White paper ↔ DevNet alignment: delegation, encrypted test sets, resampling, commit-then-reveal, governance

[umeradl]

We've written up a working summary of the white paper with a gap/alignment checklist: Developer/design/whitepaper-summary.md (§8 is the checklist). Most items map cleanly onto the P3 design push (tracking issue #36), but a handful need a scope decision rather than just implementation.

cc @abrahamnash — as the paper's author, your call on which of these are load-bearing for DevNet 2.0 vs deliberately deferred would settle scope quickly.

Needs a decision

1. DPoS delegation (paper §8) — delegation with no delegator minimum, commissions, auto-compounding. Not implemented; current mechanism design parks it to P5+. In or out of P3 token-utility scope?
2. Encrypted test dataset + per-validator key mapping (paper §5.2.3b) — single encrypted CID, symmetric key K encrypted per-validator, mapping on-chain. The current auditor test-data flow (assignAuditTestDataset + model-owner services) should be compared against this; if it's effectively plaintext/centrally shared, do we adopt the paper's scheme in 2.0?
3. Test-set resampling (paper §5.2) — reserved ~40% pool, resampled half per round, full pool in the final round. Not visible in the current evaluation pipeline. Adopt for 2.0?
4. Commit-then-reveal between auditors (paper §5.2) — prevents validators copying each other's scores. Needs confirming as in-scope for P3 scoring hardening.
5. Non-coin-based voting credits (DPPs, paper §8) — the paper's anti-plutocratic governance stance. P3 ships only onlyOwner governance hooks; full DAO design (din-dao, Start DIN-DAO #23) should note this stance early.

For confirmation / documentation only

• Terminology: the paper's single "Validator" role is split into aggregator + auditor in DevNet — matches the paper's role-delineation rationale; we'll document the mapping in Documentation/public/.
• Random validator assignment — load-bearing for the threat model; we'll verify the current assignment path and open an issue if randomness is missing.
• Market-utility over Shapley valuation — already aligned (Developer/rejected-ideas/tknn-shapley.md); we'll cite the paper in the rationale.
• DIDs/verifiable credentials for Sybil resistance — staking-only today; noted as a future complement.
• Async FL orchestration (paper §9) — imported as requirements for the P4 dind daemon design.

[umeradl]

Thanks for the calls on these, Abraham — we're wiring them into the DIN P3 specs now.

[abrahamnash]

Thank you, @umermjd11

[umermjd11]

Status update against develop @ 87413e8 (current HEAD) — here's how each checklist item stands in code today:

Needs a decision — resolved/status:

1. DPoS delegation — Not implemented. Confirmed deferred to P5+ (per the decision this thread settled). foundry/src/DinValidatorStake.sol has no delegation/commission/auto-compounding — only flat MIN_STAKE, unbonding, blacklist, and slasher registry. Tracked in Developer/design/staking-design.md:58,81 and Developer/design/MECHANISM_DESIGN.md:65,247.

2. Encrypted test dataset + per-validator key mapping — Implemented on-chain. DINTaskAuditor.sol has encryptedTestDataKey[gi][batchId][auditor] (line 248), populated in assignAuditTestDataset (883-903) / reassignAuditTestDataset (1469-1489), plus testDataCommitments + resolveTestDataDispute for dispute binding. This is on-chain plumbing only — the actual per-validator symmetric-key encryption happens off-chain in the model-owner services, matching paper §5.2.3b's shape.

3. Test-set resampling — Implemented (off-chain, Python). cache_model_0/services/modelowner.py (_select_reserved_pool, _resample_round_pool), wired into create_audit_testDataCIDs, tests in tests/test_resampling_policy.py. Developer/design/test-set-resampling-policy.md marks it done (task_210726_6 §2c). One known gap: is_final_round isn't threaded through from dincli/cli/modelownerd/auditor_batches.py:create_testdataset, so full-pool-on-final-round is dormant (fails safe — under-exposes, never over-exposes).

4. Commit-then-reveal between auditors — Implemented on-chain. DINTaskAuditor.sol commitAuditScore/revealAuditScore (~985-1017) with auditScoreCommits/hasCommittedLM, commitHash = keccak256(abi.encodePacked(score, vote, salt)) (task_210726_6 §2a).

5. Non-coin-based voting credits (DPPs) — Deferred alongside din-dao to post-mainnet (DD-1/DD-2, Developer/design/DESIGN_DECISIONS.md). P3 ships only onlyOwner hooks (DINModelRegistry.sol: approveModel, rejectModel, setOpenSourceFee, etc.). DESIGN_DECISIONS.md:103,107 notes non-coin credits need a DID/Sybil-resistance layer that doesn't exist yet — same gap as item 9 below.

Confirmation/documentation items:

6. Validator → aggregator/auditor terminology — Confirmed; used consistently throughout Documentation/public/workflows/. No standalone "paper-to-DevNet mapping" callout exists yet, only implicit consistent usage.

7. Random validator assignment — Implemented, unweighted. createAuditorsBatches (DINTaskAuditor.sol) and autoCreateTier1AndTier2 (DINTaskCoordinator.sol) shuffle the full eligible pool; stake only gates eligibility. Stake-weighted selection was formally rejected (Developer/rejected-ideas/stake-weighted-validator-selection.md, issue P3 staking: DAO-settable floor, per-model stake requirements, concurrent-registration cap, withdrawal queue #37) to avoid whale concentration.

8. Market-utility over Shapley — Already aligned, as noted; citation lives in Developer/rejected-ideas/tknn-shapley.md.

9. DIDs/verifiable credentials for Sybil resistance — Staking-only today. Open item, blocking both no. 5 above and backlog item BL-7 (evaluator diversity monitoring).

10. Async FL orchestration (paper §9) — Not yet built; captured as design input for the P4 dind daemon (Developer/ROADMAP.md P4 section, event loop / job queue / on-chain listener). Gated on the P3-DOC7 dind architecture doc before implementation starts.