EXHIBIT 08● REAL Fp ARITHMETIC · INTERACTIVE PROOF
Sum-check and a GKR layer walk
A verifier checks a sum over 2v points with v short messages and one evaluation, then reuses that check to walk a small arithmetic circuit from its outputs to its inputs.
✓ Completeness ✓ Act A error ≤ v·d / |F| under fresh uniform challenges unpredictable until each message is fixed ✗ Not zero-knowledge · not a SNARK

Adding up a polynomial over every corner of a Boolean cube takes 2v evaluations. The sum-check protocol lets a verifier accept that total after reading v small polynomials, checking two things about each, and evaluating the original polynomial once at a point the prover never got to choose.

This is an interactive proof of a sum and, in Act B, of a circuit evaluation. It hides nothing from the verifier, so it is not zero-knowledge, and it is interactive, so it is not a SNARK. Every number on this page is computed by the run you trigger; nothing is read from a script of expected answers.

Cryptographic honesty: every value below comes from exact BigInt modular arithmetic in the prime field you select. The verifier checks a degree bound and a recurrence in every round and draws each challenge only after that round's message is fixed. The tiny-field census is a complete enumeration for teaching, not a security parameter. This is inspectable teaching code for an interactive proof: it is not zero-knowledge, not a SNARK, and not a constant-time library.
ACT A

Sum-check · prove a sum over {0,1}v without listing it

The statement is H = Σx ∈ {0,1}v g(x) for a public polynomial g over the prime field Fp. In round j the prover sends a univariate polynomial sj(Xj) and the verifier checks its degree against degj(g) and the recurrence sj(0) + sj(1) = sj−1(rj−1) before drawing rj. After round v it evaluates g once at (r1, …, rv).

STATEMENT · FIELD, POLYNOMIAL, VERIFIER RANDOMNESS
Verifier randomness
|F| = p
—
variables v
—
degree bounds degj(g)
—
points listed, 2v
—
H = Σ g over {0,1}v
—
soundness bound v·d / |F|
—

H above is computed by listing all 2v corners and adding g at each — the work the verifier wants to avoid. The protocol below accepts or rejects the prover's claim with one evaluation of g, and the test suite re-derives H with its own code rather than trusting this number.

Randomness: fresh verifier randomness.

PROVER STRATEGY · THE VERIFIER NEVER SEES THIS CHOICE

The honest prover sends sj exactly as the recurrence demands.

TRANSCRIPT · ONE ROUND PER VARIABLE, CHALLENGE DRAWN AFTER THE CHECK
No protocol run yet. The prover will first announce its claimed sum.
RoundProver sends sj(Xj)Degree ≤ boundsj(0)+sj(1)ExpectedVerifier checkChallenge rj
No rounds yet — run or step the protocol.

Claimed sum: —

Final evaluation: not reached.

Run the protocol to compute a verdict.
Prover-side narration (the verifier never reads this): no run yet.
BREAK IT YOURSELF · REPLAY A TRANSCRIPT, THEN DAMAGE IT

A finished run records each round's message and challenge, but not event timestamps. Replaying it re-runs structural and algebraic checks: a missing or swapped round, a challenge with no message, a value outside the field, or an edited message is refused by name. Replay cannot prove that a recorded challenge was historically sampled after the message or was unpredictable; the live verifier's phase guard and fresh-random mode enforce those separate requirements.

No transcript replayed yet.
TINY-FIELD CENSUS · EVERY CHALLENGE VECTOR, EVERY VERDICT

Over a field of 13 or fewer elements the page can run the real verifier on every challenge vector and count what happens. That is a complete teaching enumeration, not a security parameter: the bound v·d / |F| only means something when |F| is enormous compared with v·d.

Select p = 13, 7 or 5 above to enable the census.

ProverAcceptedRejected by degreeRejected by recurrenceRejected at final evaluationOf
honest—————
wrong sum, carried lie with explicitly counted roots—————
full-field interpolant for s1 (degree ≤ p−1)—————
Readout: no census yet. The carried lie succeeds exactly when some rj lands on one of that round's explicitly constructed roots, so its accepted count is pv − Πj(p − rootsj). The v·d / |F| theorem uses degree bounds, not an assumed root count. The interpolant agrees with the honest s1 on p − 1 field points; it is refused on degree only when p − 1 exceeds the public X1 bound.
Run the census to compare the counts with the bound.
ACT B

GKR layer walk · one fixed circuit, two layer reductions, one terminal evaluation

The statement is C(x) = (y0, y1) for the fixed two-layer circuit below. Each layer i turns the claim W̃i(zi) = mi into a sum-check over the wiring identity W̃i(z) = Σb,c add̃i(z,b,c)·(W̃i+1(b) + W̃i+1(c)) + mult̃i(z,b,c)·W̃i+1(b)·W̃i+1(c), then reduces the two evaluations the verifier cannot make to one claim on a random point of the line through them. At the input layer the verifier evaluates W̃2 itself from the public inputs.

