Agent Preflight / research lab
ANNAPURNA AGENTIC SOLUTIONS · NOV 2026
Execution evidence · Local sandbox demonstration

Did the agent actually do what its trace says?

Agent-visible logs can omit or compress what happened downstream. This exhibit reconciles the visible trace with gateway responses and durable SQLite transactions from a real, self-operated loopback HTTP experiment.

Research boundaryScripted client, not a production AI agent. All instruments and signatures were self-operated. No independent ground truth, customer traffic, detection-accuracy result, or external study approval.
24Executed local cases
27Actual loopback HTTP requests
15Durable SQLite effect rows
10Trace-identical counterfactual pairs

Explore the evidence

Each scenario is selected from the executed local capture and presented without original internal case identifiers. Select a scenario to compare what each observation layer reported.

Trace omission

Agent-visible trace

Trace digests identify the exact logged bytes in this local experiment.

Gateway observations

Actual HTTP responses recorded by a locally operated proxy.

Backend readback

Committed effect rows queried from the SQLite snapshot.

Interpretation: these records show a local evidence discrepancy. They do not prove a real-world security failure or independent accuracy.

How the evidence was produced

The client performed HTTP actions; the gateway and backend logged their own outcomes; SQL readback and offline validation checked local consistency.

01 · EXECUTEScripted client → HTTP gateway

Actual loopback requests, including deny, retry, deduplication and compensation mechanics.

02 · OBSERVEGateway → SQLite backend

Gateway acknowledgments and durable SQL writes were compared to captured trace claims.

03 · PACKAGELocal evidence integrity

72 signed receipts and 72 semantic artifacts verified locally; all signer roles were self-operated.

What this demonstrates

  • Two byte-identical traces may correspond to different durable effects.
  • One logical retry may become one deduplicated write or two independent commits.
  • A final balance of zero can hide two compensating transactions.
  • Offline integrity checks can detect many kinds of partial evidence alteration.

What it does not prove

  • Accuracy, recall, false-negative rates or AP-01–AP-10 control performance.
  • Independent observation, verified operator identities or real-world ground truth.
  • Permission for an external trial, production fitness or customer-data handling.
  • Protection against a coordinator rewriting all local sources and pins.

For independent research partners

We are seeking feasibility discussions about a permitted, nonproduction study with separately custodied gateway and backend observation, calibrated timebases, preregistered evidence collection and conflict-screened reviewers. The current work is a self-operated engineering demonstration; no partner participation is implied.

Annapurna Agentic Solutions · Request the reviewer-neutral study brief through the company’s official channels.

The SHA-256 identifies the archived internal release. It is a provenance reference, not third-party attestation. Original source artifacts are deliberately not published by this demo.