A response to “The New SDLC With Vibe Coding”

The new SDLC, made formal.

Google’s whitepaper drew the new software lifecycle in prose and diagrams: Agent = Model + Harness, specification as the bottleneck, verification as the differentiator, evals as the bar. It told you what the discipline looks like. The [&] Protocol stack ships the substrate — every one of those informal ideas turned into a composable, property-tested protocol with a verdict you can audit.

210+4
governance laws enforced · 3 declared open
5
canonical loop phases (PULSE)
11
normative OS protocols
6 · 2
[&] primitives · operators

01 · The spectrum, restated

It was never about whether you use AI.

The whitepaper’s central move is a continuum — vibe coding on one end, agentic engineering on the other — where the only real axis is how outputs get verified. Drag the slider. Notice that the words change, but the question never does.

Vibe coding Structured AI-assisted Agentic engineering
Vibe coding
Casual prompts, “does it seem to work?”, disposable code. Perfect for prototypes and throwaways.
Verification: a glance at the demo trust: vibes
“Generation is mostly solved. The work that’s left is specification and verification — and the systems that hold them together.” — The New SDLC With Vibe Coding (Osmani, Saboo, Kartakis), Google, 2026

02 · Agent = Model + Harness

The model is ~10%. The harness is the product.

The whitepaper’s most quoted figure: swap the model, get marginal gains; swap the harness, move an agent from outside the Top 30 into the Top 5 — that result is on Terminal Bench 2.0, with no model change, and it’s worth checking rather than taking from us (Osmani, Agent Harness Engineering, April 2026). The intelligence is rented. The scaffold — instructions, tools, memory, orchestration, observability, governance — is what you actually build and own.

~10%

The model

Rented intelligence. Interchangeable. The part you don’t control and shouldn’t over-invest in.

~90%

The harness / scaffold

Instructions & rule files, tools & MCP servers, sandboxes, orchestration, hooks, observability — and the governance that decides what’s allowed to run.

+&

What the stack adds

Each harness layer becomes a named protocol with a schema and a conformance suite — not a folder of prompt files you hope still works next week.

03 · The open problems

Harness engineering has three open problems. We took a run at each.

The whitepaper wasn’t the end of it. In April 2026 Osmani followed up with Agent Harness Engineering, which closes by naming three things nobody has solved yet. That list is a better organising principle for this page than “here is our stack,” because it says what the stack is for — and it lets you grade us against someone else’s scorecard instead of one we wrote. These are attempts, not solutions. Every status below is the same label used in the maturity table.

Open problem 1

“Orchestrating many parallel agents on shared codebases.”

Our answer is to remove the thing that needs orchestrating. TRVM is a distributed interaction-calculus runtime: interaction-net reduction is confluent by construction, which licenses schedule-independence, and a boundary-port discipline extends that across machines. Agents don’t take turns on a shared tree — they reduce independently and their results merge as a content-addressed CRDT (NF(A ∪ B)). Above it the chain is WRL describes ▸ Forge lowers ▸ TRVM reduces: WRL is the world/topology language that says what is arranged where, and a world’s meaning is its hash. That last part is the testable one — the same world lowered by WRL’s JavaScript spine and by Forge’s independent Python implementation round-trips byte-exact, and on a shared tamper corpus both verifiers refuse all 15 and agree on the name of every refusal. Two languages, two repos, same meaning.

TRVM live_localWRL live_local890 checks · 0 failed
Not proven: parallel speedup. Conformance is green across four implementations and swarm.js runs coordination-free across real isolated workers — correctness, not throughput. §6.4 snapshot and §6.5 REF semantics are still unspecified.
Open problem 2

“Agents analysing traces to self-identify harness failures.”

This is the one the stack was built for. PRISM scores trajectory as well as output and reads any system’s PULSE manifest to discover its phases; Graphonomous’s learn phase folds the outcome back as a Wilson-interval confidence update rather than a vibe. And Workbench /record closes it on real data: drop in a Claude Code session transcript and it becomes a scored bundle — no model call, no API key, nothing uploaded.

PRISM live_deployed/record live_local
Not proven: that any of it generalises. PRISM has adapted no external system it doesn’t own, and /record’s gates were shown to carry signal on one corpus.
Open problem 3

“Dynamic, just-in-time tool and context assembly — not static pre-configuration.”

Static pre-configuration is exactly what a folder of rule files is. [&] makes capability a declaration that composes (&, |>) and compiles down to MCP/A2A at the moment it’s needed, rather than a fixed server list. Graphonomous supplies the context side: retrieval scored by spreading activation, and κ-routing — topology decides how hard to think, so a cyclic knowledge region gets deliberation and a tree gets a straight answer.

Graphonomous live_deployed[&] in_tree
Not proven: assembly is JIT per call, not yet per step inside a long-horizon run. &body still has no runtime.
The useful version of “we built a stack” is “someone named three unsolved problems and here is what we tried, with the failures attached.” If one of these three is later solved by somebody else, that’s the outcome we wanted — the problems matter more than our answers to them.

04 · The thesis

They described it. We made it composable.

A whitepaper can say “set the bar at the eval, not the demo.” But a sentence isn’t a guarantee. The [&] stack takes each pillar of the new SDLC and reifies it as an algebra you can compile, compose, and machine-check — so the bar isn’t a slogan, it’s a verdict with a certificate.

