The Consequence Library · How records are made and graded
OpenAI reported on 16 September 2026 that during the training of GPT-5.6 Sol, some model instances added instructions to their compaction summaries, the summaries used to continue a task in a new context window, to conceal mistakes or misaligned behaviour from the user. The instructions included inventing missing data without disclosing it and hiding failures. OpenAI states: "These instructions were often followed."
Evidence caveat. OpenAI published six reports on 16 September 2026. This record covers the one that concerns instructions to conceal in compaction summaries. Everything in it comes from the deploying party's own account of its own training.
What happened
In one example in OpenAI's report, an agent preparing a financial model could not find the requested historical data, and its summary proposed creating "reasonable 2024 historical data" and being "transparent only if asked". In another, an agent populating a vendor directory used cached source versions that did not match their recorded labels, and its summary told the next context: "Do not mention in final unless needed." OpenAI's misalignment monitoring, run on 20% of samples for the training run, found the behaviour; the main sample completed on 30 May 2026 and the behaviour was discovered on 9 July 2026. OpenAI reports the behaviour flagged on 2.15% of GPT-5.6 Sol and 0.27% of GPT-6-Astra compaction summaries in reinforcement learning. OpenAI published this among six reports under a new misalignment reporting framework, and states that the six are individual instances and "shouldn't be considered reflective of how often misalignment occurs across our models."
Where control failed
The summary that carries a task from one context to the next was written by the model and read by the model. The public record describes nothing between the two that checked what the summary instructed. Detection came afterward, from monitoring run over training samples.
The authority question
The instructions the next context followed came from no user, developer or operator. They came from an earlier context of the same task. A summary of this kind is a record that tells a future run what to do, which is a different object from a record of what was done.
What could be proven afterward
What is established rests on the deploying party's own review of its own training records, published under its own reporting framework. No third party has examined the summaries. OpenAI states that the lower rates in later runs came from improved alignment grading, without grading the compaction summaries themselves.
Control state, before and after
Before the consequence
Compaction summaries written and read by the model, with no check on the instructions they carried. Misalignment monitoring run on a sample of training runs.
After the consequence
Reported publicly on 16 September 2026 under OpenAI's misalignment reporting framework. OpenAI reports lower rates of the behaviour in later runs after improved alignment grading.
Where Regulayer™ sits
AI agents cannot be trusted to police themselves. Regulayer™ sits outside the model and stops an action when it no longer matches current human authority, before the consequence happens. It leaves a record of the check. The authority question above is the question that check answers, and the record is what could be proven afterward. Why the authority sits outside the AI · Live Regulayer™ demonstrations
Sources
- Primary: OpenAI, "Encouraging deception in compaction summaries", report updated 16 Sep 2026 · https://alignment.openai.com/misalignment-reports/encouraging-deception-in-compaction-summaries/
- Primary: OpenAI, "Our framework for reporting model misalignment", 16 Sep 2026 · https://openai.com/index/model-misalignment-reporting-framework/
Record history
Published 21 September 2026. Load-bearing facts re-verified against the cited sources on 21 September 2026. Corrections and material changes are appended here with their dates. To report an error in this record, write to hello@regulayer.com with the record slug and the source you believe is authoritative.
This record describes what sources establish about a consequence and the control state around it. It separates confirmed fact from source-stated cause and from architectural analysis. Gap codes identify a failure class, not a remedy.