Insight AEGI Shield

AEGI November 2026 Engineering Update: Offline Fraud Evaluation and Replayable Reports

The November integration adds a supplied-score evaluation adapter and four replayable synthetic report examples, with the engineering scope and next steps recorded.

What changed

AEGI completed a reviewed source integration on 7 November 2026 covering its controlled synthetic Fraud workflow, durable job handling and a separate offline supplied-score evaluation adapter. The tested source passed 388 regression tests on each of SQLite and PostgreSQL. All three associated integration workflows completed successfully.

The new offline route makes a concrete evaluation deliverable inspectable: a readable report, queue-change table, structured results, frozen configuration and integrity manifest. Four authored examples show how the adapter reports favourable, unchanged, negative and incomplete findings.

This advances the engineering and delivery basis for a scoped evaluation. An institution-specific engagement still needs an agreed question, suitable inputs, approved handling and an accountable decision owner.

Integration evidence

Check Recorded result What the result supports
Source integration Reviewed and merged on 7 November One integrated source history for the declared scope
SQLite regression 388 passed; no test failures or skips Behaviour covered by the SQLite test suite
PostgreSQL regression 388 passed; no test failures or skips Behaviour covered by the PostgreSQL test suite
Integration workflows All three successful Full Ops verification, fresh Fraud run and Ops–Core preview checks at the tested head
Durable-job checks Passed in the recorded test environment Restart recovery, competing requests, claim and quota behaviour
PostgreSQL restore and Docker checks Passed in the recorded test environment Restore checks and container execution
Image save and reload Identity matched The saved image reloaded as the tested image

The tested source and integration merge are identified in the internal engineering record as 8ca3965…e9456 and b283eac…543a. The merge source tree matches the tested integration tree. These abbreviated Git commit IDs identify the scope; full identifiers remain in the engineering evidence.

An image that passes local or CI checks still needs configuration and acceptance on its intended hosting target. This update does not report an AWS deployment or completed customer access journey.

The new offline supplied-score route

The adapter compares two existing scores for every case in one frozen historical cohort. It uses one common decision time, one review capacity K and one outcome cutoff. Higher scores receive higher priority; tied scores are ordered by case identifier, with boundary ties disclosed.

The institution supplies the baseline and candidate scores and their generation provenance. The adapter measures which cases each ranking places within K, the eligible positive labels in those selections and the changes between the queues. It does not train or call a model, generate a Shared Context treatment or invoke Core verification.

Input validation rejects contract errors such as duplicate identifiers or missing required scores. Unknown or immature labels are preserved. Under this version’s completeness rule, any unknown label in the cohort produces Insufficient Evidence.

Read the input and scoping guide.

Four generated reporting examples

Each example uses 12 deliberately authored cases and K=4. The scores and labels were chosen to exercise a specific output branch, so the numerical differences are demonstrations of reporting behaviour.

Example Baseline TP@4 Candidate TP@4 Difference Proposed outcome
Favourable 1 4 +3 Continue
Unchanged 1 1 0 Refine
Negative 1 0 −1 Stop
One unknown label 1 3–4 possible +2 to +3 possible Insufficient Evidence

The incomplete-label range is a bound on the possible count, not a statistical confidence interval. A positive-looking bound does not override the separately declared completeness rule.

All four examples were generated and replay-checked. The public content preparation also reproduced them on 7 November. Each output records that no human decision has been made and that no production authority is granted.

Inspect the worked package example · Understand the incomplete-label result.

What replay checks establish

Verification checks the file inventory and hashes, then recomputes the derived results from the retained input. This can detect a modified report even if someone also recalculates its manifest hashes but leaves the frozen input and configuration unchanged.

That check establishes consistency under the declared tool and inputs. It does not authenticate the original source data or independently validate the evaluation method. A party controlling all inputs can produce a different internally consistent package, which is why the engagement must preserve its approved input, configuration and execution references.

The recipient boundary also matters. Queue tables and reports can contain pseudonymous cases and scores. Their handling and permitted recipients are agreed before an actual engagement.

How this relates to earlier evidence

The October engineering article remains a historical record of the pinned 60-case synthetic workflow, its 276-per-database regression result and 12 populated recovery checks. Its draft or blocked-integration descriptions are superseded for the scope merged in November.

The BAF-003 benchmark remains a separate public synthetic representation comparison. Neither the 12-case reporting fixtures nor the regression counts extend its performance claims. The six-application illustration on the homepage remains a manually authored explanation of queue movement.

The November record is internal engineering evidence, not an independent audit. Core qualification, live model brokering, operational actions and acceptance of a customer hosting environment retain their own work and evidence requirements.

What can be scoped next

A prospective engagement can start with the workflow, current baseline, proposed candidate, review capacity and available outcome evidence. Where paired scores fit the contract, an offline evaluation can be scoped in an approved environment without depending on a hosted demo.

The agreed work covers qualification, mapping, freezing the comparison, execution, replay, outcome review and a real owner decision record. Fee, schedule and support arrangements are settled before chargeable execution. A valid finding can be Continue, Refine, Stop or Insufficient Evidence.

Scope a Fraud Review Evaluation or review the evidence hierarchy.

NEXT STEP

Bring one proposed change.

Controlled Evaluation Bring One Question