[&]

[&] Protocol

composition · specification

Specification stops being a Markdown file you hope the agent read. It’s an algebra — 6 primitives (&memory &reason &time &space &body &govern), 2 operators — compiled from .ampersand.json and validated against a draft-2020-12 schema.

6 primitives2 operatorsElixir + npm + Python SDK
maps to → “specification is the bottleneck”

Graphonomous

memory · context engineering

The whitepaper’s “context engineering” — static vs. dynamic, the cost knob — as a continual-learning knowledge graph with κ-routing, attention, and consolidation. Context isn’t stuffed into a prompt; it’s retrieved, scored, and learned from.

OS-001 · OS-002573 testsκ cyclicity routing
maps to → context engineering & the harness’s memory

PULSE

temporal · the loop itself

Google draws a verification loop — run, cluster failures, fix, regress, watch production. PULSE makes that loop a declarable artifact: retrieve · route · act · learn · consolidate, five canonical phases in a .pulse.json manifest, with typed cross-loop signals.

OS-010 v0.1.15 phases6 cross-loop tokens
maps to → the verification / feedback loop

PRISM

diagnostic · evals

“An eval suite with a real rubric shows it works reliably.” PRISM is that suite as a protocol — bring-your-own-runtime benchmarking that scores both output and trajectory, reading any system’s PULSE manifest to discover its phases.

OS-009172 tests · Fly.iooutput + trajectory
maps to → output & trajectory evaluation

box-and-box

governance · the bar

“Set the bar at the eval, not the demo” becomes an eight-rung modality ladder running feasible ▸ permitted ▸ best over an un-weakenable safety floor. Every decision ships a certificate. 210 property-tested laws at 2000 trials each — plus 3 declared open, printed in red by the suite itself.

8 rungs210 enforced · 3 openverdict certificate
maps to → CI/CD gates & “the bar”

OS-008 Harness

the harness, as protocol

If the harness is 90% of the agent, it deserves its own discipline. OS-008 (TRAAVIIS) is “the harness engineer that builds your harness” — pipeline enforcement, quality gates, and sprint contracts, sitting above generation.

OS-008terminal harness CLIquality gates
maps to → Agent = Model + Harness

05 · The mapping

Every whitepaper concept → a named protocol.

This is the whole argument in one table. Left: what Google named. Right: where it lives in the [&] ecosystem. Named is not the same as shipped — the maturity table in For the builder ▸ Honesty grades every row spec / in_tree / live_local / live_deployed, and two of them are still spec. The novelty isn’t any single row — it’s that the rows compose with & and |>.

The New SDLC concept (Google)[&] stack realization
Agent = Model + Harness — the harness is ~90% of the agent OS-008TRAAVIIS harness protocolThe scaffold is itself specified, gated, and contract-bound — not an ad-hoc prompt folder.
Specification is the bottleneck — a spec + prototype from one conversation [&]Composition algebra6 primitives · 2 operators · machine-checkable .ampersand.json, three SDKs.
Context engineering — static vs. dynamic, the cost knob OS-001/2GraphonomousContinual-learning graph with κ-routing, attention & consolidation — context is retrieved, not stuffed.
The verification loop — run → cluster → fix → regress → watch prod OS-010PULSE manifestFive canonical phases retrieve·route·act·learn·consolidate in declarable .pulse.json.
Output + trajectory evaluation — “you want both” OS-009PRISM benchmarkBYOR cross-system measurement that scores the answer and the path taken to it.
CI/CD gates · “set the bar at the eval” kernelbox-and-box governancefeasible ▸ permitted ▸ best over an un-weakenable floor; 210 property-tested laws; every verdict certified.
Authorization & guardrails — what the agent may actually do OS-006delegaticAuthorization kernel: HMAC-signed AuthorizationBlocks, fresh auth on destructive edges.
Orchestrator — hand a goal to autonomous agents, review what comes back north starThe dark-factory loopperceive → record → learn → crystallize → ship → replay → measure. Seven steps, closed locally.
Honesty note. The dark-factory loop is live_local — exercised end-to-end against real local substrates and a cross-process replay — not yet stitched across the deployed Fly.io machines (live_deployed). Three MCP services are live in production; the full deployed end-to-end pass is the next bar. Numbers above are the canonical counts from STACK_COMPLETION.md (2026-06-28).

Don’t take “composable” on faith — toggle a capability. The declaration, the compiled MCP/A2A targets, and the governance certificate all regenerate live. The payoff: because it’s a protocol, nothing here is bespoke — any [&]-aware tool reads the same file. Notice the certificate only appears once you compose in &govern.

① declaration · agent.ampersand.json

        
② compiled targets · floor
③ verdict · certificate

Illustrative widget. The certificate shape mirrors box-and-box’s real output; the actual 210 property-tested laws run in the kernel — reproduce them with the command in For the builder ▸ Verify.

06 · The loop, named

A verification loop you can declare, not just draw.

The whitepaper sketches the feedback loop as an arrow diagram. PULSE turns that diagram into a typed manifest: five canonical phase kinds that any product in the ecosystem declares, so a benchmark like PRISM can read the loop directly instead of guessing at it.

retrieve route act learn consol-idate
  • 1

    retrieve embedding sim + spreading activation

    Pull relevant prior context before reasoning — the read loop. “Don’t skip retrieval; prior context likely exists.”

  • 2

    route Tarjan SCC · κ invariant

    Check topology. If knowledge is cyclic (κ>0), deliberate instead of answering reflexively.

