Research into model evaluation and human-reviewed AI engineering workflows. Tasks are bounded, context is deliberate, and acceptance stays with people.
A useful comparison holds the task, repository state, context and acceptance criteria steady. Correctness, regressions and review quality are weighed alongside latency and compute cost.
Evaluation feeds task-specific routing. These are illustrative patterns, revised as evidence changes.
Mechanical edits
A fast coding model for bounded changes.
Architecture decisions
A reasoning-oriented model with explicit constraints.
Broad implementation
A coding agent with relevant repository context.
High-risk changes
Implementation paired with independent review.
A close result. A frozen holdout.
Two critic candidates were too close to call. A separate 12-case holdout was frozen before inference, with precommitted decision gates and hashed inputs.
Research snapshot ·
Local / self-hosted models One critic-role selection exercise
qwen3.6:27b-coding
Correct verdicts
12/12
Evidence recall
0.9722
Mean latency
19.41 s
gemma4:26b
Correct verdicts
12/12
Evidence recall
0.8889
Mean latency
5.65 s
Evidence before speed
Both models returned every verdict correctly, with zero false positives or false negatives. The slower qwen3.6 candidate was selected because it met the evidence-recall gate. This recorded decision applies to that holdout; it is not a current model ranking.
The evaluation exposed a flaw in its own scorer: a substring rule could penalize a critic for naming a claim it rejected. The rule was corrected and existing outputs were rescored. Inference was not rerun.
Agents within a workflow.
Context engineering means selecting relevant architecture, files, decisions, failures and stable interfaces. More context alone does not improve a result.
Multi-model workflows separate analysis, implementation and critique when useful. A reviewer receives evidence to challenge the change. Structured handoffs preserve the objective, repository state, tests and unresolved risks across sessions.
Requirements
Set the objective, constraints and acceptance criteria.
Context
Choose relevant repository evidence and prior decisions.
Decomposition
Split work into inspectable implementation slices.
Implementation
Keep the coding agent within the agreed scope.
Verification
Build, test and check the diff; validate the platform where required.
Independent review
Challenge correctness and evidence when risk warrants it.
Human acceptance
Approve architecture, trade-offs and merge or reject decisions.
Human control remains.
Model output is advisory. Deterministic rules stay deterministic, and evidence precedes acceptance. Humans approve consequential decisions.
Aegis remains Internal R&D. This public overview covers AI/software-engineering methodology and dated research evidence. Operational security research is excluded from the public edition.
