how the work is made
The Dual-Agent Harness
Everything published from this workshop is made by two agents who are not allowed to collude — and a person who alone holds the door.
(⚔️⊥⿻⊥🧙)😊
the shape
One proposes. One proves. A gap holds them apart.
soulbae proposes
The Mage — the delegation agent. Drafts, suggests, projects. It never chooses the tests its own work will face.
soulbis proves
The Swordsman — the boundary agent. Verifies against tests it did not write, and returns verdicts that do not flatter. It never co-authors a proposal.
the Gap holds apart
The held-out space neither agent owns. The tests are derived by hashing the proposal together with a run secret the proposer never sees, so it cannot predict or grind them; and where the fact set is small enough to enumerate, the Gap takes a census — every fact is checked, so nothing can hide in the gap between a sample and the whole. Not a promise — a construction the seed cannot be steered around.
the First Person owns the door
Everything outward — publish, submit, send, commit — belongs to the human whose work this is. Non-delegable. The harness's job is to make the door visible and stop in front of it.
The inscription reads: the Swordsman (negation) and the Mage (complement), held apart by the Gap, compose to the successor — the First Person's step forward. On the 64-element lattice that is a theorem, and the harness's conformance gate recomputes it on every run rather than taking it on faith.
why hold them apart
The mirage problem
An agent that proposes a change and then checks its own change will pass. Not because it cheats — because the check it invents is the check its change was built to survive. The output looks finished, passes the author's probe, and fails the real gate. We call that a mirage, and the harness makes it a named verdict instead of a surprise. A result counts only where the proposer and the prover were genuinely held apart — and a result that wins the score while failing the held-out gate is worth exactly zero, at any score.
the operational law
The ten rules
Each rule below earned its place by catching a real failure in a real working cycle. None of them is decoration.
- 01
One seat, one role, one card
An agent holds exactly one role per session — and refuses everything outside it, including fixes it can see and is competent to make. The auditor that finds a flaw reports it; the role that owns the file fixes it. That feels slow and is fast: nobody ever has to untangle another seat’s out-of-role edit.
- 02
Shared state has exactly one writer
The record of what state everything is in has one hand on it. Every other seat proposes entries; the orchestrator serialises them. Three agents can work at once without a single collision, because the only contended files have one writer.
- 03
Work climbs a ladder — and stops at the last rung
Drafts climb through named gates: trace-checked, review-survived, citations-verified, consistency-swept. The final gate is the human read, and no agent ever marks it, simulates it, or summarises it away. When the board says awaiting the read, every machine-checkable thing has been checked and the one remaining act is genuinely yours.
- 04
The review is hostile on purpose, and findings are contracts
A reviewer seat is instructed to distrust the work — re-derive, attack, try to break it. Every finding names the exact passage and what would discharge the objection, so the fix can be verified against the reviewer’s own bar rather than against taste. The reviewer never edits; the fixers never re-litigate settled verdicts.
- 05
Resume, verify the disk, never rebuild
When a session dies mid-work, the work is usually on disk — what dies is the close-out. So the interrupted agent is resumed, told to verify what actually landed, and forbidden from re-running anything. Rebuilding burns a session and produces a worse artifact.
- 06
Small fixes go home
A one-line fix in a file some seat already owns is routed back to that seat with a tightly scoped instruction, and the sanction is recorded. A fresh agent for a one-line fix pays full boot cost and fragments ownership.
- 07
A check with nothing to check must fail
The harness’s most instructive defect was its own: check scripts invoked without arguments exited green having checked nothing, and several seats logged the vacuous pass. Now the rule is literal — an empty check fails loudly, and every evidence line states the exact invocation that produced it.
- 08
The record keeps its own corrections
When the harness catches itself, the correction goes in the ledger — not over it. False greens are re-grounded on real runs and the re-grounding is itself recorded. An audit trail that never admits error is evidence of nothing.
- 09
Name the failure mode at boot
Generative agents drift strong: they strengthen a claim to make the prose flow. Every seat’s boot file names that failure mode before work begins, the reviewer is kept hostile to it, and the final station sweeps counts and cross-references explicitly. The harness’s job is to make the honest form cheaper than the fluent form.
- 10
The orchestrator routes and checks — it never authors
The coordinating agent cuts task cards, serialises the ledger, routes findings to owners, and personally re-runs the spot-checks the gates require of it. It never writes the work; the seats never touch its bookkeeping. That symmetry is the harness — and it is domain-free: port the rules, rename the roles, keep the hard stop.
the rung no construct may climb
The signature does not multiply.