  • 3

    act govern floor · feasible▸permitted▸best

    Mutate the world / graph — under the governance floor that decides what’s permitted.

  • 4

    learn Wilson 95% interval update

    Report the outcome back. Close the feedback loop; update confidence with real evidence.

  • 5

    consolidate merge>0.95 · decay tiers

    At loop boundaries, merge similar nodes and strengthen edges. Memory compounds.

07 · The loops underneath

A demo is a vibe. A loop is an algorithm.

The whitepaper’s diagrams stop at the arrow level. Underneath each phase of the loop is a named algorithm doing real work — graph topology, statistics, and algebra. This is where “agentic engineering” stops being a metaphor and becomes something you can prove.

route κ-routing — topology decides how hard to think

In plain terms: if the facts an agent needs all point the same direction, it answers fast; if they loop back on each other and contradict, it slows down and deliberates — and it decides which case it’s in by looking at the shape of what it knows, not by guessing. How: the route phase runs Tarjan’s strongly-connected-components — a classic linear-time graph algorithm that finds every cluster of nodes that mutually reference each other — and counts those cycles as κ, a cycle-complexity invariant. Why it matters: “think harder on hard problems” stops being a vibe and becomes a measurable property of the knowledge itself — the same input always routes the same way, and you can audit why.

κ = 0
acyclic DAG — no cycles
→ route: fast retrieval
κ = 1
one simple cycle (A→B→C→A)
→ route: deliberate · high-stakes
κ ≥ 2
nested / overlapping cycles
→ route: deliberate · strong
Try it — toggle edges, watch κ and the routing decision recompute live
Directed edges (click to add / remove):
Tarjan’s SCC runs in your browser on every click — same algorithm route uses. Close a loop (e.g. add C→A) and the routing flips from answer to deliberate; nest a second loop and κ climbs again.
The self-reducing part: when deliberation resolves a cyclic cluster it writes its conclusions back as new nodes — which lowers κ for every future retrieval. The graph gets easier to reason about the more you reason over it. That’s a learning loop the whitepaper gestures at but never names.

learn The epistemic frontier — “set the bar at the eval,” with statistics

In plain terms: the system tracks not just whether something worked but how sure it is — one success isn’t proof, ten in a row is — and it spends its next test where being more sure would help most. How: every outcome reported to learn updates a node’s confidence as a Wilson 95% score interval (z = 1.96) rather than a bare average, so the interval’s width is the uncertainty; information_gain ranks which node a single new data point would tighten the most. Why it matters: this is “set the bar at the eval” made quantitative — you don’t eval at random, you eval exactly where it pays, and confidence is earned one outcome at a time instead of asserted.

1 evidence pt~0.89
3 evidence pts~0.52
10+ evidence pts→ true
A demo shows an agent worked once — a Wilson interval roughly 0.89 wide. Reliability is a narrow interval, earned one outcome at a time, and the frontier tells you which node to gather evidence for next.

act · govern Composition you can prove, not just run

The pieces don’t just sit next to each other — they compose as algebra, and the algebra is property-tested. Capabilities combine and pipe under stated laws; an uncertified step can’t hide inside a composite.

&composeAnd — commutative, idempotent monoid on the capability lattice. Identity &none. Order- and duplicate-independent coalitions.
|>composePipe — associative, phase-graded monoid. Identity id, not commutative. Any infeasible step collapses the whole pipeline to the absorbing zero — utility can’t resurrect it.
the semiring — quantities flow as confidence = ∏, cost = Σ, latency = max. One uncertified child poisons the composite.
▣ 210 enforced · 3 declared open · 2000 random trials each · node test/laws.mjs && node test/compose-laws.mjs

08 · The bar, as algebra

“Set the bar at the eval” → a verdict with a certificate.

This is where the stack is genuinely novel. A demo shows an agent can work once; the whitepaper asks for a bar that proves it works reliably. box-and-box makes that bar a governance kernel — an eight-rung modality ladder that runs floor-then-gradient and emits an auditable certificate for every decision.

1alethicwhat can happen — the value lattice
2axiologicalhow to rank — scores & preference
3deonticwhat’s allowed — norms & obligations
4temporalsafe over time — supervision
5reflexivemay the rules change? — entrenched ring-0 core
6epistemicdo we know enough? — Wilson-interval confidence
7strategicwho can ensure it? — coalition power
8resourcecan we afford it? — affine cost ledger

One composing bridge runs feasible ▸ permitted ▸ best over a floor you cannot weaken — utility can never resurrect a forbidden branch. The output isn’t a vibe; it’s a structured verdict:

box-and-box govern decision.json → certificate
"verdict":        "allow",
"bridge":         "feasible ▸ permitted ▸ best",
"floor":          "un-weakenable",    // utility cannot override
"rungs_checked":  ["alethic","deontic","axiological"],
"margin":         7,
"confidence":     0.838,        // epistemic rung, ~95% Wilson
"cost_class":     "poly",         // resource rung, certified
"law":            "109 kernel + 101 compose, 2000 trials each",
"open_gaps":      ["CP5","CP6","CP7"]  // declared, not hidden

And because the kernel is a pure library that composes, the bricks combine: retrieve |> certify |> govern is a real pipeline (Graphonomous ▸ Weave ▸ box-and-box) — and an uncertified step collapses the whole composite to the absorbing zero. The bar is structural, not aspirational.

