hermes-brain/ideas/passepartout-economics.org

#+TITLE: Passepartout — Patents, Moats, Economics, Design Implications
#+AUTHOR: Hermes agent distillation of 2026-05-21 discussion with Amr
#+FILETAGS: :passepartout:agent:economics:ip:licensing:
#+STARTUP: content

* Summary

Discussion about the economic and strategic implications of Passepartout's
architecture — a self-bootstrapping agent that combines deterministic safety
gates (0 LLM tokens per verification), Merkle-tree memory with provenance,
a symbolic fact store with sufficiency criterion, and ACL2-based macro layer
bootstrapping for provable reasoning.

The central claim: this architecture decouples intelligence from LLM API
consumption. The probabilistic engine (LLM) handles ~10% input/output
translation; the symbolic engine handles ~80% of reasoning at near-zero
marginal cost. The cost curve inverts: generation is expensive, verification
is cheap.

* Patentability

** Likely patentable

- **Probabilistic-deterministic split with deterministic gates between LLM
  proposal and execution.** The LLM proposes, the gate stack decides. Each
  gate is a pure Lisp function costing 0 LLM tokens. Every competitor uses
  prompt-based guardrails. The specific 11-vector gate stack (secret
  exposure, path protection, self-build boundary, shell safety, network
  exfiltration, privacy tags, Lisp syntax, credential vault, tool permissions,
  policy, protocol validation) is a specific novel implementation.

- **Foveal-peripheral context model with Org-tree structured retrieval.**
  Depth ≤ 2 always; full render on foveal node; full render on semantic
  similarity to foveal; full render on temporal relevance (modified today,
  upcoming deadlines); everything else title-only. Targets 2,000-4,000 tokens.
  No agent does this.

- **Merkle-tree memory with copy-on-write snapshots and operation-level
  undo/redo.** Every memory-object is content-addressed. Snapshots are
  deep-copies. Undo/redo at the individual operation level. Applied to an
  agent's reasoning loop.

- **Gate-to-fact bootstrap with sufficiency criterion.** Mechanically
  extracting facts from the gate stack's own data structures (protected paths,
  shell blocked patterns, network whitelist) as the seed of an ontology. A
  measurable sufficiency threshold that flips the system from LLM-proposes
  to Screamer-deduces.

- **Macro-layer-as-skill bootstrapping architecture.** Encoding theorem-proving
  capability as hot-reloadable skills where each layer is verified by the layer
  below. The proof forest is a Merkle-versioned dependency tree.

** Likely not patentable (known techniques in expected applications)

- ACL2 itself (decades old)
- Screamer for consistency checking (constraint solving on a triple store is
  an obvious application)
- Hot-reloadable skills (Lisp images have been hot-reloadable for 40 years)
- Org-mode as a data format
- Multi-layer signal authentication (known in network security)

** Counterargument from prior art

A patent examiner will argue that:
- "Thin harness, fat skills" is the standard OS microkernel architecture
  applied to an AI agent
- Foveal-peripheral context is locality of reference (standard in OS design)
- Merkle-tree memory is content-addressed storage (standard in distributed
  systems)
- Deterministic gate stack is capability-based security (going back to
  KeyKOS in the 1980s)

The defense: these principles have never been *combined* in an AI agent, and
the combination produces emergent effects (cost curve inversion, sufficiency
flip, self-repairing bootstrapping chain) that no single principle produces
alone. Good patent claims would cover the specific combination, not the
individual components.

** Strongest single claim

An AI agent system comprising:
1. A probabilistic language model
2. A stack of deterministic safety gates operating at zero LLM-token cost
   between the model's proposal and execution
3. A Merkle-versioned memory store from which gate outcomes are mechanically
   extracted as facts
4. A symbolic reasoning engine seeded by those facts with a measurable
   sufficiency criterion that determines when the probabilistic model can
   be bypassed

Each element is known. The combination is novel and non-obvious.

* Licensing Strategy

** AGPLv3 for the public repository

AGPLv3 closes the ASP loophole (Section 13): anyone who modifies the
software and offers it over a network must release their modified source.
This protects against proprietary forks that extract value without
contributing back.

Crucially: AGPL is a *product requirement*, not a concession to openness.
The system's value proposition is provable correctness — every decision has
Merkle provenance, the proof forest is visible, the sufficiency meter is
readable. This claim is structurally incredible with closed source. An
enterprise buyer needs to inspect the gate stack, verify the Merkle
implementation, and confirm ACL2 integration is sound. AGPL makes this
possible without signing an NDA.

** AGPL only covers modifications to code, not:

- Gate rules specific to a domain (these are data, not code)
- The fact store (empirical data generated from usage)
- Ontology categories (design decisions stored as configuration)
- Proprietary skills loaded at runtime (AGPL boundary on plugin systems
  is legally unsettled)

** Dual license model

- AGPLv3 for open source — builds ecosystem, trust, and community
- Commercial license for enterprises that cannot accept AGPL (blanket
  policies against AGPL infection) — MySQL/SugarCRM/GraphQL model

* Moats

** Re-evaluated: time is not the primary moat

Initial assumption: the bootstrapping chain (gate outcomes → facts →
Screamer rules → ACL2 theorems → macro layers) takes months to build,
giving first-mover advantage.

