SYSTEMS RESEARCH Universal Autonomous Software Synthesis Through Reverse Engineering A Model-Agnostic Architecture for Reconstructing Software-Production Capabilities and Achieving End-to-End Natural-Language Product Creation 1 November 2026 Abstract Can a software-building agent finish a product after discovering that it lacks one of the tools needed to build it? We study this as a problem of capability admission during an active build. A controller records what an unfinished product still requires. A reconstruction worker obtains evidence from an authorized reference, synthesizes a candidate, and tests the observable behavior on which the prod- uct depends. A passing candidate becomes a versioned executable dependency. We implemented this path on Linux using pinned source modules from Reverse Engineer Anything (REA), a bounded learner, and a persistent product controller. Three byte-transform fixtures matched their compiled references on all 256 inputs each. A fourth candidate matched 254 of 255 withheld inputs and was cor- rectly refused. In a separate integrated run, a sensor-export job acquired the previously absent trans- form, survived process termination, and delivered an archive that passed tests against the compiled reference. A live coding agent given a new natural-language brief and the recipe produced a different, relocatable imaging tool that passed 16 independently frozen checks on first submission. Comparisons using a serializer and a dependency scheduler show why admission requires a distinct behavioral ora- cle for each operation. These results establish a feasible mechanism for specified, observable functions. They leave open the success rate on unfamiliar products, the automatic discovery of references, and acceptance of products that depend on external infrastructure or tacit user preferences. Keywords — program synthesis, reverse engineering, software agents, behavioral equivalence, durable execution, end-to-end testing. 1. Research problem Consider a request to produce an offline image from instrument readings. Writing a file in the correct image format solves only part of the problem. If the instrument applies a private calibration function and the software builder has no callable version of it, the builder needs an executable dependency. If it estimates the calibration from a few examples, it can produce a well-formed image with wrong pixel values. If it asks the user to discover the algorithm, the intended no-programming interaction has already failed. This case separates three obligations. The builder must identify which behavior the product needs, obtain or recover that behavior, and show that the resulting product meets its purpose at its point of use. The same distinction applies when a missing tool is a compiler backend, a browser driver or a deployment client. The complexity lies in the reference, the observation boundary and the delivery target. They can differ even for two tools with the same apparent interface. We define a capability as an executable operation with a declared input domain, state model, outcomes, 1 Autonomous software synthesis Research paper effects and operating conditions. A capsule binds an implementation of that operation to a version, a digest and the evidence supporting its use. The Autonomous Product-Completion Controller (APCC) owns the product job. REA supplies selected analysis operations to a separate worker. The LLM can be changed without retraining or rewriting the capsule if the replacement host can express the call, authorize its execution, receive its results and continue the job. 1.1 Questions and tests Question Hypothesis and decisive observation RQ1. Can the platform determine what it needs? For a declared target and catalog, dependency resolution can produce a grounded plan or identify a missing prerequisite. Removing a necessary device or authority must invalidate the affected plan. RQ2. Can an unavailable tool be recovered to a specified standard? For an observable, finite function in the selected candidate family, reference evidence plus held-out checks can admit a correct replacement. An out-of-family reference tests whether a plausible candidate is refused. RQ3. Can the original job finish after the tool is recovered? From an empty registry, successful registration must survive worker termination. The resumed job must use that digest, produce a deliverable and pass an independently executed product test. RQ4. Can another client or product reuse the outcome? A compatible second transport must preserve valid and invalid call semantics. A fresh agent given an ordinary-language brief must produce a different product using a newly materialized tool. The two experiments test different forms of portability. The universal proposition is stronger than any of these hypotheses. It quantifies over open-ended product requirements, unknown services, operating systems and human intentions. Our finite exper- iments can establish existence and identify counterexamples. They cannot estimate the probability of success for that open population. We therefore use “universal” for the proposed interface and exten- sion mechanism, and specify a separate empirical scope for each result. 