Intent layer Not a training pitch
Agents get smarter.
Your goals shouldn’t get fuzzier.
Charter locks what you want and checks whether the work matched. Human intent stays the source of truth — recorded, enforced, verified.
Why it matters
Every capable agent introduces the same failure mode: the ask drifts. Soft agreements, polite “done,” and clever detours that weren’t what you meant. More model power doesn’t fix that — it amplifies it.
- The ask is the asset. If intent isn’t recorded, nothing downstream can be honest.
- Enforcement beats vibes. Constraints belong next to the work, not in a forgotten prompt.
- Verification closes the loop. Did the outcome match the charter — yes, no, or drifted?
1 · Record
Capture the charter: goal, musts, must-nots, success checks. Plain language. Durable. Shareable with humans and agents.
2 · Enforce
Keep the charter in the loop while work runs. Scope, constraints, and stop conditions stay visible — not optional flavor text.
3 · Verify
Check outcomes against the original ask. Pass, fail, or drift — with a trail you can audit. Verification under the hood; intent on the surface.
Makes sense today — and with ASI
Today
Coding agents, ops bots, and copilots already rewrite the brief mid-flight. Charter is the sticky note that doesn’t get overwritten: what was asked, what was allowed, what “done” means.
As capability scales
When agents can do almost anything, the scarce resource isn’t cleverness — it’s a durable record of intent. Charter stays the same product: lock the goal, check the match. The smarter the agent, the more you need the lock.
Free today
Charter treats human intent as the source of truth: record what was asked, then verify whether the work matched. Free today: held-out verification tasks that dry-run offline, plus open preference-quality research artifacts on Hugging Face that power the engine underneath. Agents will get smarter; goals shouldn’t get fuzzier. Private charter audits — hudson.gouge@projxon.ai.
Hub artifacts (engine underneath)
- Strong preference split — v2.2 · 386 train · 59 eval
- Preference-quality adapter — Hub default = strong-scale
- Parent mined pairs
Free pack is a credibility surface — not a paid SKU. Headline stays intent; Hub cards stay honest about methods.
What Charter is not
- Not a DPO-dataset / training-stack public pitch.
- Not “Datadog for agents” or preference-split tooling as the product story.
- Not “trust the model.” Trust the charter, then verify.
Day-one shape (under consideration)
Early MVP direction we’re weighing:
charter.yml in the repo plus GitHub PR status checks that verify work against the recorded ask.
This is a design note — not a claim that a shipped GitHub App is live today.
Charter Pro — $49/mo
Private charters + verification reports. Billing via Kestrel Ops. Product: Charter.
$49 / month
Stripe Checkout — connect key to go live (no PayPal).
When the billing host is up, point this CTA at
{PUBLIC_BASE_URL}/subscribe?product=charter.
Waitlist
Free-tier notes + Pro interest. Signups log as NDJSON in-browser and open mailto to the Kestrel Ops inbox.
Thesis (hard bar)
Postable so everyone gets it. Makes as much sense with ASI as today. Intent is the product surface; verification is the engine underneath — not the marketing story.