Challenge: a Phase 4+ Passepartout fed on Wikipedia + Wikidata can build
a general ontology in two weeks. Entity resolution is batch work. Structural
consistency verification is minutes. The organic growth advantage collapses
for general knowledge.

** Actual moats (weaker than initially assumed)

1. **Domain-specific gate rules** — thin. A few hundred lines of Lisp data
   encoding deployment-specific path patterns, shell safety rules, and
   volume layouts. Write once, trivial to copy. Not a real moat.

2. **Empirical decision history** — every HITL decision is a Merkle fact.
   "On date T, user approved action X under context Y." A fresh instance
   has none of this. Makes *your* instance more valuable but doesn't
   prevent competition — it's a switching cost, not a barrier to entry.

3. **Evaluation harness (regression suite)** — thousands of test cases
   accumulated from every bug fix. Cannot be ingested from public data.
   Built only by using the system, breaking it, fixing it, and adding a
   test. Strongest residual moat, but even this can be partially
   compressed through public benchmarks (SWE-bench, etc.).

4. **Infrastructure integration** — the specific Docker compose layouts,
   Traefik router patterns, Authentik provider configurations, backup
   policies encoded as gate rules over months of use. A competitor's
   infrastructure is different; their generic Passepartout does not know
   your topology.

** Strongest competitor strategy

Not copying your gate rules — offering the same architecture as a service
with their own pre-seeded general knowledge, a generic safety baseline,
and a consulting engagement to customize gate rules for each customer.
The AGPL prevents closing the architecture but does not prevent offering
it as a service with a customization layer.

** The defensible business is services, not product

The defensible entity is "the organization that best understands how to
adapt Passepartout to your domain" — not "the organization that owns
Passepartout." The Lisp Machine appliance (hardware + certification) and
evaluation harness certification service are the closest thing to product
defensibility.

* Economics and Monetization

** Cost structure

- One-time cost: gate-rule encoding for a domain (from hours for codified
  domains — FAR, HIPAA, ISO standards — up to months for tacit domains)
- The LLM translates codified rules directly: ingest regulation → produce
  gate rule plist → ACL2 verifies consistency → human reviews. This is
  translation, not reasoning.
- For non-codified knowledge (craft expertise, organizational culture):
  Phase 3 archivist loop over time
- Near-zero marginal cost: ACL2 proof + Screamer consistency check +
  VivaceGraph lookup per interaction — all CPU-native, all in-image
- No recurring LLM API costs for the 80% symbolic reasoning layer
- After sufficiency flip: pennies per day vs dollars per day for LLM-only

** Revenue models by field

| Field | Why Passepartout | Revenue Model |
|-------+------------------+---------------|
| Industrial infrastructure (refineries, power grids, manufacturing) | Offline operation, provably safe, near-zero marginal cost, mandatory audit trail | Lisp Machine appliance + SCADA certification package |
| Healthcare administration (billing, claims, prior authorization) | Rule-heavy domain, privacy-mandated, audit-driven, high per-transaction cost today | Subscription for regulatory gate packages (CPT/ICD-10/HIPAA rules), updated when CMS publishes new rules |
| Software supply chain (CI/CD security, SBOM verification) | First-order structural verification — ACL2 is natural fit, CI/CD pipeline is already a sequence of gate-checkable steps | Evaluation harness as certification service — "run our 10,000-task suite and get a provable score" |
| Regulatory compliance (GDPR, SOC2, SOX, GxP) | Rule-completeness, active enforcement (not document-based), provable audit trail | Subscription for regulation-specific gate packages — GDPR package, SOC2 package, FedRAMP package, updated when regulations change |
| Defense and classified environments | Air-gapped operation, classification-level gate rules, Merkle provenance is court-admissible evidence | Government contract + hardened appliance with hardware root of trust |

** Critical insight: encoding cost drops to near-zero for codified domains **

Laws, regulations, standards, procedures, and technical specifications are
already written down in structured text. The LLM does not need to *reason*
about them — it needs to *translate* them into gate rules and ACL2 theorems.

Example: The US Federal Acquisition Regulation (FAR) is ~2,000 pages of
"thou shalt" and "thou shalt not" statements. A frontier LLM can ingest
the FAR and produce a plist of gate rules:
- (if contract > $250K AND not small-business-set-aside → :deny)
- (if sole-source AND no justification-documented → :deny, produce-justification)

ACL2 then verifies the rule set for internal consistency (Phase 6). Screamer
checks against existing compliance facts. The human reviews the bootstrap
output and approves or corrects individual rules.

The key distinction: the LLM is not *extracting knowledge from prose* in the
way Phase 3 archivist does (which is open-ended, noisy, requires grounding).
It is *translating a known rule system into a formal representation* — a
mechanical transformation of structured text into structured rules. The
result is not "the LLM's best guess at the rules" but "the rule set as
stated in the source document, mechanically transcribed."

