Skip to content

CHANGELOG.md

Latest commit

 

History

History
199 lines (141 loc) · 11.2 KB

CHANGELOG.md

File metadata and controls

199 lines (141 loc) · 11.2 KB
Raw

Changelog

All notable changes to the GRUDGE contract and the lib/chain adapter it backs. Dates are when the work landed; the contract is versioned by its on-chain storage layout (SCHEMA_VERSION).

v6.1 — Mandatory Anchored Proof (2026-07-09)

Tightens Anchored Proof from opt-in to required: create_challenge now rejects an empty proof_anchor, so every challenge registers a proof source and must verify_anchor before it can accept evidence. The "casual/unanchored" text-only proof path is gone — all evidence must link to the verified host.

⚠️ Requires a fresh deploy. No storage-layout change (SCHEMA_VERSION stays 4), but the create-time validation changed, so a redeploy makes the new rule explicit. The frontend create wizard marks the proof source required and ProofAnchor adds a copy-to-clipboard for the ownership code.

Deployed — 2026-07-09

  • GenLayer Studio: 0x6A21dc70A225dD2179F0E0EE96E0eFF7210E27f7 (schema 4, mandatory anchor). apps/web/.env.local, contracts/deployments.json, and the README deployment table all point here.
  • The prior optional-anchor v6 (0x503Cd4…9817) drops to previous in deployments.json, the README table, and the docs "previous deployments" list.

v6 — Anchored Proof (2026-07-09)

Moves the proof boundary from "validators judge submitted text" to "validators verify an owned, timestamped source" (the gap called out in the submission review): identity, account ownership, and the proof period are now enforced before/around the LLM judgment instead of being left to it.

⚠️ Requires a fresh deploy. Challenge gained two fields (proof_anchor: str, anchor_verified: bool) and create_challenge gained a parameter. SCHEMA_VERSION bumped 3 → 4. After deploying, update NEXT_PUBLIC_GRUDGE_CONTRACT_ADDRESS.

Deployed — 2026-07-09

  • GenLayer Studio: 0x503Cd4D2f88520c1f8a6455cC958199508789817 (schema 4). apps/web/.env.local, contracts/deployments.json, and the README deployment table all point here.
  • Previous deployments stay referenced (README + deployments.json + commented .env.local block) so the team can re-test earlier transactions: Studio v5 0xb9b501…379c, Bradbury 0x652789…80C6 / 0xaba1Db…819C.
  • New apps/web/scripts/seed-studionet.mjs seeds the fresh contract with demo transactions (funded throwaway accounts → create_challenge → stakes) so the ledger isn't empty on first load.

Frontend integration (landed 2026-07-09, commit 87f3d32)

  • GrudgeClient gained getAnchorInfo / verifyAnchor (genlayer + mock adapters), with AnchorInfoSchema in types.ts and useAnchorInfo / useVerifyAnchor hooks.
  • New ProofAnchor component: anchor badge + the creator's verify flow (shows the grudge-<id>-<addr> code to paste, fires verify_anchor, which is on the slow AI-writes poll path since it runs a web-fetch consensus round).
  • Create wizard: optional "Proof source URL" field wired through create_challenge's new proof_anchor arg; EvidenceTribunal blocks evidence submission until an anchored grudge is verified.
  • judgePrompt.ts mirrors the contract's new TIME WINDOW / ANCHORED PROOF prompt blocks so the mock judge matches on-chain behavior; mock client tests extended to cover the anchor origin gate.

Added — contract methods (ABI deltas)

Method Kind Signature → returns
verify_anchor write (web-fetch consensus) (challenge_id: int) -> str → {"anchor","code","verified"}
get_anchor_code view (challenge_id: int) -> str → {"anchor","code","verified"}

Contract method count: 17 → 19 (views 10 → 11, writes 7 → 8).

Changed — contract

  • create_challenge(…, proof_anchor: str) — new trailing parameter. "" = unanchored (previous behavior); a non-empty value must be one http(s) URL ≤ 200 chars and registers the challenge's proof source.
  • verify_anchor — ownership proof: the validator set fetches the anchor page and checks a challenge-bound code (grudge-<id>-<creator-addr-prefix>) appears on it, via gl.eq_principle.strict_eq (deterministic containment on non-deterministically fetched content — no LLM round).
  • submit_evidence — deterministic pre-consensus gates on anchored challenges: anchor must be verified, evidence must contain ≥1 link, and every link's host must equal the anchor host. The judge prompt gains a TIME WINDOW block (this proof period's pinned start/end) and, when anchored, an ANCHORED PROOF block requiring the verdict to rest on the fetched page — an unfetchable anchored link is deterministically REJECTED.
  • get_challenge — + proof_anchor, + anchor_verified; get_challenges_page summaries — + anchor_verified.

Added — web

  • GrudgeClient.getAnchorInfo / GrudgeClient.verifyAnchor (genlayer + mock; the mock mirrors the origin gate).
  • ProofAnchor component: anchor badge on the challenge page + the creator's verification flow (shows the code to paste, fires verify_anchor).
  • Create wizard: optional "Proof source URL" field; the tribunal blocks evidence submission until an anchored grudge is verified.

Deliberate scope note: the anchor proves account ownership, not human truth — the defensible claim is "evidence must come from an owned, platform-timestamped source," not "cheating is impossible."

v5 — GenLayer-native features (2026-06-15)

Four high-impact features on top of the hardened contract, built without weakening any existing security property: every new payable path updates total_locked and calls _assert_solvent(); every new LLM call lives in a zero-arg fn inside gl.eq_principle.prompt_comparative with sort_keys=True; every new view is bounded.

