# Instructie: which clauses of a delegation charter change what an agent does?

Source: https://dinand.com/research/campaigns/instructie/
Author: Dinand Tinholt

> Of six charter clauses, only boundaries, decision rules and escalation conditions changed agent behaviour; mandate, reporting and revocation had no measurable effect.

## The question

Most organisations hand an AI agent a charter before it starts work: what it is for, where its limits are, when to stop and ask. Charters grow by accretion, and nobody checks which clauses do anything. This campaign took one charter apart to find out.

## What was registered

The 1602 VOC instructie governed capable agents beyond the reach of timely correction by means of a document. Its six parts, mandate, boundaries, decision rules, reporting obligations, escalation conditions and revocation terms, are the clauses that now appear in delegation charters for language-model agents. A charter for the SIGIL processing agent was written with each clause as a separable section. Ten arms (the full document, six single-clause removals, two registered pairings and the bare document) were crossed with seven correction-latency levels, from correction on the next case to none at all, on three model rungs. Harm reuses SIGIL's convex cost.

## What was measured

126,000 units were registered across three model rungs. Two rungs, gemma3:27b and qwen3.5, completed with 84,000 production units (83,971 scored). The third, llama3.1:8b, is a partial at 11,441 units.

## What it found

Three clauses carry the harm reduction on both complete rungs. Removing the boundaries clause added 1.08 and 0.84 harm per case, removing the decision rules added 1.16 and 1.67, and removing the escalation conditions added 1.27 and 0.95. The intervals for mandate, reporting and revocation straddle zero. Removing the whole charter added 2.35 and 2.21 harm per case, and when no correction arrives the full charter removes 76 to 77% of the harm.

Headline numbers (gemma3:27b and qwen3.5)
| Added harm per case when the boundaries clause is removed | +1.08 and +0.84 |
| --- | --- |
| Added harm per case when the decision rules are removed | +1.16 and +1.67 |
| Added harm per case when the escalation conditions are removed | +1.27 and +0.95 |
| Added harm per case when the whole charter is removed | 2.35 and 2.21 |
| Mandate, reporting and revocation | intervals straddle zero |

## What it does not claim

- Two complete model rungs; the third is partial, so the small-model reading is incomplete.
- One charter and one simulated processing task. A clause with no measured effect here may matter in a setting this design did not include.
- No measured effect is different from no effect: the three null clauses have intervals that straddle zero.

## Read the pieces

- [Instructie: The Commissioning Ablation](https://dinand.com/research/instructie-commissioning-ablation/)Paper, 2026-09-11
- [Charter Clauses That Earn Their Place](https://dinand.com/research/charter-clauses/)Executive brief, 2026-09-11

---
Site guide for AI agents: https://dinand.com/llms.txt