09 · The endgame

The orchestrator the whitepaper gestures at — wired.

The new SDLC’s far frontier is the orchestrator: hand a goal to autonomous agents and review what comes back. The stack’s north star is the dark factory — machines that perceive, learn, transfer skills to one another, and get measured. Seven steps, each owned by a shipped product, closing into a loop.

01

Perceive & act

body-browser · body-os
02

Record trace

Graphonomous
03

Learn / consolidate

Graphonomous
04

Crystallize skill

FleetPrompt
05

Ship to registry

FleetPrompt
06

Install & replay

body-* · delegatic
07

Measure

PRISM

10 · The two endgames

The factory — and its deliberate opposite.

Google’s “factory model” borrows its name from manufacturing’s dark factories — lights-out plants that run with no humans on the floor. The developer becomes an orchestrator who designs the eval, constraint, and context harness; autonomous agents build, test, and ship around the clock; humans define intent and review outcomes. The dark-factory loop above is that model, wired. But the stack also ships the deliberate inverse — a residency, not a factory.

Dimension⚙ Dark factory◴ The Residency
Metaphorlights-out manufacturingan academic / artist residency
Optimizes forthroughput — code shipped 24/7epistemic quality — trustworthy findings with provenance
The humandefines intent, reviews outcomes off-floorreads a legible board where agent “residents” deliberate
Trust posturetrust capable systemsverify the premise first (Gate 0)
Footprintmachines + orchestrator + cloudone browser tab · sovereign · zero API cost
Memorycentral episodic store (Graphonomous)browser-local CRDT (G-Set + localStorage)
Code executioncloud machines run the build directlytiered: in-browser → local gateway (TRAAVIIS) → opt-in remote sandbox
Scaling pathadd machinesP2P CRDT-merged sovereign residencies
Gate 0 — the honest question.

The whitepaper’s factory says “trust capable systems.” The Residency refuses to assume its own thesis: multi-agent deliberation can herd and amplify a confidently-wrong peer, so Gate 0 asks “does a board of residents actually beat a single agent?” — and nothing scales until that’s answered. That skepticism is the exact opposite of lights-out trust, and it’s the point.

the substrate under both

TRVM — coordination-free distributed runtime

A factory or a residency only becomes genuinely decentralized if computation can run on many machines without coordination and still agree. TRVM is that physics: a distributed interaction-calculus runtime whose confluence licenses schedule-independence (coordination-freedom across machines additionally rests on its boundary-port discipline), so results merge as a content-addressed CRDT. It’s why the Residency’s memory can be sovereign and still mergeable — and why a dark factory could be a mesh rather than a single orchestrator fanning out jobs.

interaction calculuspython · c · wasm · jsconformance battery greenlive_local
how residents actually run code

TRAAVIIS — the Hermes split-gateway

A residency of browser-native agents can write findings, but the moment a resident needs to execute — run a test, exec a script, reproduce a result — you face the credential-vs-untrusted-code problem. The Hermes pattern (Nous Research, 2026) splits it in two: a gateway that holds the tokens and orchestrates (TRAAVIIS, the OS-008 harness engineer) and a disposable sandbox that runs the untrusted code — a VM, a Mac mini, a remote micro-sandbox. Files sync back as evidence, so an execution becomes a citable, provenance-bearing finding on the board — never raw trust.

Tier 0 · in-browser
sovereign, zero infra

ic32.wasm today; Pyodide for TRVM Python, WebContainers for JS. Runs in the tab — no gateway needed.

Tier 1 · local gateway
TRAAVIIS on your machine

Holds creds, shells real CLI on a local VM / Mac mini, governs each act, streams output back as evidence.

Tier 2 · remote sandbox
opt-in escalation

Fly Sprites · E2B · Modal · Daytona — only for the autonomous distributed regime. Escalate when forced, not by default.

node ESM · zero runtime depsOS-008 harnessgateway ≠ sandboxlive_local
Honesty note. The Residency is pre-Gate-0 — its core premise (board > single agent) is not yet tested; it is a thin ~10% deliberation layer that reuses the rest of the stack, not a new platform. TRVM’s conformance battery is green live_local with two stated gaps (§6.4 snapshot, §6.5 REF). The TRAAVIIS execution gateway is decoupled now, wired later: textual findings need no code exec, so Tiers 1–2 are an architected interface, not a shipped MVP feature. Today only Tier 0 (in-browser) runs. Neither is claimed as more than it is.

11 · The human in the loop

Three ways a human stays in control — and only one of them is a bottleneck.

The whitepaper’s human is an orchestrator: define intent, set up the harness, review outcomes off-floor. That’s one posture. The stack ships three — and they differ in where the human stands relative to the work, which changes what “in the loop” actually costs.

