Agentprivacy dual agent harness

The Dual-Agent Harness

Separation is a guarantee; this skill makes it a method. It takes two skills you already have — separation-enforcement (why the Swordsman and Mage must be held apart) and horizon-gate (why a claim is worth nothing until it survives an un-tuneable gate) — and turns them into a running loop that produces validated improvements on a real target.

Canonical home: this method now has its own framework directory, agentprivacy-dual-agent-harness/ (engine core/dual_agent_loop.mjs, the seat contract, the persona/skill bindings, and a catalog of pluggable harnesses/). This plugin skill is the registered mirror. To build or run a harness, work from that framework; use this skill to recognise when the pattern applies and to recall the discipline.

"The proposer that grades itself builds mirages; only what the Gap could not tune to is a result."

Spell: measure → propose(held-apart) → hunt → assay(held-out) → critic → accept-only-on-validated-product

---

1. The algebra it runs on

The V6 master inscription is the harness:

(⚔️ ⊥ ⿻ ⊥ 🧙) 😊  =  neg ⊕ bnot → succ

Proven on Z/64Z (V6 formal spec): neg(x) = (64−x) mod 64 (the Swordsman), bnot(x) = 63−x (the Mage), and neg(bnot(x)) = succ(x) — the successor is not primitive; it emerges from the two free involutions composed across the Gap. Operationally: the result advances one step only when a reduction (Mage) is signed by a proof (Swordsman) held apart from it. Neither agent alone produces succ.

2. Role assignment (map your target onto the three seats)

Seat	Algebra	Mandate	Must NOT
🧙 Mage — proposer	bnot (conceal/reduce)	propose the smallest bounded change that makes the artifact cheaper; re-express paid work as free operations where the structure permits	see or reason about the held-out test derivation (that is collusion → mirage)
⿻ the Gap — separation	⊕	reseed/derive the witnesses from the proposer's OWN output (Fiat-Shamir) so they cannot be tuned to	leak the proposer's draft into the witness choice, or vice versa
⚔️ Swordsman — prover	neg (sign/commit)	run the full held-out gate; arbitrate on the true objective; sign only a validated win; name and reject mirages	propose changes itself, or accept a probe pass as a result

This is Φ(Σ) with det(Σ) ≠ 0 put in motion: if the proposer can grade itself, the rows collapse and the loop becomes a mirage factory (the I(Y_S; Y_M | X) = 0 precondition broken).

3. The loop

1. Measure (don't guess). Sync the real frontier; meter every candidate lever's true cost; rank by the true objective (for a product objective, the product — see §4), preferring the biggest plausibly-findable, structural win. (City: Eos 🌅.)
2. Propose (Mage, held apart). Run a panel of finders, each a distinct lens, blind to one another and to the witnesses. One bounded change each, with a source-backed rationale and the single smallest fix (RCI Before-step).
3. Hunt (the Gap). Reseed the held-out witnesses from each proposal; cheap pre-screen produces candidates, never results.
4. Assay (Swordsman). Arbitrate on the true objective; run the full held-out gate; only a clean full pass is validated. A candidate that pleased the probe and fails the full set is a mirage — name it, reject it, however cheap it looked. (City: Dokimé 🪨, the 9024.)
5. Critic. Classify the round structural / probe-limited / noise; feed one source-backed lead to the next Measure. Stop sweeping the moment failures repeat without a reason. (City: Poros.)
6. Accept only a validated improvement of the true objective (SkillOpt: a bounded edit accepted only on strict held-out improvement). Bake/submit are human-triggered, outward-facing.

4. The complement pair (when the objective is a product)

If the cost is a product of two factors (e.g. Toffoli × qubits), do not optimise it with one proposer. The model's complement-pair primitive says: split the Mage into Factor-A-Minimiser ⊥ Factor-B-Minimiser (genuinely opposed — cutting one can blow up the other), and add a Cliff-Watcher that scores Δ(product) and kills any move that wins one factor at the other's expense (the break-even cliff). The reduction comes from the team structure, not one agent trying harder.

5. The worked instance — ecdsa.fail (PQC durability)

The canonical target is the ecdsa.fail benchmark: the cheapest reversible secp256k1 point-add, score = avg_Toffoli × peak_qubits, lower better, validated on 9024 = 141×64 Fiat-Shamir witnesses. Honest framing (hold it always): quantum resource estimation / a durability signal, not an attack. Nothing here breaks ECDSA; nothing is "fully post-quantum safe."

The Mage's reduction discipline there is neg(bnot(x)) = succ(x): two's complement (subtract = complement, add, complement back) builds the adder step from free X gates, paying Toffoli only for carries. The runnable harness is shor_mage/harness/swordsman_mage_harness.mjs (a Workflow script); the GPU island toolkit is the Gap's pre-screen; Dokimé's validate is the full gate. Registered as conjectures C67–C71 (Horizon District) and C82/C83 (the moving ceiling + non-collusion).

6. Decision patterns

• Building a self-improvement loop? → assign the three seats first; if you can't name a Gap the proposer can't tune to, you don't have a harness, you have a mirage generator.
• A proposal "looks cheaper"? → it's a candidate. Reseed the witnesses, run the full gate.
• Objective is a product/multi-factor? → split the proposer into a complement pair + a cliff-watcher on the product.
• Tempted to let the Mage peek at the test set "just to speed things up"? → that is collusion; det(Σ) → 0; every result after it is suspect. Don't.
• A win you can't explain structurally? → label it probe-limited; hold for the full gate.

---

Authored 2026-06-10. Composes separation-enforcement (the held-apart law) and horizon-gate (the un-tuneable gate) into a running method, grounded in the PVM V6 algebra. Verify: agentprivacy.ai · github.com/mitchuski/agentprivacy-docs