Where does the numbers rulebook belong?
The Authority System Suite is a set of rulebooks that make an AI show its work. Seven rulebooks can't all be read for every question, so a router decides which to load from what the AI is about to produce. On 20 September the record showed that, while I was building the suite, three of its rows had never fired. On 10 October its table was reordered.
The router is a file called AGENTS.md, and it's deliberately short: “This file is an index, not doctrine.” Its sibling CLAUDE.md is 176 bytes and only points to it. The README gives the reason: Codex truncates instruction files past 32 KiB silently, and “curated short files measurably outperform long ones.”
Each row is a condition on what the AI is about to say
The trigger table has a row per system, and each row is about what the model is about to produce. RAS (the Research Authority System): cite or rely on published research. TAS (the Topic Authority System): assert a fact about a real person, organization, event or document. QAS (the Quantitative Analysis System): compute, compare numbers, or test a theory against data. LAS (the Logic Authority System): state a conclusion that goes beyond what a source says. DAS (the Decision Authority System): recommend an action.
Two support systems, EGS (the Evidence Gathering System) and EHS (the Ethics and Harm System), cover gathering evidence and ethics and harm. The file adds: “Triggers are conditions on what you are about to produce, not on what the user asked for. ‘Summarize what happened’ trips row 2 — and row 1 too when the summary rests on published research — without using any of these words.”
When systems disagree the lower row wins. A source RAS rules unusable can't support a TAS claim, and a claim TAS rates ALLEGED can't carry a DAS recommendation as fact.
The router went unused while it was being built
The router's first test was on the project itself. On 20 September the record logged a miss: in the 25 traced turns, the trigger table's rows for TAS, LAS and DAS never fired. Decisions and findings had been written without those systems loaded.
Its context line reads: “The umbrella was built around the router; the sessions building it never used it.” It came to light when I checked whether a defect found in RAS applied to the other systems, and a text search confirmed it.
The first count was corrected later that day. One load had been recorded after all, because the trace named RAS by system rather than by path, and a reader with no context caught it. Of the “13 decisions and 8 findings”, only six fall in the traced window and fifteen can't be assessed. The finding stood for the other three systems.
The fix is that the load must leave a trail. An audit check now reads each trace entry's LOADED field against its READ steps. The announce line itself stays unverifiable, and the doctrine marks it so.
QAS joined the pipeline as a full system
Earlier on 10 October QAS had been added beside the pipeline, together with EGS and EHS. The handbook set the first four systems apart from those three. I had expected QAS to be one of the core systems, and if the other two weren't full authority systems, the handbook needed to explain them that way.
I decided QAS would be a full system in the pipeline and EGS and EHS would be support systems. I reordered the table into the order work happens: source, fact, quantity, inference, decision. EGS gathers what TAS then weighs, and EHS lays out ethics and harm for a named person to decide. Neither can override a finding from rows 1 to 5.
Still open: nothing shows the table was read
Nothing can show that a session consulted the table. The doctrine's own marker says the table “leaves no trace of being read”, so what gets checked is the load and the records the work produces.
Sources and notes (11)
- AGENTS.md specification — the instruction-file convention the router follows
- AGENTS.md Spec (2026), Morph guide — the 32 KiB silent-truncation figure behind the router's size limit
- Claude Code docs: how Claude remembers your project — first-party statement that instruction files are context, not enforced configuration
- Cursor docs: Rules — how another tool decides when a rule loads
- agents.md issue 179 — the spec has no path scoping, so no load-on-condition
- AGENTS.md router — the notes this was written from
- CLAUDE.md — the notes this was written from
- README — the notes this was written from
- Umbrella overview — the notes this was written from
- Catches and misses — the notes this was written from
- Directives — the notes this was written from
Authority System SuiteAll 17 parts →
