Looking for failures the suite has no category for
A system can't guard against a failure it has no category for. The Authority System Suite, a set of rulebooks that make an AI show its work, has a file about that: it lists seven ways of looking for such failures and says none of them guarantees anything.
The file, on unknown unknowns, begins with the problem. The audit compares the system against itself, and both failure catalogs (the lists of known ways the model and the maintainer go wrong) list known modes. A mode absent from both is invisible to every control in the repository.
The catalog of the model's failures now holds 42 behavioural failure modes. The file treats this as a count of what has been named, never of what exists.
Seven ways to look, each with a stated limit
Each method gets the same three lines: a method, what it reaches, what it cannot. One of them is an anomaly log, a log of events that fit no category. Every schema in this project forces classification, so an event that fits nothing gets forced into the nearest value or dropped. The log forbids classifying at recording time.
Its cost is stated too: low signal to noise, by design. The anomaly file holds 25 entries as the repository stands.
Questioning the frame found a missing category
On 20 September the file's sixth method, frame interrogation, which questions the assumptions the catalogs rest on, found something. Both failure catalogs assume a failure is the model's or the maintainer's. Neither covers the environment: tooling, harness, truncation, a script that silently does the wrong thing.
Two real incidents had been filed under a locus value, the field saying where a failure came from, called TOOLING, that existed in the vocabulary with no catalog behind it. No audit check could have found that, because every check operates inside the two-party assumption. Adding entries to either catalog would not have found it either.
What no control reaches
Beyond the unknown unknowns, the files name known limits. The reply gate's inventory has a section called not gateable: whether a judgement is right, whether something was omitted, whether research was exhaustive, whether a verdict is honest. A gate can't read intent, and absence has no token.
A reply isn't retained, so rules about replies are marked as prose by nature. The Paper Trail's own limits are listed the same way: a direct write is detected, not prevented; the audit's findings block nothing; the folder cannot delete.
The tests have blind spots of their own
The 21 September audit of the Decision Authority System (DAS) recorded that its behavioural arm had never been run, its rubric had never been applied, and its regression gate was “never evaluated”, because no production prompt existed. Every probe result since is also scoped to whatever model the harness, the setup that runs the tests, served when “sonnet” was requested, because the harness does not report it.
The file ends on a posture rather than a control: treat both catalogs as incomplete in every claim made about them. Its section on its own limit says: “The residual is real and permanent.”
Sources and notes (6)
- When RLHF Fails: a taxonomy of reward hacking — the published taxonomy the file names for its first method, noted there as not yet read
- Unknown unknowns — the notes this was written from
- Reply gate — the notes this was written from
- Behavioural failure catalogue — the notes this was written from
- DAS overview — the notes this was written from
- Directives — the notes this was written from
Authority System SuiteAll 17 parts →
