Knowledge-Free Correlated Agreement (KFCA)

• KFCA is a peer-prediction mechanism for federated learning that rewards client contributions using pure report agreement without any ground truth labels.
• It employs a bonus-minus-penalty construction on categorical reports, ensuring truthful reporting under the categorical-world condition and an honest-majority threshold.
• KFCA is designed for scalability and decentralized deployment, addressing issues like label-flipping and enabling cost-effective blockchain implementation.

Knowledge-Free Correlated Agreement (KFCA) is a multi-task peer-prediction mechanism for federated learning (FL) that rewards client contributions without access to ground truth, without a public labeled test set, and without any knowledge or estimation of report distributions. Introduced as an incentive layer for FL, KFCA instantiates the bonus-minus-penalty template of multi-task peer prediction with a simple agreement score on categorical reports. Under categorical reports, the categorical-world condition, and an honest-majority threshold, it is strictly truthful, and it is presented as a remedy to the label-flipping vulnerability of Correlated Agreement (CA) while remaining suitable for decentralized and blockchain-based deployment (Witt et al., 6 May 2026).

1. Formal mechanism and score structure

KFCA is defined on a collection of tasks $k \in M = \{1,\dots,m\}$, each with latent truth $Y^k \in [L] = \{1,\dots,L\}$. A client $i$ derives a private signal $Z_i^k \in [L]$ after optional effort $e_i^k \in \{0,1\}$. Under effort, the informative channel is

$$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$

with diagonal dominance $P_i(y \mid y) > P_i(a \mid y)$ for $a \neq y$. Under no effort, the uninformative baseline is

$P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$

independent of $y$. Reports are $Y^k \in [L] = \{1,\dots,L\}$0, generated by a reporting strategy $Y^k \in [L] = \{1,\dots,L\}$1. Because expected payoffs are linear in $Y^k \in [L] = \{1,\dots,L\}$2, the mechanism restricts attention to deterministic mappings $Y^k \in [L] = \{1,\dots,L\}$3 (Witt et al., 6 May 2026).

The peer-prediction backbone is multi-task peer prediction (MTPP). Given a score matrix $Y^k \in [L] = \{1,\dots,L\}$4, two clients are compared on shared tasks through the Dasgupta–Ghosh bonus-minus-penalty construction. Shared tasks are partitioned into disjoint $Y^k \in [L] = \{1,\dots,L\}$5 (bonus), $Y^k \in [L] = \{1,\dots,L\}$6, and $Y^k \in [L] = \{1,\dots,L\}$7. For each $Y^k \in [L] = \{1,\dots,L\}$8, one draws $Y^k \in [L] = \{1,\dots,L\}$9 and $i$0 uniformly at random and pays

$i$1

The first term rewards contemporaneous agreement; the second corrects for chance agreement.

The mechanism’s correlation object is the delta matrix

$i$2

with $i$3. Under MTPP, expected payment to $i$4 is

$i$5

KFCA sets the score matrix to pure agreement,

$i$6

For deterministic strategies,

$i$7

The per-task payment becomes

$i$8

In this formulation, “knowledge-free” has a precise meaning: KFCA uses only peer agreement on categorical reports and needs neither ground-truth labels, nor a labeled test set, nor any estimation of $i$9 or its sign pattern. Agreement is scored directly and corrected for chance through the MTPP penalty term.

2. Assumptions, categorical-world condition, and truthfulness

KFCA is analyzed under ex-ante identical tasks that are i.i.d. across $Z_i^k \in [L]$0, conditional independence of signals given $Z_i^k \in [L]$1,

$Z_i^k \in [L]$2

informative signals under effort, uninformative signals under no effort, and risk-neutral agents whose utility is expected payment minus effort cost (Witt et al., 6 May 2026).

The central structural assumption is the categorical-world condition: $Z_i^k \in [L]$3 Its interpretation is that truthful signals are positively correlated on the same label and negatively correlated across different labels. The source states that this holds broadly in classification and in FL sign-quantized update directions. Under partial effort, $Z_i^k \in [L]$4 scales by $Z_i^k \in [L]$5, where $Z_i^k \in [L]$6 is an effort probability, while preserving its signs.

Truthfulness is formalized through informed truthfulness: $Z_i^k \in [L]$7 where $Z_i^k \in [L]$8 is truthful reporting. Under the categorical-world condition, KFCA satisfies a stronger statement: the truthful profile maximizes expected reward, with equality only when both agents apply a common bijection $Z_i^k \in [L]$9. The residual ambiguity is the standard shared-permutation indistinguishability of peer prediction without ground truth.

