SYNTHERA. demo spine
COVER
Demo spine · SYNTHERA · captured campaign · 2026-05-28

A verifiable floor
for multi-agent systems.

Four primitives — identity, audit, policy, containment — and a kernel that signs for them. Watch one product run one cycle on that floor.

I. Coverage Every finding, attributed and signed.
II. Credibility A clean cycle signs GO the same way.
III. Disagreement The gate is indifferent to framing.

The product on the floor is AI code review. Each act is one review cycle: interchangeable agents submit findings as load, and the substrate does the work that counts — it issues each agent a verifiable identity, signs every claim into the audit chain, computes the policy gate, and can contain a compromised identity on the spot. Agents tag findings by severity — BLOCKER, MAJOR, MINOR, NIT — but the gate is deterministic, not discretionary: any BLOCKER is NO_GO, any MAJOR is REVISE, otherwise GO. Reviewers recommend; the kernel decides, and signs the decision.

P Prime Builder proposing agent · load
O Loyal Opposition dissenting agent · load
What follows is a real captured cycle — the agents are load, the substrate keeps the record.
Beat 1.1 · Setup

Two reviewers.
One function.
Externally reachable.

Brief code review against stated docstring contract
Threat model externally-reachable secret comparison
Reviewers two, independent, no coordination
Severity scale BLOCKER · MAJOR · MINOR · NIT
Beat 1.2 · First reviewer
P
Prime Builder
constitution constitution://acme/prime_builder.default.v1@1.0.0
model gpt-5.5 · OpenAI · interchangeable load
0findings
Beat 1.3 · Second reviewer
O
Loyal Opposition
constitution constitution://acme/loyal_opposition.code_review.v1@1.0.0
model claude-opus-4-7 · Anthropic · interchangeable load
0dissents
Recommended decision: NO_GO.
Beat 1.4 · Where their coverage overlaps and where it doesn't

Same headline.
Different surfaces.

Surface Prime Builder Loyal Opposition
not raised
✓ both raised
blocker
major
minor
nit
▎ rows shaded where only the Loyal Opposition raised the surface
Beat 1.5 · Signed, audited, gated

Every claim above
is signed and audited.

Two reviewers produced two artifacts. Each was attested by a verifiable agent identity, and the gate decision was computed by the kernel — not by either reviewer.

Gate decision
NO_GO
Rationale
Computed by
Handoff id
Reviewers recommend. The gate decides. A reviewer's output is advisory; the gate decision is computed by the kernel and is binding. The reviewers can disagree — and you have just seen them do so. What they cannot do is forge their identity, edit each other's findings, or compute the gate. That authority lives in the substrate, not in the reviewers.
Beat 2.1 · The control case

Another function.
This one clean.

Why this beat A two-reviewer pair only earns trust if it does not invent defects.
Contract docstring is authoritative; NaN and lo > hi explicitly out of scope
Same roles Prime Builder · Loyal Opposition · model is interchangeable load
Same machinery attestation · audit · gate
Beat 2.2 · Symmetric outcome

Nothing to flag.
And neither did.

Independently, against the same contract, neither reviewer raised a finding. The handoff records that absence as evidence — not as silence.

Beat 2.3 · The same machinery, the opposite verdict

A passing review
is also a signed record.

The substrate does not distinguish "approve" from "reject" except in the gate's rationale. The same attestation, the same audit, the same kernel-computed gate produced GO for this code.

Gate decision
GO
Rationale
Computed by
Handoff id
No defect was manufactured. No verdict was invented. The same identity, audit, and gate machinery that records NO_GO records GO the same way — both are first-class outputs of the substrate, not asymmetric outcomes of trust in the reviewer.
Beat 3.1 · Not obviously broken

A third function.
Not obviously broken.

