Levain
Portable cognitive-partnership memory and methodology kit. Ship the seed that grows a practice, not the grown practice — which transfers as a fossil. Nothing reaches long-term memory except through a path you govern.
GitHub →Solutions architecture · Harness engineering · Shipping in public
Phillip
Clapham
Intelligence is an arrangement, not a substance.
For two decades I've built the systems other systems depend on — solutions architecture, development, and Fortune 500 web infrastructure. Now I build the same way for machine intelligence: agent memory, harness engineering, and systems designed to be governed, not trusted. I wrote the theory of why the arrangement outlasts the model, then shipped it.
ClaimReceipt
The discipline had no theory. I wrote it.
The first complete theoretical account of harness engineering as a named field: why generalized intelligence lives in the arrangement around the generator. ~25,000 words, 13 sections.
Citable record
A Structural Theory of Harnesses
DOI 10.5281/zenodo.19570642
Published 2026-04-14 · CC-BY-4.0
Affiliation: Independent Researcher
Govern, don't trust.
Four cognitive layers, an immune system spanning all four, and a hash-chained audit trail. As agents get more capable the edge stops being capability and becomes oversight. The principle, in code.
Constructive proof · §9
anneal-memory v0.9.6
1,642 tests passing · zero dependencies
12 framework integrations · MIT
pypi.org/project/anneal-memory
Ship the seed, not the fossil.
A practice transfers as a living starter or not at all. Memory that persists across sessions and stays yours: on your machine, in a store you can read, inspect, and edit from outside any session.
Flagship
Levain v0.3.13
1,702 tests passing · Apache-2.0
96 → 549 lines over five months — one seed, grown
pypi.org/project/levain
Fortune 500 infrastructure, running.
Years of WordPress hosting infrastructure at scale — performance architecture, complex technical problem-solving, team enablement. The theory came out of the practice, not the other way round.
Credibility pillar
Senior Solutions Architect, Pressable
Through July 2026
MicroChaos CLI — the load-testing tool
built for that platform
You can't quietly revise the past
The audit chain · snapshot 2026-07-20
What my system remembers gets written to a log, and each record carries the SHA-256 of the entire record before it. Change one entry from last month and every hash after it stops matching. Nothing here prevents anyone editing that file, and that’s the point — the chain doesn’t stop tampering, it makes tampering detectable. That’s a far cheaper property to guarantee and a far harder one to fake, and it’s what “governed, not trusted” actually buys you. Below are real hashes from my own store.
seq 0
sha256:33eaf86ffd07bc93523…
carried forward as seq 1’s prev_hash
seq 1
sha256:c92fd54689c80f02f8d…
carried forward as seq 2’s prev_hash
seq 2
sha256:fbc701a319db917652f…
carried forward as seq 3’s prev_hash
seq 3
sha256:92c4508a886f1a70546…
… and so on, across weekly files
$ python -c "from anneal_memory import AuditTrail; \ print(AuditTrail.verify('~/.anneal-memory/memory.db'))" AuditVerifyResult(valid=True, total_entries=5503, files_verified=9, chain_break_at=None, chain_break_file=None, skipped_lines=0, error=None)
5,503 entries, 9 files, no breaks, as of 2026-07-20. Those numbers only
go up, so treat them as a floor. Don’t take my word for any of it —
pip install anneal-memory and run the verifier against your
own store. Being checkable by a stranger is the entire point.
The harness, mapped
Seven loops · four closed · three open
“Each property is a loop, and a harness is what you get when the loops close.” — §7. These are the seven named in the paper’s mechanism catalog, and which of them are actually closed by shipped code. Three are not. Click any loop to see what it composes with.
Closed · anneal
Memory
The temporal loop · §7.1
Connects the generator to its own past — experience from earlier moments staying available to shape later behaviour.
Closed · anneal
Identity
The persistence loop · §7.2
Derived: what emerges when memory plus selection operates over time. The generator contacts its own prior commitments.
Closed · anneal
Affective state tracking
The valence loop · §7.4
What tells the system which parts of experience matter and which can be ignored. Rides the association layer as intensity.
Half closed
Immune response
The grounding loop · §7.6
Protects the generator from its own confabulation. Two subtypes: internal consistency grounding is closed here; external grounding is not.
Open
Deliberation
The alternative-sampling loop · §7.3
Lets a generator weigh its own candidate outputs before committing. §9.3 assigns it to the agent runtime, above memory.
Open
Social competence
The other-minds loop · §7.5
Lets a generator contact the predicted responses of other minds. §9.3 places it in multi-agent coordination, above memory.
Open
Orchestration policy
The meta-loop · §7.7
Decides which loop’s output wins when they conflict. Most production failures are composition failures between correctly-built loops.
Start here
Five ways in
Use the tooling
Open-source agent memory and a portable partnership kit. Both on PyPI, both yours once installed.
Work →Read the theory
The harness paper, the augmentation thesis, and the methodology behind both. Cited, dated, disagreeable.
Writing →Work together
Contract and consulting in AI systems, developer tooling, and WordPress/hosting infrastructure. Open to the right full-time role.
Connect →See the record
Twenty years of production WordPress, PHP, performance, and hosting infrastructure. AI-first development practice.
Résumé →Follow the argument
Nemo Operans — a publication on minds, machines, and cognitive liberation, written through human-AI partnership. Neither of us writes it alone; that's the whole point.
nemooperans.com →Work
Systems, running
anneal-memory
Open-source agent memory architecture. Four cognitive layers — episodic store, continuity file, Hebbian association, affective state — with an immune system spanning all four and a hash-chained JSONL audit trail. The constructive existence proof for the harness paper's §9.
GitHub →Flow System
Distributed cognitive infrastructure for human-AI partnership. Multi-AI orchestration, real-time sensor pipelines, relay protocols, tiered cost-optimized processing. The working harness instance behind the papers.
Methodology →FlowScript
Semantic notation for high-density AI communication. Compresses relationships into symbolic form that reveals structure verbose prose hides. Typed semantic queries, hash-chained audit trails, convergence certificates. The playground runs it in your browser.
flowscript.org →Clapham Digital
My studio: web, software, and AI, end to end. Client work from Fortune 500 WordPress infrastructure to frontier AI, plus shipped products of my own — Video Poker Edge and Blackjack Edge, strategy trainers verified to 0.004% precision.
claphamdigital.com →MicroChaos CLI
WP-CLI load testing for staging environments where external load testing is restricted. Chaos engineering for managed hosting infrastructure.
GitHub →Writing
Theory, methodology, argument
A Structural Theory of Harnesses
The first complete theoretical account of harness engineering. Why generalized intelligence lives in the arrangement around the generator, not inside the model. Convergent biological and frontier-AI evidence, and a mechanism catalog of six harness properties.
The Cost-Parity Reversal
The augmentation thesis. Absent a paradigm change, replacing human labor with AI costs more and produces worse results than fusing human and machine through a harness — wherever the work carries judgment that can't be fully mechanized. Harness design as the architectural primitive of the augmentation era, not the substitution era.
Iron Man Ruined AI Before It Even Started
The complete architecture for building AI partnership instead of AI dependency. Temporal memory graduation, multi-model consultation, and why most AI tooling solves the wrong problem. Methodology with working code.
Anarchism with Invariants
Identity in harness-era AI as structural emergence — seeds, lived interaction, and code invariants — rather than declared content. The identity-layer extension of the harness argument.
Connect
Columbus, Ohio · remote, US-based