2. Related work and contribution Cai et al. separate tool making from tool use and cache reusable programs [S1]. Wölflein et al. supply a research repository and a short request to an agent that installs dependencies, writes an interface and corrects failures. They report successful implementations on 12 of 15 computational tasks [S2]. Thus reusable agent-generated tools and automatic wrapping of existing code are established prior art. The present research asks what must happen when the needed implementation is absent from an active product build and the available reference is an executable artifact. The outcome of interest is a recovered dependency linked to an accepted product under the same job identity. Program sketching supplies an instructive precedent. A solver fills holes in a specified program family and uses counterexamples to refine the candidate [S3]. Active automata learning can recover a tar- get under explicit assumptions about the target class and the queries the teacher answers [S4]. The present learner uses a small hypothesis family and a reference that answers selected inputs. For the byte fixtures, exhaustive comparison is also available. Most real services instead expose incomplete histories and costly, state-changing interactions. The observation regime determines the strength of an admission verdict [S4–S6]. 1 November 2026 2 Autonomous software synthesis Research paper Build-system research separates dependency scheduling from the decision to rebuild [S7]. Package research shows that version coexistence and conflicts require joint resolution [S8]. Durable execution and message-processing research explain how activities may repeat after a failure [S9–S11]. Our pro- posal composes these results around a product obligation with an independently evaluated delivery condition. The contribution is the linkage among the obligation, reference, admitted digest, resumed build and acceptance receipt, along with executable controls that can invalidate each link. We make no priority claim for the linkage. SWE-bench studies repairs in existing repositories. SWE-Lancer uses external end-to-end tests for bounded professional tasks [S12, S13]. Both inform evaluation design. A subsequent benchmark audit documents cases where tests reject valid implementations or demand behavior absent from the brief [S14]. We use two different oracles below precisely because a universal byte-for-byte comparator would have rejected most of the valid schedules in one experiment. 3. Capability closure and product acceptance Let a product request fix a target environment, an authority envelope and a set of required observa- tions. Model implementation choices as a directed hypergraph. A recipe has a set of required inputs and one output capability. Every input must be available together. Several recipes can produce the same capability. Begin with a set of grounded roots such as a working compiler, a local filesystem, a licensed library or an account-holder grant. Repeatedly add the outputs of recipes whose inputs are all present until the set stops growing. An ungrounded cycle contributes nothing. A cross compiler already available on the build host, for example, can ground a bootstrap cycle. For a finite catalog with context-free recipes and no mutually exclusive selections, this iteration termi- nates. Each productive round adds at least one of finitely many nodes. By induction, every derived node has a derivation rooted in the initial set. Conversely, induction over the height of any finite derivation shows that its output will be added. A worklist tracking the number of missing prerequi- sites can visit each prerequisite edge once in this simplified model. Count this as structural coverage. If one branch demands incompatible versions of a runtime, simple closure can overstate feasibility. The catalog also bakes in choices such as debugging and profiling access for its production profiles. A production resolver must bind a consistent version/target assignment, model consumable resources and reserve the necessary budget before executing the selected recipes [S7, S8]. For any product class, define four increasingly strong predicates. Reachable means its goal nodes have grounded derivations. Installable means their selected implementations can coexist on the target. Ex- ecutable means the authorized environment successfully performs the required operations. Accepted means the delivered artifact passes the product-level observations in the declared target and the re- quired receipt points to that artifact. Reporting the first predicate as the fourth would turn a graph calculation into a fictional deployment. 3.1 Scope of the catalog The executable catalog contains 89 nodes, with 55 declared roots and 39 recipes for 34 derived capa- bilities. Of the roots, 11 represent named target resources or authorities. Each of 12 product profiles has explicit goals and a release observation. The table gives abbreviated examples. The full graph and per-profile observations are included alongside the blueprint. 