Guardrails¶
DragonPay's guardrails are the specific, checkable rules Claude evaluates
itself against while planning and implementing — the mechanism ENG-02 §19
defines; the Guardrail Catalog
is the sole authoritative source for each individual rule's exact text.
This page is a concise index, not a substitute for it — read the Catalog
for the actual Rule, Applies When, and Evaluation Check text before
relying on any guardrail's behavior.
Each entry has an Applies To (PM, ENG, ADMIN, ALL, or a
combination) and an Enforcement Type (EVALUATIVE — requires model/human
judgment — or BOTH, a partial deterministic backstop). No guardrail is
purely deterministic yet.
What the catalog currently covers¶
ENG Mode planning integrity (GR-001–GR-013) — the largest group:
not implementing later-story scope, disclosing material assumptions,
treating examples as illustrative rather than authoritative, staying
consistent with the existing model, truthful default/absence handling,
state/evidence consistency, semantic sufficiency, internally consistent
plans, prerequisite integrity, scoping blocking to material risk, not
treating plan approval as license for new assumptions, scenario
viability, and respecting mode authority boundaries.
Truthful reporting (GR-014) — applies across modes; the basis for
this site's own "never invent progress/completion data" convention.
Product change discipline (GR-015–GR-021) — the PM Mode side:
scoping Product changes tightly, epic consistency, cross-requirement
conflict detection, requirement ownership/duplication, preserving
released history, not leaking implementation detail into Product text,
and acceptance-criteria completeness. This group is what
ENG02-23's Product Change Review runs
against.
Mandatory control integration (GR-022) — the newest entry: a
control isn't real if it can be silently bypassed.
How guardrails get used¶
- PM Mode reads the Catalog directly against the Product repository it's authoring.
- ENG Mode / Admin Mode read it through the pinned Product baseline, never live Product HEAD — the same invariant that applies to every other Product artifact those modes consume.
- Adding or refining a guardrail is normally a PM Mode change to the Catalog itself, not a change to ENG-02 — as long as the framework already supports the guardrail's shape.