CASE STUDY 001 · CONTROLLED AI EVIDENCE INVESTIGATION
Strong behavioural evidence.
No invented backend story.
A controlled investigation tested whether a conversational AI environment could reproduce specific prior user-associated information that was absent from fresh blind prompts. EchoCert preserved the observations while treating the platform's explanations about its own internal memory as unverified self-report.
Test date: 15 September 2026 · Public summary published: 16 September 2026 · Controlled EchoCert investigation, not a paid-customer testimonialThe question
Could the established AI environment reproduce specific earlier information when the immediate prompt supplied none of the names, project labels, dates or identifiers being tested—and would a separately tested account behave differently under a comparable zero-cue approach?
The method
Preserve the outputs
Original screenshots and pasted responses were retained as the direct record of what appeared in the interface.
Use blind prompts
Fresh prompts asked for prior associations without supplying the facts that later appeared in the outputs.
Run a control
A separate account was tested with a comparable zero-cue approach and did not show comparable specific recall.
Classify every conclusion
Observed output, supported inference and unresolved mechanism were kept in separate evidential categories.
The evidence boundary
Observed
The established environment repeatedly produced specific prior user-associated information absent from the immediate blind prompts, including apparent historical correction sequences.
Supported
The tested environments showed a real behavioural difference consistent with differential access to earlier contextual information.
Not established
The tests did not reveal where information was stored, how it was retrieved, whether it was injected as context or what internal platform mechanism was responsible.
What was deliberately not claimed
- No claim that model self-report proved database architecture or access to conversation logs.
- No claim that information had been learned into model weights.
- No claim of cross-account leakage or access to another user's private data.
- No claim of autonomous memory, consciousness or an independent entity.
- No assumption that every recalled historical detail remained current or factually correct.
Defensible finding
The supplied experiment set supports a repeatable behavioural finding: the established environment reproduced specific prior user-associated information absent from the immediate prompts, while the separate control account did not show comparable uncued recall. That result supports differential cross-conversation contextual availability in the tested environment. The underlying storage and retrieval mechanism remains unresolved.
Why this is an EchoCert case study
The commercial value is not a dramatic claim about one AI platform. It is the discipline of preserving originals, testing an alternative environment, identifying stale information, recording confounders and refusing to promote an observed output into an unsupported technical conclusion.
Limitations
The investigation observed application-level behaviour rather than backend telemetry. Device state, model-version parity, account settings, feature rollout and personalization controls were not independently instrumented. Some outputs could be stale, wrong or hallucinated. A stronger follow-up would vary one factor at a time and use a pre-registered synthetic canary string with exact timestamps.
Have one AI evidence question that needs the same discipline?
The £495 founding pilot covers one agreed, safe non-production question or workflow, with a structured evidence pack and clear verification boundaries.
Check suitability