1 November 2026 3 Autonomous software synthesis Research paper CLI editor GUI editor Edit source OR Board access external fi xture Firmware image prior build Physical validation measure on target AND Simulator check emulator available Figure 1. Two kinds of dependency. Either editor satisfies the editing step. A phys- ical validation claim requires an image and a board. A simulator produces different evidence. Profile Material dependency Distinguishing delivery observation Web application HTTP runtime, persistent data, browser and target host A real user flow succeeds at the requested origin and prior state survives a restart. Native desktop OS-specific packager, signing path and target installation A clean target installs, launches and can remove the specified package. Mobile SDK, signing identity and device or emulator The intended signed build installs and completes a named flow on its target. Game Engine/export templates, assets and target runtime A packaged build responds to input and reaches the required gameplay state. 2D graphics Vector or raster generation and an independent renderer A consumer renders the expected geometry, pixels and dimensions. 3D graphics Scene representation, loader and rendering context Scene topology, transforms and visual output survive interchange within tolerance. CAD Geometry kernel and exchange/import path A second operation on the imported solid reproduces topology and dimensional invariants. Multimedia Codec, container and decoder Decoded samples, timestamps or perceptual tolerances meet the contract. Data processing Parser, schema, transformation and serializer Expected rows, nulls and numerical semantics survive an independent reader. Embedded Cross toolchain, board interface and physical fixture The programmed target produces the measured response under the declared conditions. Distributed service Multiple failure domains, observability and recovery Injected loss or restart preserves the required durable outcome. AI-enabled software Model runtime, permitted model artifact or provider, evaluation data The shipped inference path satisfies held-out quality, cost and latency requirements. The source concordance for these profiles is keyed [C1–C19] in the companion bibliography. The model includes language and runtime toolchains, compile/link, builds and dependency resolution, 1 November 2026 4 Autonomous software synthesis Research paper editing and version control, GUI and media authoring, browsers, testing, debugging and profiling, OS/network interfaces, databases, identity and APIs, simulation, containers, virtual machines, de- ployment, telemetry and recovery. These functions are activated by the product contract. An IDE can expose useful operations while a headless editor satisfies an equivalent build obligation. A hosted pay- ment network, publisher signing authority and physical instrument are different kinds of prerequisite. Reconstructing their client libraries supplies no claim on those resources. With every root declared available, all 12 structural profiles close in six rounds. Removing the 11 target-resource and authority roots leaves five profiles reachable under their specified local-delivery goals (2D, 3D, CAD, media and data). Removing only the four authority roots leaves seven. Removing the physical board leaves a simulation path while eliminating the physical firmware acceptance path. These are counterfactuals in a declared model. Different release contracts change the counts. A local web archive and a web service deployed to a named host are distinct goals, even when they share all the source code. 4. Reconstruction as a scoped inference problem We distinguish five routes. Direct invocation uses the existing executable. Source reuse uses identified licensed source, preserving notices and the distribution conditions. Protocol compatibility implements the expected messages and state transitions at a specified boundary. Functional reconstruction inde- pendently implements the required observable behavior from a specification or evidence. Complete replication would require every in-scope behavior of the reference, including operational dependen- cies and state. Success under one route supports only its own claim. A replay may be an adequate test double for a protocol exchange and an inadequate implementation for the next product. The source study pinned REA 6.8.0 at revision d3ab0faf52a02ff5a618f3ed7beb608b78752abd [R1]. Its ca- pability composition declares provider identities separately from observations of host availability. Its reconstruction ledger already represents claims, evidence, scoped dependencies and unresolved obli- gations. Its coverage assessment can leave an authenticated but insufficient claim unknown. Recorded environments can replay captured interactions and model some observed changes [R2–R4]. Those prior mechanisms matter because a controller can consume a structured result and keep its exact ev- idential scope. This study exercised selected modules and did not run those higher-level ledger or recording workflows. The proposed worker first reduces a missing tool to the smallest externally testable operation sufficient for the current consumer. Given several authorized references, it ranks them by behavioral fit, compat- ible execution environment, available oracle, permission to inspect or redistribute, and expected cost. Discovery can inspect a trusted package index, a standards catalog, an existing project inventory or an owner-supplied binary. Record rejected candidates. If the full build depends on reproducing a ser- vice’s unexposed state or on an unavailable right, return the unresolved requirement. The experiments supplied