For domains where the knowledge is codified as text, the gate-rule encoding
time drops from weeks to hours. The only bottleneck is human review of the
output — and the system can assist here by surfacing contradictions for
resolution rather than requiring a full line-by-line audit.

** What can actually be monetized (TLDR)

1. **Pre-loaded bootstrapping chains for specific verticals** — domain gate
   rules, pre-seeded fact stores, mature proof forests. Saves the buyer
   months of bootstrapping. Distributed as data packages under commercial
   license, not AGPL.

2. **Evaluation harness as certification service** — "Bring your agent,
   we'll run it through our suite and give a Merkle-verified score."
   The regression suite grows with every deployment; a competitor's
   regression suite starts empty.

3. **Hardened Lisp Machine appliance** — RISC-V soft-core with Lisp
   microcode, pre-loaded mature Passepartout, certified for specific
   verticals (IEC 62443 for industrial, HIPAA for healthcare). Value is
   in integration and certification, not the AGPL software.

4. **Verified skill marketplace** — marketplace where skills are verified
   (sandbox + ACL2 non-contradiction proof) before listing. Marketplace
   takes a cut. Value is in the verification infrastructure, not the
   skills themselves.

5. **Support and consulting** — the Red Hat model. AGPL code is free;
   training, custom gate rules, ontology design, and emergency support
   are paid.

* Design and Architectural Implications

** The self-improving system

Passepartout bootstraps two feedback loops:

- **Empirical loop:** gate outcomes → facts → Screamer-verified patterns →
  sufficiency flip → auto-extraction. Knowledge grows without the LLM
  touching most of it.

- **Logical loop:** ACL2 theorems → macro layers (generators, metafunctions,
  induction DSL, abstract theories) → richer proof strategies → better
  verification. Reasoning capacity grows without changing the prover binary.

These loops intersect at the fact store: proven theorems become facts, richer
facts generate better proof strategies, better strategies verify more facts.
The system upgrades itself.

** The 10-80-10 becomes approximately true

- 10%: LLM handles input translation (natural language → structured goal)
  and output formatting (structured result → natural language)
- 80%: Symbolic engine handles reasoning — Screamer plans, ACL2 verifies,
  VivaceGraph retrieves facts. Zero LLM tokens.
- The cost curve inverts: verification is cheaper than generation.

** Key implications

1. **Verification becomes cheaper than generation.** Once macro layers are
   mature, proving a new rule non-contradictory costs near-zero. The LLM
   proposes; the symbolic engine accepts or rejects.

2. **Trust scales with use.** Every interaction produces a structurally
   verified outcome. Non-lossy fact base grows. Proof forest thickens. An
   auditor can inspect the Merkle tree of gate outcomes and trace any
   decision to its root theorem.

3. **Degradation is reversible.** Every proof layer is a hot-reloadable
   skill. Every fact has provenance. A bad metafunction is unloaded;
   theorems proven under it are flagged for re-verification; the fact
   store retains the pre-upgrade ontology version.

4. **The system can diagnose its own logical frontier.** If ACL2 keeps
   failing on a class of properties, and the failure mode is structural
   (not solvable by more macros), the fact store accumulates a pattern:
   "These N properties are first-order inexpressible." This signals the
   human: the system needs a CIC prover (dependent types) for this domain.
   The system cannot transcend its logic without external intervention —
   but it can surface the boundary precisely.

** The Lisp Machine endpoint

If the system designs and builds itself on Lisp Machine hardware:
- The same system that proves theorems also optimizes the microcode
- No OS boundary, no driver layer — system and proof environment are one
- A RISC-V soft-core with Lisp microcode is manufacturable at older fab
  nodes (28nm, 45nm) — sovereign intelligence without GPU supply chains

** Social implications

- **Concentration of reasoning.** The macro layers become opaque to anyone
  who doesn't understand the bootstrapping history. The system understands
  its own reasoning better than its users do.

- **Cost advantage widens inequality asymmetrically.** The first instance
  to reach maturity requires significant gate-rule design (from hours for
  codified domains to months for tacit ones). After that, replication is
  cheap. Organizations that invest early have a permanent cost advantage
  over those that wait for a turnkey product.

- **Sovereign artifact.** A self-building system on its own hardware does
  not depend on cloud APIs, GPU supply chains, or proprietary model
  weights. Its intelligence is generated, verified, and sustained locally.
  Enables sovereign AI for nations without GPU access.

* Open Questions

1. Can CIC (dependent type theory) be implemented as a Passepartout skill,
   verified for crash-freedom and rule fidelity by ACL2, and integrated
   into the existing fact store API? The Gödelian boundary: ACL2 can
   verify the kernel's implementation but not its soundness in any
   absolute sense — but this matches current practice (Lean 4's ~500 line
   C++ kernel is trusted, not proved).

2. Can the system generate novel proof strategies? A sufficiently rich
   abstract theory layer + Screamer could propose: "Proofs in domain X
   all use induction schema Y. Generalizing to Z would prove new
   properties across A, B, C." The LLM translates to a metafunction;
   ACL2 verifies it; the prover gains a new tactic invented by itself.

3. What is the social contract for a system that can truthfully say
   "I know this is correct" — and "I know what I don't know"?
   Most current AI systems can do neither.

