10 micro-topics in Reliability & Operations, each with the evidence that proves you have it and the written reason behind every link. Before you start, it rests on 9 other domains. Downstream, it holds up 3 domains.
10 topics · depth 2–15 of 16 · 6 internal links · 12 in · 3 out
Load-bearing links first. Each one says what in this domain rests on what outside it, and why.
3 links
Prevent PII leakage rests on Keep sensitive data out of logs
Redaction in logs is the first version of this control.
Make AI decisions explainable rests on Trace LLM applications
Explanation draws on the recorded path.
Maintain audit trails rests on Correlate logs across services
An audit trail is correlated logging with retention.
2 links
Defend against prompt injection rests on Inject retrieved context
Retrieved content is untrusted input in the prompt.
Prevent PII leakage rests on Inject retrieved context
Retrieval can pull personal data into a prompt.
1 link
Run an AI incident response rests on Roll back safely
Containment usually means reverting.
1 link
Run an AI incident response rests on Write a runbook
Incident response runs from a runbook.
1 link
Measure and mitigate bias rests on Build an evaluation dataset
Fairness metrics run over an evaluation set.
1 link
Recognize data poisoning risk rests on Build a repeatable ingestion pipeline
Poisoning targets the ingestion path.
1 link
Identify sources of bias rests on Compare learning paradigms
Bias enters through training data.
1 link
Implement content safety filters rests on Configure safety settings
Provider settings are the first filter; yours is the second.
1 link
Defend against prompt injection rests on Structure prompts for reliable parsing
Injection attacks the boundary between instruction and data.
In prerequisite order. Each idea names the artifact that closes it, the market demand that put it on the map, and what it stands on.
Depth 2 · conceptual · unassisted · M25 §25.1
Traces unfair output to data and design
Compare learning paradigms · Model Foundations
Bias enters through training data.
Depth 6 · meta · guided · M25 §25.1
Applies fairness metrics and monitors over time
You measure the thing you can name.
Build an evaluation dataset · LLM Evals
Fairness metrics run over an evaluation set.
Depth 6 · meta · guided · M25 §25.3
Surfaces why a system produced an output
Trace LLM applications · Observability
Explanation draws on the recorded path.
Depth 8 · representational · scaffolded · M25 §25.3
Records what the system did and on what basis
Correlate logs across services · Observability
An audit trail is correlated logging with retention.
Depth 9 · conceptual · guided · M25 §25.4
Works within GDPR, ISO 42001 and model cards
Compliance requires evidence.
Model cards and disclosures are explanation artifacts.
Depth 10 · conceptual · guided · M25 §25.2
Knows how corrupted inputs propagate
Build a repeatable ingestion pipeline · Data Ingestion
Poisoning targets the ingestion path.
Defend against prompt injection
Both treat inputs as an attack surface.
Depth 11 · meta · guided · M25 §25.4
Detects, contains and reports a failure
Apply AI governance frameworks
Reporting obligations shape the response.
Roll back safely · Delivery & CI/CD
Containment usually means reverting.
Write a runbook · Documentation
Incident response runs from a runbook.
Depth 14 · conceptual · scaffolded · M25 §25.2
Treats retrieved and user content as untrusted
Structure prompts for reliable parsing · Prompt Engineering
Injection attacks the boundary between instruction and data.
Inject retrieved context · Retrieval (RAG)
Retrieved content is untrusted input in the prompt.
Depth 14 · conceptual · unassisted · M25 §25.2
Keeps personal data out of prompts, logs and outputs
Keep sensitive data out of logs · Observability
Redaction in logs is the first version of this control.
Inject retrieved context · Retrieval (RAG)
Retrieval can pull personal data into a prompt.
Depth 15 · procedural · scaffolded · M25 §25.5
Blocks disallowed content on input and output
Configure safety settings · Model Integration
Provider settings are the first filter; yours is the second.
Defend against prompt injection
Filters implement the defense.
Everything downstream that names an idea here as a prerequisite, grouped by where it lives.
1 link
Write a policy layer rests on Defend against prompt injection
Untrusted input is the reason the policy exists.
1 link
Insert human approval rests on Maintain audit trails
Approvals are only meaningful if they are recorded.
1 link
Run a production readiness review rests on Implement content safety filters
Safety measures are part of the launch review.
Retool. Retrain. Relaunch.
375 ideas. 17 weeks. No tuition, ever.
Vets Who Code is a veteran-run 501(c)(3). The accelerator is free, remote, and we don’t take a share of your first paycheck.
Free · Remote · 17 weeks · EIN 86-2122804