the references. Automatic discovery is part of the design. For a binary, bind the executable digest, architecture and ABI before using static inspection. Treat recovered instruction sequences, constants and call relationships as hypotheses about behavior. Test them with controlled inputs at the callable boundary. For a stateful application, record the pre-state, action, response, post-state, principal and observation point. Replay sequences after a reset. Vary ac- tion order, duplicate requests, permissions, identities and failures. For a network protocol, separate bytes on the wire from authoritative server state. For source reuse, build from pinned source with its dependency graph, compare its public behavior and document what has been carried into the deliv- erable. The full engineering blueprint specifies the applicable route, evidence, test and stop condition 1 November 2026 5 Autonomous software synthesis Research paper for each modality. A useful formal definition of equivalence names an input domain, allowed starting states, observation projection and comparison relation. A serializer may require exact bytes. A topological sorter can return any linear extension of the same precedence relation. A floating-point geometry kernel needs tolerance bounds, topology checks and degenerate cases. A stateful service can require equivalent histories up to permissible orderings and matching effects after retries. For stochastic behavior, test a declared property or distribution with an explicit sampling plan. A seed shared between implemen- tations may aid diagnosis while hiding distributional differences. A finite transcript generally leaves ambiguity. For a deterministic binary-valued function on 16 dis- tinct inputs, four distinct labeled observations leave exactly 4,096 unrestricted truth tables consistent with the transcript. Every assignment to the other 12 positions makes a different table. In the byte experiment, an explicit family narrowed the problem and full enumeration settled the chosen finite domain. For large or evolving products, the alternatives are verification over a justified model, tar- geted falsification and a narrower claim. Even exhaustive returned-value equality for a pure function says nothing about side channels or a hostile implementation’s effects unless those observations are in scope. LLM host intent + typed calls Controller job + obligations Receipt store state + causal history REA + reference observations / traces Candidate builder recipe + evidence Registry admitted versions missing requirement admit Acceptance runner user-visible outcomes Isolated executor pinned recipe + grants Artifact / target local or authorized host resolved dependency Figure 2. The registry admits a particular implementation. The acceptance runner tests the product that consumed it. The controller preserves the relationship between those decisions. 5. Controller, interfaces and authority Each product obligation has an identity, a revision, prerequisites, an acceptance predicate, an autho- rized target and a ledger of attempts. Discovery returns a pinned compatible capsule or creates a child reconstruction job. Eligible product branches continue independently. A child can return an admitted digest, a counterexample, a missing right or resource, or an exhausted search budget. The parent can then resume, try another eligible recipe or expose a specific unresolved dependency. No new user prompt is required to repair an ordinary technical failure. The production design adds immutable artifact storage, a registry with revocation, a policy broker 1 November 2026 6 Autonomous software synthesis Research paper and separate workers for construction and acceptance. The builder receives the published contract and diagnostic tests. The acceptance worker receives the candidate digest, reference access and sealed inputs. A failed promotion test can become a development example only after it has been retired from the sealed evaluation and replaced. Signed statements link actor, policy, source, environment and result [S15, S16]. An authentic statement still needs to cover the intended operation and use a suitable oracle. A publication transaction links the passed candidate, selected dependency and waiting job. Pin be- fore execution, invalidate descendants when behavior, authority, environment or acceptance criteria change, and retain successful unrelated branches. Production workers need expiring leases, fencing tokens, immutable outputs and conflict handling. The local controller demonstrates a single-worker journal and recovery at a selected commit boundary. It looks up the capability name again on restart. The registry stayed unchanged in the recorded run. The design guarantee of an original-digest pin across concurrent upgrades requires an explicit per-job binding. Our design does not infer distributed concurrency correctness from that experiment. Use a host-neutral internal call with capability identity, digest, typed arguments, preconditions, grant, effect category, job and logical action ID. A stdio MCP adapter can project discovery and invocation into MCP messages. An HTTP adapter describes its operations using OpenAPI and maps tool errors to HTTP responses. A CLI adapter can transport the same contract as bounded input and structured