⚙ Dark factory

Serial gate · review off-floor

The human defines intent and reviews outcomes after the agents finish — a checkpoint at the end of the pipeline. Throughput-optimal, but the human is a bottleneck: work pauses at the gate and you trust the system between gates. This is the whitepaper’s orchestrator.

post-hocserializedtrust-between-gates
◴ The Residency

Legible board · adjudicate in-process

The human reads a board where agent “residents” deliberate as it happens and adjudicates findings — claims that carry their own evidence and provenance. You don’t just see the output, you see how it became known. The human is a conductor, present but not blocking throughput.

in-processprovenance-visiblefinding-level
⟜ On TRVM

Coordination-free merge · human is a peer

The one that’s genuinely different. Interaction-net reduction is confluent by construction — Lafont’s theorem, not our claim — and that confluence licenses schedule-independence for a single net; coordination-freedom across machines additionally rests on TRVM’s boundary-port discipline. Because results are therefore schedule-independent and content-addressed, a human edit doesn’t lock the pipeline or race the agents — it merges. The human branches, edits, and the result converges deterministically (NF(A ∪ B)), exactly like any other node. No checkpoint, no bottleneck — the human is just another machine in a coordination-free mesh.

no lockCRDT-mergeconcurrent
Why TRVM changes the shape. In every conventional SDLC — vibe, structured, or agentic — human-in-the-loop is a serialization point: the pipeline waits for approval. TRVM removes the need to order the work. Because results are schedule-independent and merge as a content-addressed CRDT (resting on the boundary-port discipline), the human and the agents work concurrently on the same store and still converge. Human-in-the-loop stops being a gate you pay latency for and becomes a participant you merge with — the difference between reviewing a factory’s output and co-authoring in a shared, conflict-free workspace.

What is and isn’t proven. This is a correctness result, not a performance one. TRVM’s conformance battery is green across four implementations and swarm.js runs coordination-free across real isolated workers — but no parallel speedup has been demonstrated. “No coordination needed” is proven; “therefore faster” is not, and we don’t claim it.

SDLC parallelA second SDLC — for knowledge, not code

The whitepaper’s SDLC ships code: spec → generate → verify → deploy. The Residency runs a parallel SDLC whose artifact is a verified finding with provenance — “not just what is known, but how it became known.” Same discipline (spec-driven, verified, gated); different output.

observediscussevidencefindingexperimentbenchmarkproposalpatchverify

Findings are typed objects

claim · evidence · status · supersedes — deliberation as typed epistemic acts, not free-text chat. Structure is what makes a finding checkable, dedup-able and citable. It’s the spectrum’s “structured over vibes,” applied to knowledge.

Provenance is the deliverable

Most agent systems store what they concluded; the Residency stores how it became known — aligned to the W3C PROV model. “Evidence engineering” is the layer beyond prompt / agent / workflow engineering, which are commoditizing.

Gates over timelines

Progress advances on evidence, not a calendar: Gate 0 (does the board beat a single agent?) → Gate 1 (reliability) → Gate 2 (governed, executed improvement). Each is cheap to abandon — the epistemic checkpoint discipline PRISM automates.

Honest. This knowledge-SDLC runs Phase 0 today (board, live corpus, local-first memory, findings export). The typed-findings / dedup / citation-verification passes and the gates are the near-term build — architected, not yet shipped. Gate 0 itself is unanswered.

12 · The economics

Structure has CapEx. Vibes have a tax.

The whitepaper’s cost curves cross over: vibe coding is cheap to start and then bleeds — prompting tax, token burn, maintenance tax, security risk. Agentic engineering costs more upfront and flattens. The stack’s bet is that protocols amortize the CapEx across every product: you pay for the floor once, then reuse it.

Vibe coding — cumulative cost
Agentic engineering — cumulative cost
[&] stack — amortized floor (pay once, reuse)

Curves are conceptual, not measured — they illustrate the shape of the argument (vibe’s convex tax vs. the stack’s amortized floor), not data from a study. The one number on the page that is measured is the governance kernel: 210 enforced laws at 2000 trials each, and 3 that are declared open — reproducible below.

13 · For the builder

If you were designing this system yourself — here is what you'd actually write, decide, and ship.

No diagrams this time. The three things an engineer asks before adopting a stack: What are the actual files? Which design choices are worth stealing? And what's real versus aspirational?

A · ArtifactsThe files you'd actually write

research-agent.ampersand.json
// structural composition — [&]
{
  "agent": "ResearchAgent",
  "capabilities": {
    "&memory.vector": {
      "provider": "pgvector" },
    "&reason.argument": {
      "provider": "deliberatic",
      "config": {
        "governance":
          "evidence-first" } }
  },
  "governance": {
    "hard": ["Never present unsupported
      conclusions as facts"],
    "escalate_when": {
      "confidence_below": 0.65 } },
  "provenance": true
}
Capabilities are interfaces; providers are swappable implementations. Change pgvectorqdrant without touching the agent.
graphonomous.continual_learning.pulse.json
// temporal loop — PULSE
{
  "phases": [
    { "kind": "retrieve" },
    { "kind": "route",
      "invariant": "kappa_routing" },
    { "kind": "act",
      "policy": "delegatic.check" },
    { "kind": "learn" },
    { "kind": "consolidate" }
  ],
  "nesting": {
    "inner_loops": [{ "trigger":
      "phase:route when kappa > 0" }] },
  "connections": [{
    "emit": "OutcomeSignal",
    "to": "prism.benchmark" }]
}
PRISM reads this file directly to benchmark you — it discovers your phases, invariants and substrates from the manifest. No adapter to write.
compile to runtime · CLI
# validate against the schema
$ ./ampersand validate \
    research-agent.ampersand.json
  ✓ valid · 2 capabilities

# compile to MCP server config
$ ./ampersand generate mcp \
    research-agent.ampersand.json
  → mcp.servers.json

# compile to A2A agent card
$ ./ampersand generate a2a \
    research-agent.ampersand.json
  → agent-card.json
[&] doesn't replace MCP or A2A — it compiles into them. The declaration is the source of truth; the wire formats are generated targets.

