Skip to content

feat(ui): add a miner-vs-human breakdown to the Wave-3 gate-precision / reversal-rate / findings-by-category dashboard cards #4520

Description

@JSONbored

Context

A maintainer review-stack × Autonomous Miner System (AMS) integration audit (2026-07-09) found that the Wave-3 maintainer analytics dashboard (29 planned cards, 19 MCP tools) has zero miner-vs-human split anywhere in its scope, and the underlying gate-eval engine is structurally incapable of one today.

Evidence

  • Issue 🧩 Wave 3 — Dashboard UI & MCP (HELD tracker) #2013's body lists all 29 dashboard cards and 19 MCP tools verbatim — none mention miner/contributor-origin.
  • src/review/parity.ts (computeGateEval) selects only project/target_id/decision from review_audit — no actor/login column.
  • migrations/0049_review_audit_parity.sql states outright: "Privacy: project + target_id + head_sha + decision + reasonCode + timestamp ONLY. No actor logins" — confirming review_audit, and therefore every stats.ts/GateEvalReport aggregate keyed off it, has zero actor-origin signal, not even a boolean flag.
  • The one login-keyed substrate that does exist, contributor_gate_history (migrations/0126, src/review/contributor-calibration.ts), is explicitly scoped only to feed a future internal per-actor confidence adjustment (maintainer: personalized gate-prediction tuning per contributor/miner history #2349) — its own design note bars it from ever appearing on a public dashboard or any contributor-facing output, and it isn't wired to the Wave-3 dashboard/MCP surface regardless.

Why this matters for the Autonomous Miner System

Gittensory's own miner fleet (AMS) will be a fast-growing PR source for repos that adopt it. If none of the Wave-3 cards — gate-precision, reversal-rate, findings-by-category, acceptance-rate, gate-outcome breakdown, top-contributors-by-quality-band — can be sliced by contributor origin, a maintainer running a mixed human+miner repo cannot tell whether the gate is systematically mis-calibrated for one cohort, and neither #2348's value-weighted calibration nor #2350's anti-farming detector can be validated per-cohort from the maintainer's own dashboard. Note this is distinct from #2352 (which wires a source-scoped computeGateEval pass into the auto-tune circuit-breaker specifically) — that's an automated enforcement action, not a maintainer-visible dashboard view of the same split.

Deliverables

  • Add a privacy-safe, aggregate-only "miner-originated" dimension to the tables/aggregates the Wave-3 dashboard cards read from (without reintroducing actor logins into review_audit itself — e.g. a boolean/cohort tag rather than identity)
  • Extend at minimum the already-shipped gate-precision card (feat(ui): gate-precision analytics card (precision + confusion matrix) #2191) and reversal-rate/findings-by-category cards with an optional miner-vs-human split view
  • Tests: cohort split reflects real data without leaking actor identity; existing blended aggregates unaffected when the split isn't requested
  • Docs: note the new dimension in the dashboard/analytics documentation

Surfaced by a maintainer review-stack × AMS integration audit (2026-07-09). Related: #2013 (Wave-3 Dashboard UI & MCP tracker, which this extends), #1967 (maintainer analytics dashboard umbrella), #2352 (miner-source-scoped auto-tune circuit-breaker — an automated consumer of a similar split, not a dashboard view), #4265 (the self-improve design note that already draws the audience/data-source distinction this issue's dashboard view needs to respect).

Metadata

Metadata

Assignees

Labels

maintainer-onlyOwner-only work — yields no Gittensor points.orbGittensory Orb related - maintainer self-hosting analytics.

Projects

Status
Done

Relationships

None yet

Development

No branches or pull requests

Issue actions