Skip to content

Latest commit

 

History

History
40 lines (33 loc) · 3.11 KB

File metadata and controls

40 lines (33 loc) · 3.11 KB

Threat model

The failure-mode files each take one attack in isolation. This is the attacker's-eye synthesis: what an adversary wants, what they can do, which defense answers it, and what residual risk survives. The mechanism is not claimed unbreakable — it is claimed legible: every residual ends at "route to a human," never at "silently assigns the wrong owner and hides it."

Attacker goals

Goal Vector Primary defense Residual
Seize a lane I don't earn self-nomination assignment is computed from merged history, not claimed none at mechanism level; social pressure on maintainers persists
Keep a lane after I stop contributing stale dominance time-decay + periodic re-derivation + opt-out (01) a contributor active just enough to refresh decay; surfaced as low-margin
Inflate my weight cheaply commit-count farming reviewed-and-merged proxy + reviewer-diversity, not raw count (02) merge/co-author is a proxy; true signal needs forge API (stated gap)
Destroy a rival's weight revert-war (06) credit is append-only; reverts net out; unjustified revert scores bare a genuine two-maintainer war; surfaced as churn, escalated to humans
Split a rival's weight identity fragmentation / drive-by (05,08) .mailmap + email-cluster; split-credit across touched lanes aliases sharing neither name nor email; rare cross-cutting refactor under-rewarded
Claim many lanes at once breadth-farming (08) per-commit credit split across its lanes legitimate wide refactor catches up over cycles, not instantly
Block a settled lane (DoS routing) manufactured "contested" (08) fails safe — routes to a human who dismisses a one-touch claim costs a human a glance
Turn the map into a target publish who owns crypto (07) computation public, output-identity redactable per-lane sensitivity weakens auditability for sensitive lanes — a stated trade, not free
Game the area boundaries granularity (04) path-glob areas v0; co-change-derived areas as research coarse areas hide sub-owners until the research tier lands

The two invariants every defense preserves

  1. Auditability — every assignment is recomputable by a third party from public history. Weakened only for 07-flagged sensitive lanes, and there only at the level of published identity, never published method.
  2. Fail-safe routing — when the mechanism is unsure (low margin, contested, manufactured conflict), it routes to a human. It never silently picks. Every residual above terminates here, which is why "the graph surfaces it, humans adjudicate" is the honest ceiling and not a dodge.

What is explicitly out of scope

  • Roadmap authority. The owner gets triage and first-review, never decision power over direction. Most governance-capture attacks are out of scope because there is no governance prize to capture.
  • Sybil identity creation at the forge level (fake accounts passing review). That is the host platform's account-integrity problem; this mechanism assumes the forge's identity layer, it does not replace it.