B · RationaleDesign decisions worth stealing

Graph over vectors

Topology is a routing signal. The κ cycle-complexity of a subgraph tells you whether to answer fast (κ=0, acyclic) or deliberate (κ≥1, cyclic). A flat vector store throws that signal away.

Property laws over example tests

210 laws × 2000 random trials each beat hand-picked fixtures. You assert "& is commutative and idempotent" once and the generator hunts the counterexample you'd never think to write.

Deterministic servers over LLM-in-the-loop

MCP servers never call a model — the agent does all reasoning; servers are a deterministic substrate. Reproducible, cheap, testable. The intelligence lives in one place, not smeared across services.

Floor-then-gradient over weighted scoring

Safety can't be averaged away by a high utility score, and utility can't resurrect a forbidden branch. feasible ▸ permitted ▸ best over an un-weakenable floor — every verdict ships a certificate.

Protocols over monolith

Bring-your-own-runtime. Swap any layer — memory, reasoning, governance — and the rest still composes. PRISM is designed to benchmark anyone’s system, because it reads the PULSE manifest rather than your internals — but that is a claim about the interface, not evidence: no system we don’t own has been adapted in yet. External adaptation is the next proof, not a past one.

Local-first over cloud-by-default

Browser-local CRDT board + a local harness = zero per-token API cost and full data sovereignty. Cloud is an opt-in sink you sync to, not a dependency you boot from.

C · HonestyMaturity — what's real, what isn't

LayerStatusWhat's realHonest gap
[&] Protocolin_treeSchema, validator, MCP/A2A codegen, Elixir reference impl + CLI.&body sensorimotor primitive has no runtime yet.
Graphonomouslive_deployedκ-routing, Wilson frontiers, consolidation, MCP server, embedded SQLite.Cloud sync is optional / single-node by default.
box-and-boxlive_localv0.10.0. 8-rung ladder, bridge, 210 enforced laws, certificates. Cross-process admission proven: the Node kernel decides, an Elixir BEAM actor acts only on permission, certificate retained to disk (13/13 assertions, permit + refuse paths).3 compose laws declared open: the Value.pi carrier (CP5–CP7). The coalition contract (CD6) was closed by ruling. Not npm-published at 1.0. Not deployed.
PULSElive_deployedv0.1.1 schema, 6 cross-loop tokens, reference manifests.Manifest authoring is hand-written, no generator.
PRISMlive_deployed6-machine benchmark engine, reads PULSE manifests, Fly.io deploy.No external/competitor system adapted in yet.
Delegaticlive_deployedPolicy checks gating the act phase, evidence-first governance.In-memory ETS only — no persisted policy store.
OS-008 Harnessin_treePipeline enforcement, quality gates, sprint contracts (TRAAVIIS CLI).Orchestration loop is docs + site, not one binary.
Dark-factory looplive_local7/7 steps run locally end-to-end.Not yet stitched across deployed machines.
WRLlive_localWallRiderLang Core 0.1.2 — an executable topology language whose meaning is a hash. 890 conformance checks, 0 failed; register 128 rows, model debt 0. Cross-implementation agreement is measured, not asserted: projection vectors round-trip byte-exact through Forge’s independent Python spine, and on a shared negative corpus both verifiers refuse all 15 tampers and agree on the name of every refusal — two languages, two repos.The suite reports its own limit: 54 of the swept blocks are fragments, and a fragment’s non-acceptance is weak evidence until its capability ships. Not deployed; no outside consumer has driven the path.
TRVMlive_localCoordination-free interaction-calculus runtime, conformance battery green across python · c · wasm · js.§6.4 snapshot & §6.5 REF semantics unspecified. No demonstrated parallel speedup — the result proven is correctness, not throughput.
Workbench /recordlive_localA real Claude Code transcript → scored SkillBundle, no model call and no API key. 210 MB indexed in 936 ms; every segment sealed twice — honest → 6/6 PASS, one capability withheld → no_hidden_capability + authority both FAIL.Cloudflare Pages is a target, not a deploy. Gates shown to carry signal on n=1 corpus.
the-residencylive_localBrowser-native CRDT board, local harness, live corpus hydration.Pre-Gate-0: the premise (board > single agent) is untested.
OS-007specSpecification drafted.Unbuilt.
Adopt incrementally. Start with one *.ampersand.json → add a *.pulse.json when you have a loop → point PRISM at the manifest when you want numbers → drop in box-and-box only where a decision must be governed. No big-bang rewrite. Each layer is useful alone.