output. Schema validation covers shape and ranges. A separate semantic layer checks invariants such as time zones, ordering, resource identity, idempotency and units. When a host cannot preserve one, declare the operation unavailable there or provide a verified conversion. An intermediary can confer interoperability on a host that has a tool bridge. Model weights alone furnish no executor [B1–B4]. Separate three rights: observing a reference, producing a candidate from permitted information, and distributing the resulting material. Keep owner-supplied fixtures and compatible source notices iden- tifiable. Canadian and U.S. interoperability provisions specify conditions. They supply no general authorization to copy or bypass protections [S18, S19]. At invocation, bind the grant to a tenant, an audience, a destination, an effect and an expiry. Treat an instruction inside an analyzed program, pack- age manifest or tool result as data. Untrusted candidate code should run with limited CPU, memory, filesystem and network access, under a principal unable to alter its evaluator, registry or production secrets [S16, S17, B7]. The experimental runtimes execute trusted local fixtures and do not validate containment of malicious code. The receipt for an externally visible action needs special treatment. A provider can finish a write immediately before the worker dies. The local history then looks identical to a write that never reached the provider. Record the immutable intent before dispatch. Reuse its operation key only when the destination enforces key/payload identity, or reconcile against an authoritative remote record. When neither facility exists, hold the operation as uncertain. A local promise to retry cannot ensure both one external effect and guaranteed completion in those two indistinguishable histories [S9–S11]. 6. Experimental design The experiments test different links of the architecture. We fixed domains and expected outcomes per trial and retained raw observation files. Product authorship occurred within this AI-assisted study. “Independent” below means a specified separation such as an external reference, a different consumer or a grader frozen before submission. The experiments were conducted by one AI-assisted study and do not constitute a replication by another laboratory. 1 November 2026 7 Autonomous software synthesis Research paper Trial Experimental unit Comparator or falsification A. Binary reconstruction Four authored Linux x86-64 ELF fixtures Compiled C ABI oracle, withheld values, single-value faults, raw backend control B. Functional interchangeability 137 binary payloads and 161 directed graph cases GNU tools, pinned standard-library source, separately written code, transcript replay and mutants C. Interrupted product build One REA-backed job and two timed product recipes Empty, preverified and reconstruct-on-demand registries, real process exit, extracted product checks D. Live natural-language transfer One fresh-context coding-agent submission Protocol and input cases frozen before assignment, independent byte oracle, moved installation E. Interoperability and side effects Two wire adapters and 15 recovery configurations Typed successes/errors, protocol tests, two independently committed databases and hard process exits F. Domain reach Dependency graph and locally available graphics/media/CAD programs Root removals, exact root-support enumeration, pixel/signal/geometric oracles The core host was Linux x86-64 with CPython 3.12.14, Node 24.22.1, GCC 13.3.0, GNU coreutils 9.4 and SQLite 3.53.1 in the controller. The executable study includes source hashes and per-trial versions. All product release targets were local files or loopback services. Timing uses monotonic clocks and is reported separately for in-process calls, fresh processes and complete jobs. The number of invocations is not a number of independent products. No token billing, cloud cost or multi-host service reliability was measured. 6.1 Binary study and synthesis procedure An authorized assembly fixture exports a pure one-byte transformation made from multiplication, ad- dition, exclusive-or and rotation. Three variants change constants or order. A fourth includes an extra branch. The experiment compiles the fixtures, runs the pinned REA ELF-header parser and objdump- command construction code, and executes GNU objdump on the target symbol. Four upstream Type- Script files and their original license are included with Git blob identities. The study harness imports the modules directly. The full REA server, commercial disassemblers and dynamic instrumentation products were not needed for these trials [R1, R5]. The extractor deliberately retains four operation/immediate pairs without using instruction order. The learner enumerates their 24 permutations and computes their candidate byte tables. It chooses the byte whose possible answers minimize the largest remaining candidate group, queries the compiled refer- ence and retains agreeing candidates. Once the surviving candidates agree on the entire finite domain, the emitter writes Python code. A separate admission pass compares it with the original shared object at every unqueried input and checks declared error handling. This construction isolates the logic of identification and promotion. It intentionally discards