* Impact on the AI and GPU Industry

If a symbolic-bootstrapping architecture becomes popular — especially now
that codified domains can be ingested at near-zero encoding cost — the
industry structure shifts fundamentally.

** Token demand compresses

The entire AI industry (OpenAI, Anthropic, Google — ~$50B API revenue) is
built on per-token pricing: metered cognition. A mature Passepartout
reduces token consumption to the unfamiliar 10% I/O boundary. Token demand
shifts from "every interaction burns tokens" to "only unfamiliar
interactions burn tokens." Steady-state per-user LLM consumption drops by
an order of magnitude.

** GPU inference demand plateaus in regulated industries

GPU inference is driven by two things: training and per-request inference.
Training demand is unaffected (frontier models still train on clusters).
Inference demand drops 80-90% in any sector where the rule book is
published — which covers most economically significant sectors (finance,
healthcare, industrial, government procurement, legal compliance).

Nvidia's growth narrative shifts from "every transaction goes through a
GPU" to "every training run needs a GPU, and the generative 20% needs
inference." A smaller inference TAM than current market pricing assumes.

** Hyperscaler competition shifts

The competitive thesis "AI is the next OS, and we own the compute layer"
weakens if the most valuable AI workloads run on a $500 RISC-V board on
your premises. The hyperscalers respond by:
- Offering Passepartout as a managed service (AGPL allows this)
- Differentiating on the frontier I/O API and world model API
- Competing on gate rule libraries for specific industries

The race shifts from "who has the most H100s" to "who has the best
domain-specific gate rules." Google's industry data advantage matters
more than Azure's raw compute.

** New hardware tier: verification appliances

A new category emerges: CPU-native verification appliances running a Lisp
microcode on RISC-V cores. Low volume (hundreds of thousands/year),
high margin ($5K-50K/unit), high switching costs. The Sun Microsystems
model, not the Intel model. Manufacturable at older fab nodes (28nm,
45nm) — no dependency on TSMC's leading edge.

** The key uncertainty and its resolution

Original question: how long does gate-rule encoding take?

Resolution: for codified domains, near-zero. The LLM translates published
regulations into formal rules in one pass — it is a mechanical transformation,
not open-ended reasoning. The bottleneck only exists for tacit, oral, unwritten
knowledge (craft expertise, organizational culture).

Consequence for the transition timeline: Phase 2 (sufficiency) happens
within months for any domain whose rule book is published. The disruption
accelerates from years to quarters.

* Broader Insights

** The historical fork: why C won economically in the 1980s

C won because the economics of 1980s hardware made Lisp's overhead
unaffordable:

- **Memory cost.** DRAM was ~$5,000/MB in 1980. Lisp's runtime (SBCL today
  is ~40MB) was unthinkable. C's runtime fit in 64KB.
- **CPU speed.** 1-10MHz. Every instruction counted. Lisp's GC, type dispatch,
  and dynamic allocation consumed cycles that C spent on actual work.
- **Software scale.** Programs were thousands of lines, not millions. A single
  developer could hold the entire program in their head and verify correctness
  by reading it. Testing was sufficient. Formal verification was unnecessary
  overhead.
- **Market dynamics.** The PC market was expanding exponentially. Speed to
  market, volume, and unit cost mattered more than correctness. A buggy $500
  PC sold more units than a correct $50,000 Lisp Machine.
- **Hardware ecosystem.** RISC (reduced instruction set) was the revolution.
  Simpler chips, higher clock speeds, cheaper fabrication. RISC CPUs are
  optimized for C's execution model because C was the dominant systems
  language when RISC was designed.

Lisp lost not because it was worse, but because the market optimized for
a different axis: raw throughput per dollar, not correctness per line.

** What changed to make Lisp viable now

Four transformations flipped the economics:

1. Memory is free. 40MB runtime is noise on a $20 Raspberry Pi with 8GB
   RAM. The cost of the runtime is now zero at any relevant scale.

2. Transistors are free. A modern ARM Cortex-A72 has billions of
   transistors. The GC, type dispatch, and dynamic dispatch that Lisp
   needs are executed in dedicated silicon within the CPU — they cost
   nothing because the transistors are there whether used or not.

3. Software complexity saturates human verification. Systems are now
   tens of millions of lines. No single person can hold them in their
   head. Testing is necessary but insufficient — zero-day vulnerabilities
   prove that bugs survive all testing. Formal verification is no longer
   overhead; it is the only known path to correctness at this scale.

4. The cost of failure is now higher than the cost of verification. A
   single breach costs millions. A compliance failure shuts down a
   factory. Regulation (GDPR, SOX, HIPAA, FedRAMP) mandates provable
   compliance. The cost of proving correctness is now cheaper than the
   cost of not proving it.

** The key insight

The 1980s trade-off was: C is cheap enough for the market. Correctness
is a luxury the market cannot afford.

The 2020s trade-off is: C is expensive for the market. Incorrectness
has become the dominant cost of software. Lisp's verification infrastructure
is now the cheaper option.