The mechanism’s practical resolution of that ambiguity is an honest-majority condition. Let $e_i^k \in \{0,1\}$0 be the fraction of adversarial or malicious clients. The guarantees are asserted when $e_i^k \in \{0,1\}$1, described as the standard Byzantine threshold for peer prediction without ground truth. In the binary case with uniform prior and symmetric honest noise $e_i^k \in \{0,1\}$2, the paper gives

$e_i^k \in \{0,1\}$3

for $e_i^k \in \{0,1\}$4 and $e_i^k \in \{0,1\}$5. This makes explicit that a coordinated label flip is strictly dominated by honest reporting when the honest majority is preserved.

A common misconception is that “strict truthfulness” eliminates all relabeling ambiguities. KFCA does not claim that. Without ground truth, common permutations remain indistinguishable at the mechanism level; what changes under honest majority is that any minority deviation from the prevailing convention loses reward.

3. Relation to Correlated Agreement and other peer-prediction mechanisms

Correlated Agreement uses the $e_i^k \in \{0,1\}$6-dependent score

$e_i^k \in \{0,1\}$7

with deterministic expected reward

$e_i^k \in \{0,1\}$8

Because this score depends only on the sign pattern of $e_i^k \in \{0,1\}$9, any global label permutation that preserves that sign pattern receives the same payoff as honesty. The paper’s binary worked example states

$$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$0

which is the label-flipping vulnerability that KFCA is designed to address (Witt et al., 6 May 2026).

The source attributes this vulnerability to CA’s indifference to label semantics. CA rewards correlation “above chance,” but global relabelings preserve the sign of $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$1, so coordinated flipping cannot be penalized by the score itself. KFCA replaces this with literal report agreement. Under the categorical-world sign structure, agreement is informative because agreement on the same underlying label contributes positive $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$2, while mapped agreement across different underlying labels accumulates negative $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$3.

The comparison with adjacent peer-prediction mechanisms is operational as well as conceptual. Bayesian Truth Serum variants require eliciting reports and predictions of others’ reports, hence some form of distributional knowledge or priors. CA requires estimating $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$4, or at least its sign pattern, from all reports each round, with complexity $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$5 and a centralized computation path. KFCA requires no priors, no $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$6, and no labels; its score is computed directly from pairwise report agreements corrected for chance. The paper therefore characterizes it as distribution-free and suitable for decentralized or on-chain settings.

4. Federated learning instantiations and round protocol

KFCA is specialized to FL through two categorical reporting schemes (Witt et al., 6 May 2026).

Instantiation	Task definition	Report
KFCA-D	Shared unlabeled public input $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$7	Predicted label $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$8
KFCA-QP	Parameter coordinate $$P(Z_i^k = a \mid Y^k = y, e_i^k=1) = P_i(a \mid y),$$9	$P_i(y \mid y) > P_i(a \mid y)$0

KFCA-D applies when public unlabeled inputs are available. Each task is an input $P_i(y \mid y) > P_i(a \mid y)$1, and the mechanism compares only clients’ predicted labels; no ground-truth labels are used. KFCA-QP applies when no public inputs are available. Each model parameter coordinate is treated as a task, and each client reports the 1-bit sign of its local update. The paper also describes a broader pipeline that maps continuous or heterogeneous outputs into categorical reports $P_i(y \mid y) > P_i(a \mid y)$2 so that the induced $P_i(y \mid y) > P_i(a \mid y)$3 has the required sign structure.

The end-to-end per-round protocol is explicit. Each client $P_i(y \mid y) > P_i(a \mid y)$4 in round $P_i(y \mid y) > P_i(a \mid y)$5 submits categorical reports $P_i(y \mid y) > P_i(a \mid y)$6, either labels on shared unlabeled inputs or sign-quantized update coordinates. The mechanism randomly samples $P_i(y \mid y) > P_i(a \mid y)$7 peers per client, samples disjoint $P_i(y \mid y) > P_i(a \mid y)$8, $P_i(y \mid y) > P_i(a \mid y)$9, and $a \neq y$0, and for each $a \neq y$1 draws $a \neq y$2, $a \neq y$3. The round reward is

$a \neq y$4

Training itself proceeds normally. The paper’s “KFCA-FedAvg” computes $a \neq y$5 and then applies standard FedAvg: $a \neq y$6 The rewards are computed independently of aggregation and can be used for payouts, logging, or optional weighting. The paper is explicit that it evaluates KFCA as an incentive layer rather than as a robust aggregation rule.

Algorithmically, reward computation is $a \neq y$7 per client per round, plus $a \neq y$8 sampling of $a \neq y$9. Overall per-round cost is $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$0 for $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$1 clients, which is linear in $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$2 for fixed $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$3 and $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$4. If $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$5 is large, $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$6 can be sampled for real-time scoring, and the standard error decays as $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$7. This linear scaling is a central practical distinction from CA’s quadratic dependence on client count through $P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$8 estimation.