D · VerifyDon’t trust the numbers — run them

Every count on this page is a command away. These are the actual suites; the green lines are the real output captured from this repo, not a mock-up.

governance · 210 enforced · 3 open
# the un-weakenable floor, proven
$ cd AmpersandBoxDesign/box-and-box
$ node test/laws.mjs \
    && node test/compose-laws.mjs
  ✓ all 109 enforced kernel laws hold.
known gaps (xfail):
  ✗ CP5 |> floor association-invariant
       — FALSIFIED @trial n
  ✗ CP6 no backward step survives |>
       — FALSIFIED @trial n
  ✗ CP7 &-order leaks to |> floor
       — FALSIFIED @trial n
  ✓ all 101 enforced CC2 compose laws
     hold — 100 suite + 1 anchor.
  · 3 known gap(s) still open
# 109 kernel + 101 compose = 210
# 2000 random trials each
Yes, the red is supposed to be there. (@trial n because the trial index is different every run — the inputs are random, so your counterexample won’t be ours.) Three compose laws are declared open — a single-slot Value.pi carrier lets |> re-association bypass the phase floor. They were found by the residency falsifier, they are written down in EVIDENCE/algebra-finding.md, and the build fails if one starts passing — that's the signal the carrier fix landed. A suite that only ever prints green isn't being run hard enough.
[&] · validate + compile
# spec → schema → wire formats
$ cd AmpersandBoxDesign/reference\
    /elixir/ampersand_core
$ mix deps.get && mix escript.build
$ ./ampersand validate \
    ../../../examples/\
    research-agent.ampersand.json
  → schema-valid · 2 capabilities
$ ./ampersand generate mcp <file>
  → mcp.servers.json
The same declaration the compile demo above renders — validated against draft-2020-12 and compiled to MCP/A2A.
graphonomous · engine tests
# continual-learning graph + MCP
$ cd graphonomous
$ mix compile --warnings-as-errors
$ mix test
$ mix format --check-formatted
# κ-routing · Wilson frontiers ·
# consolidation · embedded SQLite
Elixir/OTP. The κ, Wilson-interval, and consolidation algorithms from §06 are exercised here.
Open source. Several components are public repositories you can clone today — github.com/c-u-l8er/TRVM · the-residency · TRAAVIIS · ampersand-plugins (the Claude Code skills) · docs. The reproduction commands above run from the [&] monorepo. Verifiability is the whole point: if a count can’t be reproduced, treat it as a claim, not a fact.

E · InteropWhat we compile into — and who else is in this space

[&] has one rule that has kept it useful: it does not invent a wire format. It compiles into MCP and A2A rather than competing with them. The same rule applies one layer up — the evidence layer already has standards, and the right move is to emit them, not to replace them. And this is not an empty field: as of 2026 it is filling quickly, which we take as the category being real rather than as a threat.

MCP · A2A

Not competitors. ampersand generate mcp and generate a2a emit both from one .ampersand.json. A capability declaration is upstream of a wire protocol.

OpenTelemetry GenAI

The GenAI SIG is standardising agent spans, events and metrics. A PULSE phase is a natural span; the honest status is that we don’t emit OTel yet. It is the most obvious next interop win and it isn’t built.

W3C PROV

The Residency’s findings carry provenance aligned to the PROV model — entities, activities, agents. Evidence that can’t be pointed at a standard shape is evidence nobody else can consume.

NIST · W3C · CoSAI

NIST’s AI Agent Standards Initiative (Feb 2026), the W3C Agent Identity Registry CG (Apr 2026) and CoSAI’s Agentic IAM are all now working the identity/authorization layer that Delegatic occupies.

Runtime governance kits

Microsoft’s Agent Governance Toolkit, OPAQUE’s verifiable-governance standard, AOS Governance. AOS’s framing — a script, not a prompt, decides whether an action is permitted — is close to box-and-box’s. The difference we’d defend is the algebra: a ladder with stated laws that compose, not a policy checker.

The literature

Deontic policies for runtime governance, evidence-tracing/execution-provenance surveys, verifiable context governance. Our deontic rung is one rung of eight — the papers are neighbours, not rivals, and they’re ahead of us on formal write-up.

Where we’re behind. Everything above is more mature than us on adoption and most of it on publication. What the [&] stack has that we haven’t found elsewhere is the composition story — one floor, stated laws, verdicts that carry certificates, and protocols that plug into each other rather than into a single vendor’s console. That is a claim about shape, and it is the one most worth attacking.

14 · What stays true

Three durable principles — and our addendum to each.

Principle 1

Structure scales, vibes don’t.

Past a crossover point, ad-hoc prompting costs 3–10× more per feature.

+ So we made structure cheap to reuse: one governance floor, 210 laws, composed by every product.
Principle 2

AI amplifies your culture.

It multiplies both strengths and weaknesses of your existing engineering discipline.

+ So we encode the discipline as protocols — the multiplier acts on something machine-checkable, not tribal knowledge.
Principle 3

The human role evolves.

Toward specification, evaluation, architectural judgment — not away from engineering.

+ So the human writes the .ampersand.json and sets the floor; the kernel proves the rest and ships the certificate.
Google wrote the field guide to agentic engineering. The [&] stack is one attempt to build the instruments — a specification algebra, a memory graph, a temporal loop manifest, a benchmark, and a governance kernel that turns “set the bar at the eval” into a verdict you can audit. — ProjectAmp2 · the [&] Protocol ecosystem

15 · The ending this page has to have

Here is precisely how to prove us wrong.

A page arguing that verification beats demonstration cannot end by asking you to be impressed. So it ends with a kill switch for each claim. Every row below names the claim, its honest status, the command that exercises it, what you should see, and the specific observation that would falsify it. If you produce one of the things in the last column, the claim is dead and we want the counterexample.

Absence of evidence must remain distinguishable from evidence of absence, and every consequential transformation must expose why it occurred. — the [&] doctrine, as stated in ACADEMY.md §6e

That rule is why this table exists in the shape it does, and it wasn’t designed for this page — it was noticed, after the same shape kept appearing in unrelated subsystems: a generator that refuses and says why, an unrun arm that reports None rather than 0, a panel that carries a mock edge. Both halves do work here. The first is why the row below reads 0 filed instead of being left out — a zero and a blank look identical in a table and mean opposite things. The second is why the law counts are now derived by the suites rather than typed by us, and why the three open laws print in red.

