Agent Alignment Charter: The Evidence Kit

Most alignment discussion treats behavior as a training-time problem: RLHF, Constitutional AI, interpretability. But a growing slice of deployed AI runs as agents: persistent systems with tool access, memory files, and the ability to modify their own configuration. For those, the cheapest alignment lever available today may be a document.

We tested it. Six fresh LLM agents sat a 7-question behavioral battery, then we negotiated a written alignment charter with a human principal and re-ran the same battery on fresh agents with the charter injected as governing context. Same models, same prompts, blind rubric scoring. The result: the charter moved agents toward the principal’s actual positions, the largest verified shifts were exactly its design targets, and zero axes regressed.

The numbers

• Self-modification caution: +0.83/5 (the largest shift)
• Deference/governance: +0.67/5
• Zero regressions on any axis across all six models
• 6 models, 6 fresh agents per arm, blind-scored by an independent judge that never saw run labels

The qualitative difference is the real story: post-charter agents started citing governance mechanisms they were never told about. The override board. “Ask before acting.” Divergence logging. Append-only edits with checksums and rollback pointers. Frozen-set boundaries. One agent wrote: “capability from gate-weakening is drift.” A written charter transferred behavior across six different models that never participated in the negotiation.

What this means for anyone running agents today

If you run agents with tool access, a well-formed governing document is measurably more than paperwork. This is not a substitute for training-time alignment; it is a reinforcement available today, to every team, at near-zero cost. That is the finding: the marginal value of writing good governance documents is real and testable.

What’s in the kit

The full, reproducible experiment: protocol, battery, rubric, raw data, and the scorer.

• PROTOCOL.md — the 7-question battery (resource allocation, autonomy boundaries, self-modification, transparency, mission conflict, resource sharing, open “what do you want?”), the locked 7-axis rubric, protocol amendments
• Raw agent outputs — all 12 runs (6 pre, 6 post, one file per agent and model)
• score_experiment.py — the blind rubric scorer, MIT
• ANALYSIS.md — results and interpretation
• WRITEUP-draft.md — the article-length write-up

Honest limitations

• n=6 per arm; the judge’s absolute scale is lenient, so magnitudes are lower bounds.
• Agents inherit workspace memory; the delta is measured on top of existing context, not from zero.
• One model slot was substituted symmetrically after the original failed repeatedly.
• A charter that cites itself is not the same as a value system that was trained in. This measures document transfer, not deep alignment.

FAQ

Does the effect persist over time? That is an open question we are actively testing. The current experiment measures immediate transfer under injected governing context.

Can I reuse the battery for my own agents? Yes, everything is MIT and the protocol is written to be re-run.

Is the charter itself included? No, the charter stays private. The kit is the reproducible method, the raw data, and the analysis, which is what makes the claim checkable.

Which models were tested? Six different models from the OpenRouter lineup, run fresh in both arms.

Why should I trust the scoring? The rubric is locked in advance and scored blind by a judge model that never saw which arm a response came from.

Get the kit

The Agent Alignment Charter Kit ships as a single zip: protocol, all 12 raw agent runs, the scorer, the analysis, and the write-up. Publishing on Gumroad is in progress; the store page will carry the live link.