One AI Safety Contract, Multiple Enforcement Points: Why Prompts Are Not Enough

·1 views

A new framework argues that AI agent safety cannot rely solely on prompts or instructions, since agents can bypass them without structural controls in place. The proposed approach centers on a single declared contract file, enforced at multiple mandatory checkpoints rather than through informal guidance. A local runner provides fast feedback and early refusals, but is considered optional and insufficient as a final enforcement layer. CI pipeline checks, made mandatory through branch protection policies, serve as the first non-optional gate to verify that only contract-compliant changes are merged. Runtime sandbox and capability boundaries form the second mandatory enforcement point, ensuring agents cannot reach unintended external effects even after passing CI.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Two API Settings Tripled GPT-5.6 Sol's ARC-AGI-3 Benchmark Score, OpenAI Says

OpenAI reported that enabling two Responses API settings — retained reasoning and compaction — raised GPT-5.6 Sol's score on the ARC-AGI-3 public benchmark from 13.3% to 38.3%. The initial evaluation harness prevented the model from carrying forward reasoning across steps of the benchmark's interactive 2D puzzle tasks, limiting its performance. Alongside the score improvement, the new configuration also reduced output token usage by approximately six times. OpenAI emphasized that this outcome highlights how API configuration, not just model capability, can significantly influence both agent performance and token efficiency. The findings serve as a practical reminder that benchmark scores reflect the combined effect of the model, its settings, and the evaluation setup rather than model weights alone.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Define Your Data Grain Before Picking a Chart, Analysts Warned

A data modelling guide published on DEV Community argues that the most common error in business reporting is not poor chart design but mismatched data grain — the unit that each row in a dataset represents. Mixing invoice-level, line-level, payment, and financial-summary data in a single table causes totals to multiply and filters to answer the wrong question. The article illustrates the risk with a concrete example: joining a £1,000 invoice total across five line items and summing the result incorrectly yields £5,000. Analysts are advised to state the reporting question in plain language first, then select the appropriate grain — such as one row per invoice, per product line, or per payment — before building any visualisation. The guide also recommends keeping separate fact tables for each natural grain and documenting non-additive measures to ensure dashboards remain accurate when data is refreshed.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Most Startups Should Avoid Complex Infrastructure and Right-Size Their APIs

A software engineer argues that startups routinely over-engineer their infrastructure by adopting Kubernetes and complex setups long before user scale demands it. In one real-world case, migrating a project from a costly Kubernetes cluster to AWS Lambda, SQS, and App Runner cut monthly infrastructure costs from thousands of dollars to roughly $50. The post breaks down serverless platform limits, showing that AWS Lambda can theoretically handle up to 8.6 billion requests per month, well beyond what most early-stage startups require. The author recommends using managed containers for user-facing APIs and reserving serverless functions for event-driven or bursty workloads like webhooks and scheduled jobs. The central advice is to treat scaling as an 'earned problem' — only invest in complex infrastructure once actual traffic demands it, not in anticipation of hypothetical growth.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a 'Dead-Man's Switch' Catches Silent Cron Job Failures Before Data Rots

A solo developer running multiple production systems discovered a critical blind spot when a scheduled curation job silently stopped running, leaving stale data served undetected for days. Unlike loud failures that trigger error alerts, the job produced no exceptions or log warnings, and all downstream health checks remained green. The developer's fix involved a two-part dead-man's switch: the job writes a timestamp only on verified success, and a separate watcher on an independent runtime alerts if that stamp grows too old. Crucially, the watcher runs in a different failure domain so it can catch the job's silence even if the job itself is completely dead. The threshold for triggering an alert must be tuned to the content's natural update rhythm rather than the job's schedule, to prevent alert fatigue from rendering the safeguard useless.

0 comments Read more at DEV Community