information a general binary analyst could use. A straightforward order-preserving instruction translator could be simpler for these fixtures. Binary condition Candidate fitting Held-out comparison Disposition Default constants/order 1 reference query 255/255 equal Admitted 1 November 2026 8 Autonomous software synthesis Research paper Binary condition Candidate fitting Held-out comparison Disposition Changed constants/order 1 reference query 255/255 equal Admitted Alternate constants/order 1 reference query 255/255 equal Admitted Additional hidden branch 1 reference query 254/255 equal Quarantined, no manifest For the three admitted fixtures the fitting byte plus all 255 withheld bytes exhaust the declared input domain. Each admitted capsule also rejected six malformed Python inputs. Injecting a one-value mutation after fitting produced an observed mismatch for each positive fixture. The fourth fixture provides the stronger control: the actual reference departed from the family, the candidate still fit the chosen probe, and the admission test caught the error at input 255. Allowing a 99% threshold would have shipped the wrong calibration. An additional comparison on the default fixture evaluated direct use, a separately built copy of the authorized source, a specification-based implementation and the REA-assisted capsule. Each matched all 256 values. Raw objdump supplied the same normalized disassembly as the integrated command path. The experiment establishes that actual REA modules can participate in the chain. It demonstrates no incremental accuracy advantage over invoking this backend directly. Reusing the original assembly produced a byte-identical ELF build. The specification-based implementation shared study authorship and was not a clean-room independence trial. 6.2 Two incompatible definitions of correctness The first comparison requires canonical Base64 bytes for arbitrary finite input on the tested cases. The second requires a valid precedence-respecting topological order for an acyclic graph and rejection of a genuine cycle. Inputs, graph conventions and error normalization are described in the comparative protocol [T1–T3]. We invoked GNU originals, adapted unmodified licensed CPython modules and supplied separately written implementations. A lookup table replaying only construction transcripts acts as a negative control. An additional bounded Base64 learner identifies its alphabet and pad symbol from 65 executable probes under an explicit sextet-based family. Route Base64, 137 withheld Dependency order, 161 withheld GNU original 137 161 Compatible source reuse 137 161 Independently implemented algorithm 137 161 Alphabet/padding recovered from observations 137 Outside this trial Literal transcript replay 0 0 Deliberately defective variant 44 17 Of the graph cases, 90 were acyclic and 71 cyclic. On the 90 acyclic cases, the independent implemen- tation returned a different valid byte sequence from GNU in 69 cases. The reused module did so in 65. With edges from u to both v and w , either sibling can follow u first. Exact-output comparison would 1 November 2026 9 Autonomous software synthesis Research paper falsely reject correct schedules. Conversely, ignoring order entirely could accept an edge violation. The correct check verifies the vertex set, uniqueness and each precedence edge, and uses a separate cycle condition. Replay scored zero by design on deliberately disjoint inputs. This finding applies to exact lookup. It does not characterize stateful simulation or every replay-assisted implementation. Cold invocation can dominate a small tool. On 30 timed fresh processes per arm after five warmups, median Base64 execution took 1.38 ms using GNU, 31.96 ms with the Python source-reuse route and 26.27 ms with the independent Python implementation. Corresponding schedule times were 1.38, 29.42 and 26.95 ms. These are complete process-boundary costs on one host. A batching or persistent- worker decision needs timings at its own boundary. The raw samples and negative controls are in- cluded. 6.3 Same-job reconstruction, interruption and delivery The integrated controller begins the sensor-export job with an empty registry. A supplied product recognizer identifies a byte-calibration obligation. The worker calls the reconstruction pipeline into a fresh directory, validates the candidate and commits its digest with the job transition. The process then exits deliberately with code 75. A fresh process reopens the same job, resolves the admitted de- pendency and packages the product. The reference shared object stays outside the deliverable. An evaluator tested the extracted program against the original C ABI using an empty stream, all 256 byte values and 9,033 seeded bytes. All returned the expected bytes. The journal has one event each for planning, registration, acceptance and delivery for this job. Removing the synthesizer or the admis- sion record would break the causal chain. Two further products exercised a black-box URI encoder under a bytewise composition assumption. A table initially escaped every byte. Querying the reference on all 256 singletons corrected 66 mismatches in one round. The emitted function then passed the empty string and 200 fresh multibyte strings. A candidate with a special rule for the two-byte string ab can