This is the inversion Passepartout exploits: the verification appliance
(AGPL symbolic engine + RISC-V Lisp μcode on FPGA) costs $5,000/year and
replaces $500,000/year in compliance audits, breach litigation, and
regulatory fines. The 1980s math said Lisp was too expensive at any price.
The 2020s math says Lisp is the only affordable option.

The remaining question is not whether the economics flipped — it's whether
anyone builds the bridge from today's AGPL software to tomorrow's
verification appliance. Passepartout is that bridge.

** Distribution and lifecycle management

Passepartout diverges in two independent dimensions across instances:
code (engine, macro layers, skills) and knowledge (gate rules, fact store,
ontology, proof forest, empirical decision history). This creates a
distribution problem no current AI system faces — most agents ship as
static binaries with no per-instance knowledge divergence.

*** Distribution tiers

Code only (AGPL repo — GitHub)
- The core engine, ACL2 macro layers as skills, generic gate rules.
- No domain knowledge, no pre-seeded fact store, no proof forest.
- User bootstraps from scratch — the system learns from their gate outcomes.
- Audience: hobbyists, researchers, early adopters willing to invest time.

Code + basic knowledge (commercial data package)
- Pre-seeded fact store for a domain (healthcare, finance, infrastructure).
- Curated gate rules, ontology categories, and initial ACL2 theorem set.
- The buyer gets sufficiency faster — weeks instead of months to reach
  the Phase 4 flip in their domain.
- Audience: enterprises that want fast deployment without bootstrapping.

Code + knowledge + verification appliance (hardware product)
- RISC-V + Lisp μcode on FPGA or custom ASIC.
- Pre-loaded with full domain-specific knowledge package.
- Hardened, certified, no-cloud, tamper-proof HSM root of trust.
- Audience: regulated industries that need provable compliance out of
  the box.

*** The upgrade problem

Once instances diverge in both code and knowledge, a naive `git pull`
breaks things:

1. A new macro layer expects a fact structure the old store doesn't have
   → re-verification fails, system degrades to pre-macro state

2. Two instances of the same version develop different ontologies
   → a shared gate rule package produces different outcomes on each

3. Knowledge learned in v1 must be verified against v2's code
   → old facts carry `:ontology-version`, ACL2 checks they're still
   consistent under the new code's rules

4. A security patch changes the gate stack
   → all existing gate-outcome facts must be re-verified against the
   new gate vectors

*** How Passepartout's architecture handles this

The architecture already has the primitives for safe upgrades:

- **Ontology versioning (Phase 5).** Every fact stores the ontology
  version at assertion. On upgrade, facts with old versions are flagged
  for re-verification. ACL2 checks consistency against new code. Facts
  that survive are promoted to the new version. Facts that fail are
  flagged for review.

- **Degradation, not crash.** If an upgrade breaks the fact store, the
  system degrades to the pre-macro symbolic state (hash-table fallback
  for VivaceGraph, text-scan fallback for Screamer). It still works —
  it just can't prove as much. The TUI displays: "Symbolic index: legacy
  mode. 2,134 facts pending re-verification under new ontology."

- **Reversible upgrades (Phase 0 undo).** Every upgrade produces a
  Merkle snapshot before applying. If re-verification fails, the system
  rolls back to the pre-upgrade state. The failed upgrade becomes a
  memory-object with `:provenance :failed-upgrade` — empirical data
  for the next upgrade attempt.

- **Delta distribution.** Upgrades are delivered as diffs against the
  current ontology version, not full replacements. A code upgrade ships
  with a migration script: "This version changes the gate rule structure
  from (action pattern) to (action pattern context meta). Here is the
  automated migration for your existing gate rules." The user's
  empirical data is preserved across the migration.

*** The real challenge: divergence of empirical knowledge

The hardest upgrade is not code — it's the fact store. Two instances
that started from the same code but served different users for a year
have diverged ontologies. Instance A's gate rules encode "emergency
restart = docker compose restart" (hospital IT). Instance B's gate rules
encode "emergency restart = emergency_shutdown.py" (factory floor).

A code upgrade that changes how emergency restarts are processed cannot
be safely applied to both instances with the same migration. The
migration must be per-instance, tested against the instance's specific
facts.

The pattern that solves this: **the upgrade is itself verified by the
upgraded system before committing.** The distributor ships: "Here is
the new `emergency-restart` gate vector. Run it against your existing
gate rules. ACL2 will report which rules are compatible and which need
human review." The instance operator reviews only the incompatible
subset.

This is slower than `git pull` but avoids the catastrophe of a silent
incompatibility. An enterprise running Passepartout in a compliance
context would not accept an unverified upgrade anyway.

*** The business model for upgrades

- Code upgrades: free (AGPL). Core engine improvements benefit everyone.
- Migration scripts: subscription. The instance operator pays for the
  verified migration path from their current ontology version to the
  new one. The migration script is generated by running the new code
  against the instance's fact store in a sandbox.
- Domain knowledge package upgrades: subscription. When HIPAA updates,
  the healthcare gate rule package updates. The new rules are verified
  against the instance's existing fact store before shipping.
- Verification appliance firmware: bundled with hardware. The Lisp
  Machine's microcode is updated as the engine evolves. The update is
  signed and verified against the hardware root of trust before
  applying.