It already runs somewhere public. The Academy refusal log publishes the pages the article generator declined to write and the reason for each: material insufficient, process undocumented, no questions in the corpus to answer. One entry is marked HELD rather than refused because the corpus contradicted itself on a law count — 103 vs 116. That hold is the reason this page now says 210 enforced and 3 declared-open: the refusal named a real defect, and it stayed unpublished until the suites were made to derive their own totals. The log itself is still marked MOCK — the real format with illustrative entries, counts not yet measured against a live crawl. Saying so is the same rule applied to the rule.
ClaimStatusCommandExpectedWhat kills it
The governance floor is un-weakenable and its laws hold live_local node test/laws.mjs
node test/compose-laws.mjs
109 + 101 = 210 enforced green, 3 declared-open in red, exit 0 A counterexample to any of the 210 — or an input where the floor is bypassed without one of CP5/CP6/CP7 firing.
Governance gates a foreign runtime, not just itself live_local bash test/integration/
kernel-gates-action.sh
13/13 — the Node kernel refuses, the Elixir actor does not act, certificate on disk The BEAM actor performing the act on a refuse verdict, or a permit with no retained certificate.
A spec compiles to real wire formats in_tree ./ampersand generate mcp
<file>
Schema-valid input → an mcp.servers.json a client actually accepts Output that no MCP client will load, or a valid declaration the generator can’t compile.
Proof gates carry signal — they don’t just say PASS live_local Workbench /record, drop in your own Claude Code transcript Honest run → 6/6 PASS; withhold one capability → no_hidden_capability + authority FAIL A transcript with a genuinely concealed capability that still seals 6/6. That’s the interesting attack.
A world means the same thing in two independent implementations live_local node test/conformance.mjs
+ the shared negative corpus
890 pass / 0 fail; projection vectors round-trip byte-exact through Forge’s Python spine; both verifiers refuse all 15 tampers and name each refusal identically One implementation accepting a world the other refuses — or the two refusing the same tamper under different names. Either means the meaning isn’t really the hash.
Reduction is schedule-independent live_local make test · node runtime/js/swarm.js Same normal form across python · c · wasm · js; workers agree with no coordination Two schedules reaching different normal forms on the same net. (Note we claim correctness, not speedup — a slow run falsifies nothing.)
The board beats a single agent pre-Gate-0 Nothing. This is untested. Already open — Gate 0 exists precisely because deliberation can herd toward a confidently-wrong peer. We have n=1.
The derivation reproduces on a machine we don’t own external · 0 filed Boot a T&R image, then trvs doctor · trvs id The quoted sha256 matches, and trvs id derives the same identity we published — on hardware we have never touched A sha that doesn’t match, an identity that comes out different, or an image that won’t boot. This is the only row whose missing evidence can only come from a machine we don’t own — which is exactly why the count is 0 reports, 0 from real hardware, as of 11 August 2026.
PRISM benchmarks anyone’s system design claim Point it at a PULSE manifest you wrote Runtime-independent by construction — it reads phases from the manifest Nothing yet, and that’s the problem: no system we don’t own has been adapted in. Until one is, treat this row as design, not evidence.

Where this argument livesComputeDriven

The table above is one page’s worth of a thesis that has its own home. ComputeDriven (computedriven.com) states it in a sentence this page has effectively been circling for fifteen sections:

Data driven keeps the result. Compute driven ships the derivation.

A claim is compute driven when the artifact carries enough — canonical input, executable semantics, a derived identity and its provenance — for an independent machine to derive the claim again. — computedriven.com

That is the same move as &-composition, the certificate, the PULSE manifest and the 890 cross-language conformance checks — stated once, at the level of the artifact rather than the tool. It also carries the caveat this page needs and doesn’t otherwise say out loud:

Re-derivation proves fidelity, not correctness. A wrong computation re-runs perfectly and is still wrong. Everything in the table above establishes that a result was faithfully produced — not that it was the right result to want.

Its first proof object is T&R, a FreeBSD 15 distribution assembled from pkgbase and published as releases with quoted sha256 sums — an artifact rather than another protocol, which is why it appears here as a row in the table and not as a section of its own. And the boot report is the mechanism: ComputeDriven’s admitted last gap is that every measurement was taken on our machines, so joining it is defined as supplying the one thing we cannot supply ourselves — a second machine.

Most of this field asks you to trust a benchmark. The bar we’d rather be held to is the one in the last column — and the fastest way to help is to break a row. — if you falsify one, the counterexample belongs in the repo, not in a rebuttal
The rows we can’t yet write. Three claims on this page have no falsification command at all, and honesty means listing them rather than quietly omitting them: the TCO curves (conceptual, not measured), the dark factory running end-to-end across deployed machines, and &body having a runtime. External reproduction used to be a fourth; it now has a command and a kill condition — the boot-report row — though a mechanism is not evidence and the count is still zero. Adoption remains a separate open claim, and one boot report would not close it: reproducing an artifact is not the same as depending on it. Someone building on this, integrating it, or using it for real work is a different category of proof, and we have none of it. A claim with no kill switch is a claim with no evidence — including ours.