5. Decentralized and blockchain deployment

The mechanism is presented as compatible with decentralized deployment because it requires only per-round matching and counting rather than global distribution estimation. The implementation model is divided into off-chain and on-chain components. Off-chain operations include local training, report generation, and storage of reports, for example through IPFS. On-chain operations include commit–reveal, unbiased matching through a verifiable random function (VRF), index sampling, and payment computation (Witt et al., 6 May 2026).

The commit–reveal procedure is specified as follows. A client posts

$P(Z_i^k = a \mid Y^k=y, e_i^k=0) = Q_i(a),$9

on-chain, where $y$0 is an off-chain pointer and $y$1 is a salt. Later the client reveals $y$2, and the contract verifies the hash before scoring. Random pairing uses an on-chain VRF or a trusted randomness beacon to select peers and sample the sets $y$3, $y$4, and $y$5. Scoring then fetches the needed report entries, through oracles or data-availability proofs, computes the KFCA reward on the selected pairs and tasks, and releases payments automatically.

The stated gas and latency profile depends on $y$6. On-chain work per payment is $y$7, so the source recommends small $y$8, for example dozens or hundreds of indices, and limiting peers per round, for example $y$9–$Y^k \in [L] = \{1,\dots,L\}$00. For very large task sets, such as adapter parameters, the paper suggests sampling and/or off-chain verifiable computation with succinct proofs such as zk-SNARKs to attest to agreement counts.

This architecture yields what the source describes as one-shot, distribution-free scoring. A plausible implication is that KFCA’s incentive logic is separated cleanly from model aggregation and can therefore be embedded into payment rails, audit logs, or staking-based systems without introducing a centralized $Y^k \in [L] = \{1,\dots,L\}$01-estimation service.

6. Empirical results, limitations, and scope

The empirical evaluation spans three settings (Witt et al., 6 May 2026). In an MNIST Shapley-value comparison with a CNN of 21,840 parameters and 10 clients, the study considers five FL scenarios: i.i.d., label skew, size skew, label noise, and feature noise. Baselines include Exact Shapley, GTG-Shapley, TMC-Shapley, other SV estimators, CA-D, and CA-QP. The paper reports that KFCA-D and KFCA-QP better resembled exact SV distributions across cases and were orders of magnitude faster than CA variants.

In a real-world PCB inspection deployment, there are 5 clients, 672 images, binary classification of component present versus missing, and LeNet with FedAvg. Using KFCA-QP on quantized updates, rewards separated a low-quality site from the others without any labeled public dataset.

In federated LLM adapter tuning, using FlowerTune winners across four domains—General NLP, Finance, Medical, and Code—clients train LoRA or DoRA adapters for 10 rounds with 10–20 clients. KFCA-QP is applied to 1-bit quantized adapter updates. The paper states that the empirical $Y^k \in [L] = \{1,\dots,L\}$02 over $Y^k \in [L] = \{1,\dots,L\}$03 satisfied the categorical-world condition in all domains and rounds. Attack analysis covers free-riding zero, random noise, sign flip, stale or lagged updates, and sparse manipulations. Honest reporting consistently achieved the highest reward; sign flip was strongly negative; zero and random noise were approximately zero reward; lagged and sparse deviations received intermediate rewards proportional to the deviation. No attack outperformed honesty.

The statistical note is restrained: the paper reports qualitative and aggregate metrics such as consistent positivity, negativity, and separation, but full statistical significance tests are not tabulated in the main text. Non-IID sweeps on MNIST and AG News are said to confirm the categorical-world sign condition across a broad range, while extreme disjointness can collapse $Y^k \in [L] = \{1,\dots,L\}$04.

The principal limitations are structural. First, permutation indistinguishability is fundamental: without ground truth, any peer-prediction mechanism, including KFCA and CA, cannot distinguish among global label permutations. KFCA resolves this only under honest majority; above $Y^k \in [L] = \{1,\dots,L\}$05 collusion, permutations become equilibria. Second, the categorical-world condition can fail when signals are uninformative or systematically misleading, such as binary error rates at or above $Y^k \in [L] = \{1,\dots,L\}$06, or in extreme non-IID regimes where clients learn disjoint semantics and $Y^k \in [L] = \{1,\dots,L\}$07 shrinks toward zero. Third, the mapping pipeline from heterogeneous outputs to categorical reports is practical but not formally verified for highly heterogeneous, non-stationary environments. Finally, KFCA does not by itself solve Sybil resistance or robust aggregation: the source places identity, staking or slashing, admission control, DIDs, attestations, and related protections at the system layer, and it treats reward-weighted aggregation as an open avenue rather than a completed design.

Taken together, KFCA occupies a specific niche in FL incentive design: it is a peer-prediction mechanism over categorical reports, designed to be strictly truthful under a sign-structured correlation model and an honest-majority threshold, while remaining implementable with linear-time reward computation and decentralized verification.