⚠️ Requires a fresh deploy. Storage layout changed (Evidence gained two fields; a new reputation map). SCHEMA_VERSION bumped 2 → 3. studionet history resets, so this is a clean redeploy — there is no in-place migration. After deploying, update NEXT_PUBLIC_GRUDGE_CONTRACT_ADDRESS.

Added — contract methods (ABI deltas)

Method Kind Signature → returns
appeal_verdict write · payable (challenge_id: int, evidence_index: int) -> str → {"verdict","reason","confidence"}
get_reputation view (address: str) -> str → reputation object (below)
explain_verdict view (runs consensus) (challenge_id: int, evidence_index: int) -> str → {"verdict","explanation"} (explanation ≤ 600 chars)
suggest_evidence_policy view (runs consensus) (statement: str) -> str → {"policy","rationale"} (each ≤ 280 chars)

Contract method count: 13 → 17 (views 7 → 10, writes 6 → 7).

Changed — contract

  • create_challenge(statement, evidence_policy, category, duration_days, required_proofs) — same signature, new behavior: an empty evidence_policy is now allowed and triggers AI policy design (the contract fills it via suggest_evidence_policy). A non-empty policy still enforces the 4–280 char bound.
  • get_challenge / get_challenges_page's nested evidence — each entry gains two fields: appealed: bool and appeal_bond: str (wei).
  • get_evidence_page — same two new fields per entry.

Added — storage (layout change → SCHEMA_VERSION 3)

  • Evidence dataclass +2 fields: appealed: bool, appeal_bond: u256 (appended; existing fields unmoved).
  • New reputation: TreeMap[Address, Reputation] map.
  • New Reputation dataclass: challenges_created, challenges_won, proofs_verified, proofs_rejected, doubts_made, doubts_correct (all u32).

Added — constants

  • SCHEMA_VERSION = 3
  • MIN_APPEAL_BOND = 10**17 (0.1 GEN)
  • CONVICTION_DAMPENER = 3

Feature 1 — Appeals

When a proof is REJECTED, the creator can appeal with a bond. A fresh consensus round re-judges the evidence with an appeal notice prepended.

  • Flip → VERIFIED: the proof counts (verified_count += 1), reputation moves rejected→verified, and the bond is returned to the appellant (credited to claimable).
  • Upheld: the bond is forfeited to the doubter pool.
  • One appeal per entry; bond >= MIN_APPEAL_BOND; creator-only.
  • Solvency: the bond is held like any stake (total_locked += bond on deposit, then only moves between buckets), so conservation holds — proven by test_appeal_conservation.

Feature 2 — AI-designed evidence policy

The contract's LLM proposes a concrete, anti-gaming evidence policy from the statement. Used two ways: the suggest_evidence_policy view (preview, creates nothing) and the blank-policy autofill branch in create_challenge.

Feature 3 — Explain verdict

explain_verdict returns the referee's full reasoning for an existing verdict via consensus, without re-judging or mutating state (verified by test_explain_does_not_mutate). Output bounded to 600 chars.

Feature 4 — Conviction rating

On-chain reputation derived from kept/broken/doubt history. Counters update in settle (creator created/won; each distinct doubter made/correct), submit_evidence, and dispute_evidence. get_reputation returns the raw counters plus two deterministic 0–100 scores (no LLM):

{
  "address": "0x…",
  "challenges_created": 0, "challenges_won": 0,
  "proofs_verified": 0, "proofs_rejected": 0,
  "doubts_made": 0, "doubts_correct": 0,
  "conviction_score": 0,   // 100·won/(created+3) — dampened so 1/1 ≠ 100
  "doubter_accuracy": 0     // 100·correct/(made+3)
}

Frontend (apps/web/lib/chain + UI)

New GrudgeClient methods (implemented in both genlayer.ts and mock.ts): getReputation, explainVerdict, suggestPolicy, appealVerdict.

New Zod schemas: ReputationSchema; EvidenceEntrySchema extended with appealed + appealBond.

New hooks: useReputation, useExplainVerdict, useSuggestPolicy, useAppealVerdict.

New / updated UI:

  • ConvictionBadge — on profile, challenge creator, and each doubter.
  • ExplainVerdict — "Explain this verdict" disclosure with a typewriter reveal.
  • AppealAction — bond modal on rejected entries (creator only).
  • Create wizard — "✨ Suggest a fair policy" button + rationale preview; the policy field may be left blank to let the AI design one.

Tests

  • Contract: +18 Direct Mode cases (41 total) — F4 (5), F3 (3), F2 (4), F1 (6), all run the real contract in-memory with the LLM mocked.
  • Web: 35 vitest cases (mock client extended for all four features).
  • Gates green: ruff --select ALL, mypy --strict, genvm_lint, genvm-lint check/typecheck, the wallet-lib guard.

v1–v4 — hardened contract baseline

The contract before v5: paginated bounded views (no unbounded reads), total_locked solvency bookkeeping + _assert_solvent(), the claiming reentrancy guard on claim(), and deterministic time via gl.message_raw["datetime"]. Public surface: create_challenge, stake, submit_evidence, dispute_evidence, settle, claim, and the get_challenge* / get_*_page / get_claimable / get_solvency views.