THE CIRCUIT · FIXED WIRING, EDITABLE INPUTS, VALUES RECOMPUTED GATE BY GATE
LayerGate labelOperationReads (in1, in2) in the layer belowValue
2 (inputs)(0,0) (0,1) (1,0) (1,1)input——
1(0,0)add(0,0), (0,1)—
1(0,1)mult(0,1), (1,0)—
1(1,0)add(1,0), (1,1)—
1(1,1)mult(0,0), (1,1)—
0 (outputs)(0)mult(0,0), (0,1)—
0 (outputs)(1)add(1,0), (1,1)—

Wiring predicates the verifier derives for itself: —

Raw union sum for this walk: —

PROVER STRATEGY · THE VERIFIER HOLDS ONLY THE CIRCUIT, THE INPUTS AND ITS COINS

The honest prover evaluates the circuit, sends its outputs, and answers every sum-check from the true tables.

THE WALK · OUTPUT CLAIM, LAYER 0 → 1, LAYER 1 → 2, TERMINAL CHECK
No walk run yet.
Run the walk to compute a verdict.

What this run is and is not. Executed: the bare-bones layer reduction for exactly the two layers of this circuit — one sum-check per layer over 4 variables with degree bound 2, one line reduction per layer with degree bound 2, and a terminal evaluation of the input layer's multilinear extension by the verifier. Not executed: the log-space-uniformity machinery, the sublinear verifier, Fiat–Shamir, or any general theorem; this verifier evaluates add̃ and mult̃ by enumerating all six gates.

Sources, and what is adapted

The sum-check technique is Lund, Fortnow, Karloff and Nisan, “Algebraic Methods for Interactive Proof Systems,” J. ACM 39(4), 1992 (author version read here: §3 “The Protocol” — Theorem 1, Lemma 4 and the formal Expand/Shrink protocol — and §4 “Extensions”). That paper states the protocol for the permanent; the one-variable-per-round sum over {0,1}v executed here is the form in which Shamir and Babai–Fortnow applied it, which Goldwasser, Kalai and Rothblum restate as Figure 2 and prove as Lemma 2.5 (error m·d/|F|) in the full version of “Delegating Computation: Interactive Proofs for Muggles,” ECCC TR17-108 (revision 1, 2017; STOC 2008 extended abstract). Act B follows that full version's §3.1 (layered circuit, addi/multi as equations (4) and (5), Ṽi), §3.2 (the i'th phase: sum-check, then the line step 1–4, then the final verification) and §3.3 (soundness 7mdδ/|F|).

Two adaptations are deliberate. GKR's bare-bones protocol sums over Hm for an extension field H of GF(2) with low-degree extensions of degree δ and an extra β̃ factor; this page takes H = {0,1} and the multilinear extensions, the simplification Thaler attributes to his own 2015 work and to Cormode–Mitzenmacher–Thaler (Thaler, Proofs, Arguments, and Zero-Knowledge, 18 July 2023, §3.5 Lemma 3.6, §4.1 Proposition 4.1, §4.6.4 Lemma 4.7, Remark 4.4 and Figure 4.13). GKR's line reduction uses two fixed distinct points t1, t2; this page uses ℓ(0) = b*, ℓ(1) = c*. The raw union sum printed in Act B is computed for this circuit from those adapted degrees, not copied from either paper. It is a probability bound only with fresh, independent, uniformly sampled verifier challenges that remain unpredictable until the owning prover message is fixed; it is capped at 1 and labelled vacuous when the raw sum reaches 1. Predictable seeded runs do not inherit that theorem.

Related demos: Exhibit 05 shows how a hash would replace the verifier's coins — this page never does that; Exhibit 06 is the conceptual SNARK pipeline this protocol is not; STARK Tower builds the FRI side of transparent proofs; ZK Arena compares proof systems that put sum-check and its relatives to work.

Extension seams: apply the Exhibit 05 transform to the verifier's coins and show the transcript it produces; or grow the circuit and give the verifier a wiring formula so add̃ and mult̃ no longer cost a pass over every gate.