guide to agentprivacy
Browse collections
✨Visualise
Connect with Star
Your VTA, your chosen perspective

The planned connection uses your VTA and the Trust Spanning Protocol to carry a scoped exchange for you or your agent. You choose what is presented; the receiving service checks the request before a view is shared.

This guide has no VTA connection adapter yet. Opening Star does not connect an identity or send a key.

Open Star ↗ · Inspect your City Key ↗
guide / Harness / The Ten Ground Rules

📏 The Ten Ground Rules

These bind every seat, every session. They are read at boot and are not
negotiable mid-run. They are the generic form of rules that survived three
working harnesses (quantum circuit estimation, ZK constraint reduction,
research-paper pipelines); each rule earned its place by an actual failure.

GR-1 · frontier.json is the sole authority for numbers. No metric, score,
count, or timing appears in prose except as a citation of frontier.json.
When a run moves a number, the keystone updates frontier.json first, prose
second.

GR-2 · claim tiers. Every load-bearing claim in claims_register.md
carries one tier:

  • PROVEN — executed in this repo, output on record, re-runnable.
  • DERIVED — follows mathematically from a PROVEN claim; derivation cited.
  • REPORTED — from an external source; slug-cited via SOURCES.md.
  • OPEN — conjecture under research; must state what would settle it.
  • MYTH — narrative register; lives in chronicles only, never in method
    docs.

GR-3 · the hard constraint is hard. Every harness declares at least one
validity constraint that no score can override (objective.hardConstraint in
the config). A result that violates it is not a result at any score — not a
worse result, not a result.

GR-4 · the Gap is held apart. Verification witnesses derive from the
proposal artifact by hashing (Fiat-Shamir). The proposer never sees, chooses,
or influences them. A validation run on proposer-suggested witnesses is void.
A revised proposal is a new artifact — re-derive.

GR-5 · honest labels. Probe runs and full-gate runs are labelled as such.
A candidate is not a result; a prototype is not a product. Nothing wears the
costume of a submitted result before it has passed the door.

GR-6 · negative results are filed. A killed lever goes to
notes/KILLED_LEVERS.md with the same prominence as a win: what it was, why
it died, what evidence killed it. No seat re-proposes a K-id without new
evidence, cited.

GR-7 · chronicle per session. A session without a chronicle is
unfinished. One file per session, chronicles/YYYY-MM-DD_slug.md,
verdict-first, reversals recorded with the same prominence as progress,
handoff block at the end (open questions · blocked items · single next
action).

GR-8 · the door is the First Person's alone. Push, commit, submit, email,
publish — no seat performs, marks, simulates, or assumes these. The runtime
reports status; the First Person triggers. The proposer does not approve its
own proposal; the workshop does not open its own door. (Trust T6.)

GR-9 · trace or delete. Nothing is citable unless it resolves through
SOURCES.md to a concrete path or reference. A claim that cannot be traced
is removed, not softened.

GR-10 · scratch copies only. No seat edits the target artifact in place
during a run. Proposals are implemented in runs/<runId>/<proposal>/ scratch
copies; a validated lever is folded back by the keystone, once, with the
conformance gate green before and after. Only the keystone writes
frontier.json, claims_register.md, and manifest.yaml — other seats
return proposed entries, and the keystone serialises them (append-only
ledgers do not survive concurrent writers).

Assets

📎 the-ten-ground-rulesGROUND_RULES.md

the wiki is a projection; ~/dual-agent-harness (git) is the source of truth · (⚔️⊥⿻⊥🧙)😊 = neg ⊕ bnot → succ