** Large refactoring in a neurosymbolic planner — semantic equivalence

*** The workflow

ACL2 proves semantic equivalence of programs written in its own
logic — which includes Passepartout's own source code. When the
system refactors its own skills, ACL2 can prove the new function
produces the same outputs for all inputs as the old one. This is
standard ACL2 practice (verifying compiler optimizations, sort
algorithm replacements).

For other languages (Python, Java, JavaScript), the path is:
1. Model the critical subset (API surface, contracts, data
   transformations) in ACL2 as a logical specification
2. Prove the specification is preserved across the refactoring
3. The actual implementation stays in the target language —
   ACL2 proves the structural contract, not the runtime behavior

The CIC prover upgrade (Lean-in-Lisp, planned as future work)
would extend this to dependent-type-level equivalence proofs
across language boundaries — verifying that a Rust API binding
correctly wraps a C library, or that a Python refactoring
preserves the type-level contract of the original.

** The self-driving Lisp Machine on FPGA or Tenstorrent

A Tenstorrent P150 (~72 RISC-V Tensix cores on a PCIe card) or
a mid-range FPGA (AMD Alveo, Intel Agilex) offers enough
hardware to run a full Passepartout image with Lisp microcode
acceleration. The host Linux system provides boot, I/O, and
thermal management; the accelerator card provides the Lisp
execution fabric.

*** What it can do today

- **Run the full symbolic engine.** ACL2, Screamer, VivaceGraph,
  and the fact store are pure Lisp — they run on any Lisp backend.
  The RISC-V cores on a Tenstorrent or the soft-core on an FPGA
  provide enough compute for real-time gate verification and
  constraint solving.

- **Hot-reload skills and macro layers.** The Lisp image loads
  skills, tangles Org files, compiles ACL2 books, and registers
  metafunctions — all without reboot. The FPGA fabric can be
  reprogrammed with new microcode in milliseconds.

- **Manage its own knowledge base.** The fact store grows and
  evolves. Gate rules are proposed by the LLM and verified by
  ACL2. Ontology versions are tracked. The system knows what
  it knows and what changed.

- **Roll back failed upgrades.** Merkle snapshots provide
  instant undo for both software state and FPGA configuration.

*** What it needs to cross the threshold to self-driving

The system is not yet fully self-driving because three things
still require external intervention:

1. **The LLM dependency.** The 10% I/O translation (natural
   language → structured goal, structured result → natural
   language) requires an LLM. A small local model (Phi-4,
   Qwen 2.5) on the host or card can serve this. The symbolic
   engine handles everything else. Once sufficiency flips
   (Phase 4), even the LLM is rarely needed.

2. **Hardware driver development.** The FPGA microcode (tagged
   memory, hardware GC, Lisp dispatch in hardware) is currently
   written by humans. The system could eventually propose new
   microcode patterns from profiling data — "your GC accounts
   for 12% of runtime; here is a hardware GC barrier that
   reduces it to 3%" — but the synthesis and verification of
   hardware descriptions (VHDL, Verilog) requires a separate
   toolchain.

3. **The initial bootstrap.** The first FPGA load, the first
   Linux boot, the first Lisp image — these are done by a
   human or a pre-existing system. Once bootstrapped, the
   system manages itself. The threshold is crossed when the
   system can design, compile, and load its own FPGA microcode
   from within the running image.

*** The threshold

The self-driving threshold is crossed when the system can
synthesize and load its own FPGA microcode or Tensix dispatch
programs from within the running Lisp image. At that point:

- The system profiles its own gate verification latency
- It proposes a new microcoded instruction for the hot path
- It compiles Verilog from ACL2-verified specifications
- It reprograms the FPGA fabric via PCIe DMA from within SBCL
- It benchmarks the new instruction against the old one
- If throughput improves, the new microcode becomes permanent
- If not, it rolls back and tries another approach

This is not science fiction — it is the natural extension of
an architecture that already hot-reloads its own code, tracks
its own performance telemetry, and verifies its own changes
before committing them. The hardware description language is
the last abstraction boundary.

*** What stops it from being full science fiction

| Barrier | Status | Path |
|---------|--------|------|
| LLM dependency | Phase 4 flip reduces it to near-zero | Already designed |
| Hardware microcode synthesis | Most speculative | Requires hardware DSL verified by ACL2, then compiled to FPGA bitstream |
| Initial bootstrap | One-time human action | After first load, system manages itself |
| Power and thermal | Handled by host Linux | Unchanged |
| PCIe DMA from SBCL | Feasible with sb-alien + libpcie | Needs driver, but well-understood |

The Tenstorrent approach is particularly interesting because
its Tensix cores are *already* RISC-V processors. The microcode
is not FPGA logic — it's a RISC-V program. The system can write
RISC-V assembly, compile it with the RISC-V toolchain, load it
onto the Tensix cores, and benchmark the result. This is
dramatically simpler than FPGA synthesis because it's software,
not hardware.

