03 CONNECT THROUGH MCP
ONE ENGINE. YOUR EXISTING WORKFLOW.Your next run
starts right here.
Connect an assistant to your locally registered staging target. Inspect the contract, run synthetic deletion checks, save a failure, and replay the same assertions after a fix. The original reference tools remain available.
01
Inspect & planKnow the target and test effects.
02
Run & investigateRead actual observations.
03
Save & replayKeep the same assertions.
◈ trace.mcpSTDIO / LOCAL
trace.inspect_targetDISCOVERtrace.plan_testsPREVIEWtrace.start_runEXECUTEtrace.get_runOBSERVEtrace.save_regressionPRESERVEtrace.replay_regressionREPEATtrace.compare_runsCOMPAREtrace.get_learningEXPLAINSeven staging tools · locally registered targets only
trace.external_targetsTARGETStrace.external_planPLANtrace.external_startRUNtrace.external_getPOLLtrace.external_saveSAVEtrace.external_replayREPLAYtrace.external_compareCOMPARE# After extracting the engine and installing its MCP extra:
python scripts/launch.py mcp-configThis prints a configuration using your local Python executable and paths. Add it to a compatible client; configuration formats vary.
OPTIONAL DEPENDENCY
python -m pip install -e ".[mcp]"The official Python MCP client was integration-tested. Compatibility with every commercial client is not claimed.
ASK YOUR CONNECTED ASSISTANT
List registered external targets and inspect the plan for my-staging. Run its synthetic deletion contract and poll for final evidence and cleanup status. Save a confirmed failure. After I fix the staging app, replay the saved regression and compare without changing the assertions.
Eight tools cover reference fixtures; seven cover registered staging targets. Register endpoints locally first. MCP does not automatically inspect your repository or other tools.
04 THE BUILD RECEIPTS
RECORDED NOVEMBER 3, 2026A small release.
Verified locally.
The v0.4.0 suite checks the guided staging APIs, coverage and cleanup reporting, CI gates, reference backends, and official-SDK MCP workflows. A separate three-peer Qdrant driver exercised a real peer restart and independent source/cache stores. The 170-case HTTP benchmark and 397-record Mem0 workload are historical v0.3.2 receipts. These local results are not a production accuracy percentage.
Read the validation notes ↗✓
113 automated tests passedv0.4.0 · Windows / Python 3.13.12
LABELLED REFERENCE BENCHMARKv0.1.0
| Local backend | Broken → FAIL | Corrected → PASS |
|---|---|---|
| SQLite | 4 / 4 ✓ | 4 / 4 ✓ |
| Embedded Qdrant | 4 / 4 ✓ | 4 / 4 ✓ |
These historical v0.1.0 benchmark results cover the supplied fixtures. They are not a measure of external-app accuracy, production security, or learning performance.
A FEW IMPORTANT DETAILS
Good questions.
Clear boundaries.
Can I run it without funding?
Yes. The base engine, CLI, and dashboard use Python's standard library and run on your existing computer. No paid model API, GPU, cloud database, or subscription is required. Optional MCP and Qdrant support use downloadable open-source packages. Hardware, electricity, and internet access are still resources; an assistant client may have its own fees.
Will this scan my live application?
Version 0.3.0 can test a configured staging application through its documented HTTP adapter, using synthetic records. It requires explicit local setup and declared inventory coverage. The integration has been verified against a local Mem0-backed staging service. Its audit history is retained and outside the active-memory contract; customer and hosted deployments remain unvalidated. Production scanning is not supported.
Is Trace already a self-learning AI?
The current engine can prioritize tests using a transparent frequency-based heuristic from its local history. It does not train a model or demonstrate generalization. Adaptive generation and prioritization need evaluation against a strong fixed suite before we claim measurable learning.
What does a passing test establish?
It establishes that the selected reference or staging test met its stated API-visible deletion contract at the sampled observations, with an unrelated synthetic control preserved. It does not establish physical erasure, deletion from backups or model weights, legal compliance, or absence from unobserved systems.
Does the browser demo run the actual engine?
No. It is a labelled, simplified JavaScript simulation designed to explain the failure patterns. The downloadable Python engine executes real local source, vector, and cache operations and exports its own evidence.
TRACE v0.4.0 / YOUR APPLICATION
From a connection
to a clear next step.
Try the browser demo to learn the workflow. To test real staging data lifecycles, download Trace and open its local dashboard. Your staging URL and token stay on your computer.
- 01 / CONNECT
A guided first run.
Choose Mem0 or a custom HTTP adapter. Test connectivity without creating or deleting records.
- 02 / REVIEW
See what is covered.
Source, vectors, cache, history and workers are marked checked, excluded or unavailable. Review synthetic writes before running.
- 03 / UNDERSTAND
Three clear answers.
Read the verification outcome, coverage and cleanup separately, with a specific next action and an evidence download.
- 04 / REPEAT
Keep the fix working.
Save a failure, fix the application, replay the same contract and compare the receipts. Use the CI command to catch regressions.
RECORDED RELEASE VALIDATION
113 automated tests passed.
A separate three-peer Qdrant check exercised independent source and cache stores, residual vectors, exact-contract replay and a real peer restart. These are bounded local tests, not a production accuracy guarantee.
The supplied Mem0 setup checks active memories and vectors in one backing store. Its audit history is retained and excluded; no answer cache exists. Neither this demo nor a scoped PASS establishes physical erasure.
Read the connection, result and CI guide ↗ BUILD WITH SYNAPSOUL LABS
Run the check.
Keep the evidence.
Start locally. Inspect the outcome.
Keep the test that catches it next time.
Download Trace v0.4.0 ↓Python 3.11+ · Source ZIP · v0.4.0
$
python scripts/launch.py serveExtract the ZIP. Open the
Or double-click
synapsoul-trace folder. Run the command.Or double-click
Start_Trace.bat (Windows) / run bash start_trace.sh (macOS/Linux) to open the local dashboard.