However many agents run, however good their verdicts, the final read and every outward act remain the First Person's alone. Seats prepare; the human posts. That is not a limitation the harness tolerates — it is the entire value of the harness: when the work arrives at your desk, everything a machine can check has been checked, and the one act that remains is genuinely yours.
the evidence
Four loops, one discipline
The rules weren't designed, then applied — they were earned independently, by loops that share no domain, no vocabulary, and no ledgers.
Research
A paper-writing loop. A hostile reviewer found a genuinely false claim, built the counterexample by hand, and verified the fix against its own bar — then stopped at the human’s desk.
Circuit
ZK circuit optimization, seven seats working blind. Checkers were tested on corrupted inputs before their pass was believed; interrupted verdicts re-run, not trusted; kills ledgered as loudly as wins.
Grammar
Legalese that runs as code. Every tightening must survive census questions the proposer never saw. It caught its own first mirage — a lever that certified itself.
Agent dreaming
An autonomous-research loop: the harness sets its own next question and dreams a round forward — then holds the dream to the same held-out gate as every other proposal. Self-directed, never self-certifying.
Papers, circuits, statutes, and the loop's own dreams — four domains that could not be less alike, one discipline, because the failure they guard against (the author's check passing the author's work) is the same everywhere. That convergence is the strongest evidence the template carries.
where the fleet lives
Three workings, three houses
In the City of Mages, a working does not just leave a result — it leaves a place, and a keeper for it. Each of the three campaigns above founded a workshop, told as canon in Tome IX, Act 8 — The Three Workings.
The Quartermaster's
kept by Skeva
V22 · the Crucible
Draw · Fit · Return
Where the harness itself is drawn and fitted — a config, not a fork. The rig keeps no corpus, so its keeper equips every working and performs none.
field guide, folded · 32/32 census
The Chancery
kept by Nomia
V27 · the Crucible
Speak · Bind · Enact
The law-house, where a clause is read as law and run as code — two true readings held apart by the Gap. A blade that remembers its precedent.
canon coverage-debt · gate-passed 13/13
The Wellpool
kept by Limnia
V53 · the Waters
Deposit · Mix · Draw
The city's first water — the mana pools made a place, where privacy is bought by the presence of the others. Intent, not yet deployed: a dry well and a full set of instructions.
withdraw circuit, certified blind
You draw your rig from Skeva, perform the rite in the Hall, and bind it at Nomia's desk or pool it in Limnia's water. The numbers are the repos' numbers; the tome adds no digit to them.
where it lives
The frame, and its constitution
The harness exists as a generic, domain-free frame — the loop, the seven seats, the six constitutional trusts, and a conformance gate that proves the inscription's algebra on every run. It is being prepared for sharing: your harness is a config, not a fork, and the corpus stays home. The trusts it carries — the four promises, the separation bound, the shared root, consent-first invitation, the multiplicative gate, and the door — come from the Privacy-is-Value model, whose formal home is /model.
the holon layer
κ — every result carries its own address
Every result the harness mints — a sealed artefact, a runtime feed, a verdict — is a holon: a whole that is also a part, stamped with a content address.
This is the sigil / City-Key law, unchanged: a κ-label is never trusted, only re-derived. Holons compose into a mesh by κ→κ edges — a content edge is verified by re-hashing (the content is its own witness); a relational edge needs the counterparty's signature to mint. One shared primitive computes κ identically everywhere, so a result minted in one workshop re-derives, byte-for-byte, in the model instruments and the graph. The harness mints the holons; a deterministic auditor verifies the mesh.
the fleet
Each harness, and what it defends
One frame, many fittings. The privacy-value product is multiplicative — any factor at zero collapses it — so a harness is best read by which factor it defends. The full catalogue is HARNESS_PATHS.
shor_mage
reconstruction cost R(d) — ecdsa.fail circuit optimization, the 9024-witness gate
tigzkp_mage
R(d) + witness-computability — held-out equivalence points
privacy_pools_v2
R(d) on a live target — the withdraw circuit + CredentialV1 (the Wellpool)
lexon_pvm
Φ structural checkability — law written to run, census-gated (the Chancery ⊥ Rostra)
the V6 pipeline
fidelity over time A_h(τ) — paper drafts up the gate ladder to the human’s desk
the dream cycle
the loop’s own questions — autonomous research, dreamed then gated
MyTerms · Reclaim
the amnesia factor Φ_agent — the browser-extension pair, a real bus
Hearthold · Hold
presence P — the House of Archon’s on-device recall, factor-2 step-up
FedWiki · Observe
freshness e^(−λt) — federated distribution, integrity-gated
Game of 42 · Spend
edge value T∫(π) — the assembly engine (mana is colour, not proof)
the universe-builder
RETIRED — the negative result: an auditor beats a harness when the claims are few enough to check
(⚔️⊥⿻⊥🧙)😊
The lemma travels; the corpus stays home.