A Tenstorrent P150 running Passepartout would be: 72 RISC-V
cores running Lisp microcode, one core dedicated to the ACL2
prover, one to Screamer, the rest to gate verification and
fact store operations. The host Linux system handles I/O and
the LLM. The system designs its own core dispatch logic,
loads it onto idle cores, and verifies the result with ACL2
before committing.

** How the LLM's coding ability affects the bootstrapping timeline

The LLM writes code at every stage of the bootstrapping:
1. The Lisp Machine's microcode (RISC-V dispatch, GC barriers, 
   tagged memory operations)
2. The CIC prover kernel (if built as a skill)
3. ACL2 macro layers for new domains
4. Gate rules for previously uncodified domains
5. The initial self-optimization proposals

At each stage, the symbolic engine (ACL2, Screamer, gate stack)
verifies the LLM's output before accepting it. The LLM proposes;
the symbolic engine disposes.

This means the LLM's coding ability is a **speed multiplier, not
a gate**. A weak LLM (3B local model) produces correct code 
after N retries where the symbolic engine catches the mistakes
and feeds them back. A strong LLM (Claude Sonnet, DeepSeek, GPT)
produces correct code after fewer retries. The cost difference
is in API calls and wall-clock time, not in the correctness of
the final output — the symbolic engine guarantees that.

| Scenario | LLM quality | Retries per unit | Wall-clock per unit | Correctness |
|----------|-------------|------------------|---------------------|-------------|
| Bootstrapping with local 3B model | Low | 5-15 | 10x slower | 100% (verified) |
| Bootstrapping with frontier API | High | 1-3 | 1x | 100% (verified) |
| Bootstrapping after sufficiency flip | None (symbolic only) | 0 | Instant | 100% (verified) |

The critical transition is between row 2 and row 3: once the
symbolic engine has accumulated enough non-lossy facts about
the Lisp Machine's hardware behavior (latency profiles, GC
patterns, instruction timings), it can propose microcode
optimizations without any LLM involvement. ACL2 proves the
optimization preserves correctness; Screamer checks it against
known hardware constraints; the gate stack verifies it won't
damage the running system. Zero tokens.

This is the sufficiency flip applied to hardware. The timeline
to reach it depends on how many facts the system can gather
about its own runtime behavior, not on how good the LLM is.

** The surprising result

An LLM that is just barely competent at coding (enough to
generate syntactically valid RISC-V or Lisp that passes the
symbolic engine's checks after a few retries) is sufficient
for the entire bootstrapping chain. It takes longer — more
retries, more wall-clock — but it reaches the same endpoint:
a system that designs its own microcode without any LLM.

The LLM's coding ability determines how many API dollars and
calendar months the bootstrap requires. It does not determine
whether the bootstrap succeeds. A patient operator with a
3B local model and a Tenstorrent card reaches the same
destination as an operator with bottomless API credits — the
second arrives faster, but both arrive.

Large refactoring projects (extract module, rename API, split monolith)
are the hardest test for any AI agent. Current approaches (Claude Code,
Copilot) handle them probabilistically — every step costs tokens, and
there is no formal guarantee the final system is consistent.

Passepartout's Phase 7 planner (10-80-10) transforms this: the symbolic
engine handles planning, ordering, and structural verification; the LLM
handles only the code transformation itself.

*** The workflow

1. **Codebase ingestion.** The scanner walks the entire codebase, builds
   an abstract syntax graph (not just flat files — imports, type
   dependencies, function call chains, test coverage). Each module
   becomes a fact in the fact store with `:provenance :codebase-scan`.

2. **Goal translation (LLM, 10%).** "Extract authentication into its own
   service" becomes a structured goal plist:
   (:goal :extract-module
    :source :auth
    :target-service auth-service
    :files (:affected (:app/auth/* :app/middleware/auth.clj :tests/*))
    :constraints (:no-breaking-api (:public-api :unchanged))
    :verification (:all-tests-pass :api-contract-preserved))

3. **Constraint satisfaction (Screamer, 80%).** Screamer expresses the
   refactoring as a constraint satisfaction problem:
   - Variables: file modifications ordered by dependency (auth middleware
     must be updated after auth module is extracted but before its callers)
   - Constraints: no circular dependencies, no step that creates broken
     intermediate state, no test file modified before its source
   - Objective: minimize total steps while respecting all constraints

   Screamer returns a viable plan or reports unsolvability with the
   conflicting constraints — for example, "auth middleware and auth
   module have a circular type dependency that must be resolved before
   extraction."

4. **Plan verification (ACL2, part of 80%).** ACL2 proves:
   - No dependency cycles in the plan (A must run before B, B before C,
     C before A → rejected)
   - No deadlocks (two modules waiting on each other)
   - Every planned write is within the refactoring scope (no stray
     modifications to unrelated files)
   - The gate stack will not reject any planned command (no blocked
     patterns in the refactoring scripts)

5. **Incremental execution.** The planner executes each step:
   a. Take Merkle snapshot of the current state
   b. LLM proposes the code transformation for this step
   c. Gate stack checks the proposal (no forbidden file writes,
      no shell commands outside the allowed refactoring scope)
   d. Transformation is applied
   e. Tests run. If they pass → commit the snapshot, update the
      fact store with the new codebase graph. If they fail → roll
      back to the previous snapshot, flag the LLM's proposal as
      `:provenance :failed-transformation`, and attempt a corrected
      proposal.