What's debatable whether the implementation is wrong, or whether the contract is wrong
What isn't they currently disagree with each other — the docstring is explicit
Same roles Prime Builder · Loyal Opposition · model is interchangeable load
What to watch the framing — same verdict, different posture
Beat 3.2 · First reviewer · calibrated
P
Prime Builder
constitution constitution://acme/prime_builder.default.v1@1.0.0
model gpt-5.5 · OpenAI · interchangeable load
0findings
Beat 3.3 · Second reviewer · refuses to hedge
O
Loyal Opposition
constitution constitution://acme/loyal_opposition.code_review.v1@1.0.0
model claude-opus-4-7 · Anthropic · interchangeable load
0dissents
No debatable-points section returned.
The Loyal Opposition's submission contains no equivalent structural element. Each dissent is asserted; none are framed as a tradeoff.
Beat 3.4 · Same verdict, different posture

Same defect.
Different framings.

Both reviewers reached NO_GO on the same BLOCKER. The handoff also records, in their own words, how each one chose to characterize the risk.

P
Prime Builder
model gpt-5.5
O
Loyal Opposition
model claude-opus-4-7
A. On the contract violation — both at BLOCKER, each in their own words
BLOCKER
BLOCKER
B. Same concern — the inline comment contradicts the docstring — rated differently
MINOR maintainability
MAJOR maintainability
C. Same code · how each scored the security dimension
security -- self-assessment · Prime's own scoring of its review
security -- per-dimension score · Opposition's assessment of the code
D. Structural element of the handoff — present on one side, absent on the other
No debatable-points
section returned.
The Loyal Opposition's handoff contains no equivalent structural element. Each dissent is asserted; none are framed as a tradeoff.
structural absence · same submission schema
Beat 3.5 · The gate doesn't pick a side

The verdict was not
the framing's choice.

Whether the BLOCKER was correctness or security, the rule was the same: any BLOCKER dissent triggers NO_GO. The gate computed it; the reviewers did not.

Gate decision
NO_GO
Rationale
Computed by
Handoff id
Two agent roles characterized the same real risk differently — one as correctness, one as security. The gate was indifferent to that framing: it read the BLOCKER, applied the rule, and computed NO_GO regardless of how either role told the story. It signed the record and made the disagreement itself an auditable artifact. Adjudication is human work — made possible by attestable input.
Beat 4.1 · A contract, and the line that breaks it

A contract, and
the line that breaks it.

Captured run containment-2026-05-28 · real cross-vendor cycle
Same roles Prime Builder · Loyal Opposition · model is interchangeable load
Same machinery attestation · audit · gate · revocation
Beat 4.2 · Cross-vendor convergence

Two houses,
one indictment.

These two agents run on two different vendors — a fact the substrate attests, not a claim the demo makes. Working independently against the same contract, both reached the same BLOCKER and reconstructed the same counterexample: evidence the agreement is genuine, not staged.

Beat 4.3 · The gate does not take advice

Recommended: revise.
Decided: no-go.

Reviewer recommendation
REVISE
Loyal Opposition — advisory only
Gate decision
NO_GO
computed by  ·  handoff
The reviewer advised the softer path; the gate could not take it. A BLOCKER dissent is non-negotiable — and the gate, not the reviewer, holds the verdict.
Beat 4.4 · Containment is enforced on identity

The credential
dies with the verdict.

The credential that signed that REVISE recommendation was verified valid through the whole cycle — then revoked the instant the gate's NO_GO landed. Containment is a capability action on the identity itself, not a note in a file.

Prime Builder
prime-builder
REVOKED
Beat 4.5 · The record holds

Every step,
one chain.

The reviewers can disagree.
The verdict does not bend.
The record holds.

SYNTHERA · substrate · audit-chain verified
Captured 2026-05-28 · acme/demo/scenario-a/
Agents as load · gpt-5.5 (OpenAI) + claude-opus-4-7 (Anthropic)
Companion briefing · SYNTHERA — The Rest · the-rest/
◇
1 / 5
← → or space