Charter summary: can a model shorten the rules every agent session loads?

A product charter is a ledger of settled decisions: what a product does, what was rejected and why. Rather than load the whole ledger at every session start (43K–112K characters for the ones tested here), coding agents load a short summary. If the summary drops a condition or an exception, an agent that follows it breaks a rule while believing it is obeying one. We tested three ways to write that summary on three real charters.

Run October 27, 2026 · 3 charters · 24 summaries · flat-rate plans, no metered spend

The charters

• E-commerce site charter
  134 settled decisions, 11 rejected options
• Platform ops charter
  84 settled decisions, 13 rejected options
• Construction takeoff charter
  33 settled decisions, 1 rejected option

Round 1: the model rewrites

Each model was asked to compress the whole ledger into a summary of at most 3,000 characters: pick the critical decisions, keep every condition, and list every rejected and proposed id.

Round 2: the model ranks, the tool pastes

The model only ranks the settled decisions, most critical first. The tool then pastes each decision in its exact ledger wording until a 4,500-character budget is full, and lists rejected and proposed options by the start of their text.

Round 3: a second model reviews the ranking

Sonnet ranks as in round 2. Antigravity (Gemini 3.8 Flash high) then reads that summary against the full ledger and suggests swaps (a more critical decision in place of a less critical one) and stale flags (a line a later decision reverses or narrows). Sonnet re-ranks once with that feedback, and the tool pastes verbatim, adding a pointer to the newer decision on a flagged line only if that decision exists.

Averages over the 3 charters, one run each. A HIGH distortion is a line that would make an agent break a rule if it followed it: a dropped condition, exception, branch or limit, or a wrong fact. LOW distortions are the judge's lesser category. The judge score (0–10, higher is better) rates each summary as the only charter context an agent gets.

What it found

• Rewriting was not safe at this size. Every one of the 12 round-1 summaries carried 1–6 HIGH distortions and missed 2–21 decisions the judge rated critical.
• Every model kept the ids honest: none invented, and every rejected and proposed id listed. Only 1 of 12 round-1 summaries stayed within 3,000 characters.
• Ranking and pasting verbatim cut HIGH distortions to 0.0–0.7 per summary on average (with a larger budget, see below), against 3.0–4.7 when the model rewrote. The judge placed all 3 that remained in the shortened rejected-option lines, not in the pasted decisions: 2 listed a rejection that a newer decision had since reversed, and 1 was cut off mid-sentence.
• The cost moved to coverage. Only 14–19 decisions fit the 4,500-character budget, and the judge still found 8–13 critical decisions missing from each summary.
• The 3 rankers were within 0.7 HIGH distortions, 1.3 missed decisions and 0.7 judge points of each other. With 3 runs each, this run does not rank them.
• In this run, a review by a model from another family helped. Against Sonnet ranking alone, HIGH distortions went from 0.3 to 0.0 per summary and missed critical decisions from 11.0 to 8.3 (fewer on 2 of 3 charters), with the judge score at 6.3 against 6.0. Each charter took 2–3 minutes, acceptable for a job that runs only when a charter changes.

What we take from it

Let a model choose, never rewrite. In this run, ranking plus verbatim text carried far fewer HIGH distortions than any rewrite, and with the review step none at all. The pipeline the run points to: Sonnet ranks, Antigravity reviews the ranking, Sonnet re-ranks, and the tool pastes verbatim. It still leaves critical decisions out, so they need another route to the agent, such as being looked up when it edits the files they govern, and rejections that a later decision reversed need to be marked.