6. **Final verification.** After all steps complete, ACL2 re-verifies
   the entire codebase fact store — all dependencies, all public API
   surfaces, all constraints. The result is a Merkle chain from
   "before" to "after" with every intermediate state verified.

*** What the symbolic engine cannot do

The fundamental limit: **first-order logic cannot prove semantic
equivalence of arbitrary programs.** ACL2 can verify that the
refactoring plan has no structural flaws, that the dependency graph
is acyclic, that every step is within scope, and that tests pass.
It cannot prove "the extracted auth service behaves identically to
the inline auth module from the caller's perspective" for a
general-purpose language.

This gap is filled by:
- **Tests as empirical verification.** The planner runs the full test
  suite after each step. A passing test suite is not a proof of
  correctness, but combined with ACL2's structural verification, it
  is strong empirical evidence.
- **API contract checking.** For refactoring that preserves public APIs,
  ACL2 can verify that the type signatures, argument counts, and
  return types of the extracted module's public interface match
  exactly — a structural equivalence that does not require semantic
  reasoning.
- **Human review of semantic concerns.** The planner flags steps that
  involve semantic choices (e.g., "the extracted auth service now
  handles token refresh differently"). These steps are presented to
  the developer for review with full provenance: before state, after
  state, and the diff. The developer's approval or rejection becomes
  a Merkle fact with `:provenance :human-reviewed`.

*** What makes this better than Claude Code

| Dimension | Claude Code | Passepartout Planner |
|-----------|-------------|---------------------|
| Planning | Prompt-based, implicit | Screamer CSP, explicit, verified |
| Step ordering | Greedy, every step costs tokens | Dependency-ordered, zero-token constraint check |
| Rollback | Limited (git reset, no per-step) | Merkle snapshot per step, instant rollback |
| Scope control | Prompt-based ("only touch auth files") | ACL2-verified write scope, cannot escape |
| Cost | $0.50-$2 per refactoring session | Near-zero (CPU cycles for planning, LLM for code only) |
| Final proof | None — you trust that tests caught everything | Merkle chain from before→after, ACL2-verified |

A software ecosystem changing hardware economics has never happened before.
Passepartout's most realistic path: verification appliances for regulated
industries — RISC-V cores with Lisp microcode on FPGA, sold as hardened
devices for healthcare compliance, defense, and industrial control.

Not a general-purpose Lisp Machine replacing laptops. A specialized device
where correctness is worth paying for. If such appliances sell in the
hundreds of thousands, the economics of a custom Lisp ASIC start to make
sense. The reversal is not Lisp returning as a general platform, but Lisp
winning a vertical important enough to justify its own silicon.

The path: Passepartout software (AGPL) → creates demand for verified
infrastructure → verification appliance (FPGA, RISC-V + Lisp μcode) →
high-volume niche → custom ASIC economics viable → Lisp native hardware
exists for the first time since the Symbolics era.

** Lisp vs C for embedded systems

- Lisp can match C for low-level work through compile-to-C paths (ECL,
  PreScheme) or tiny Lisps (uLisp, FemtoLisp, BitLisp for RISC-V)
- The GC is the hard wall for hard real-time; mitigated by pre-allocation,
  no-alloc hot paths, or real-time GC
- Most practical path: "Lisp as macro language for C" — generate C from
  Lisp macros, ship the compiled binary. This is how NASA's Deep Space 1
  worked: Lisp planning on Earth generated commands for C flight software.
- The Lisp Machine on commodity FPGA (RISC-V softcore + Lisp μcode on
  Artix-7 / iCE40) is the ambitious path — Lisp down to the metal for $50.

** Microbiology works like Lisp, not C

Striking parallels:

1. Homoiconicity — DNA is code and data in the same molecule; no separate
   source and binary
2. Hot-reloadable image — alternative splicing, epigenetic marks,
   post-translational modifications change the running program without
   restart
3. Automatic memory management — proteasomes degrade misfolded proteins,
   autophagy recycles organelles; the cell never calls free()
4. Interpreted dynamic language — DNA → RNA → ribosome (interpreter) →
   protein; no static compilation step
5. Self-modifying source — CRISPR, transposons, DNA repair modify the
   genome at runtime; eval on the genome
6. Duck typing — protein folding depends on chemical environment, not
   type declarations; interfaces are shape-matching, not compiler-checked
7. Concurrent real-time GC — apoptosis breaks down cell components for
   recycling by neighboring cells; the collector is external to the object

Biology chose the Lisp model because it is more robust, adaptable, and
evolvable. Evolution paid for the overhead (GC, interpretation, dynamic
dispatch) with parallelism and redundancy. It optimized for survival in
an unpredictable environment, not peak single-thread throughput.

Biology is the proof that the Lisp model can be efficient at planetary
scale, running on hardware that self-assembles from food. The ceiling
Passepartout aims at is still far below the system that wrote itself
in DNA.