agree on all singletons. That counterexample explains why success on the alphabet alone cannot establish bytewise composition for an unrestricted encoder. The web product accepted and retrieved bookmarks through a real HTTP form endpoint and preserved data after a server process restart. The native C utility examined files and produced encoded paths and lengths. Their archived releases passed eight web and four native checks respectively, including error paths. The sensor product passed three. The web checks exercised HTTP and HTML content. They do not establish browser rendering or accessibility. The lab also rejected a publish operation outside its grant, detected a changed capsule digest and verified redelivery of an existing local artifact. Product Missing, factory disabled Previously verified Reconstructed at runtime Web bookmark shelf 0/8 delivered 8/8, 216.839 ms 8/8, 273.649 ms Native file manifest 0/8 delivered 8/8, 149.827 ms 8/8, 239.942 ms Each row summarizes eight fresh state directories per condition. The six conditions were shuffled within rounds. Successful wall times include the process launch, production steps, evaluator and local archive delivery. Preloading was outside the timed window. With the factory enabled the median additional time over a ready registry was 56.810 ms for web and 90.115 ms for native. In a separate in-process comparison on 1,000 inputs of 48 bytes, the standard-library encoder took a median 1.956 microseconds per call and the reconstructed capsule 8.269 across 12 alternated batches. The driver’s maximum child high-water memory reading was 67,376 KiB. The setup tests cost and reuse for these recipes only. 1 November 2026 10 Autonomous software synthesis Research paper 6.4 Fresh natural-language product trial To distinguish product synthesis from selecting a prescribed template, a fresh-context coding agent received one ordinary-language brief and a machine-readable catalog entry for the calibration recipe. The brief asked for a relocatable offline contact-sheet maker, with raw bytes on standard input, a user- selected width, corrected pixels in reading order and literal zero-padding in the final row. Empty input, bad widths and unknown options had to fail without partial images. The agent was instructed to read only the public brief and catalog, its own output and newly generated tool data. Files were in a shared workspace. This was a procedural separation, not an enforced filesystem boundary. The grading program and brief were hashed before assignment. The agent ran the catalog’s recipe in a fresh temporary directory, copied the verified capsule and manifest, and wrote a new PGM-producing CLI. We froze the first submission, moved it to a different directory and compared every output pixel with the compiled reference. Nine valid cases exercised width boundaries, complete and incomplete rows, and up to 10,003 readings. Seven invalid cases exercised input and option errors. All 16 passed on the first submission. The grader checked dimensions, pixel bytes, padding, failure status and absence of output on failure. It recognizes the P5 headers emitted here and is not a general PGM conformance suite. The recorded hashes verify artifact identity. Their chronology was observed within this session without an external timestamping service. No external model was called during deterministic replay. This is one observed live-agent product creation with a supplied tool catalog and a specific but non- technical brief. It extends the three controller templates. It does not establish that the same controller dynamically selects arbitrary analysis backends, that the agent can discover the right reference on the open internet or that a second model family would succeed. The controller and live-agent trials test complementary paths. 6.5 Interface and recovery trials The interface experiment ran a real stdio MCP process, pinned to protocol revision 2025-11-25, and a separate loopback HTTP process described by OpenAPI 3.1.1. Both mapped to one integer-ordering contract. Each passed six valid calls and rejected nine invalid calls, for 30 behavioral calls in total. Ad- ditional checks exercised protocol/description properties, envelope validation and content bindings. A separate URI test using a custom envelope exercised a Python stdio client and a Node HTTP client on 33 cases each. These are transport tests. They do not measure how often two independent LLM hosts discover and choose the correct tool. A controlled side-effect experiment used two independent SQLite databases, one for the controller and one standing in for a provider. In each of 15 configurations the worker started, then ran again after a normal return or a hard process exit at one of four program points. A blind-retry variant, a provider-keyed variant and a hold-on-uncertainty variant exposed how the same local history can require different recovery policies. Crash position Blind retry Destination honors operation key Hold on unresolved dispatch After intent, before external effect One effect, accepted One effect, accepted Zero effects, uncertain After effect, before local receipt Two effects, accepted One effect, accepted One effect, uncertain Before intent, after receipt, or uninterrupted control One effect, accepted in each case One effect, accepted in each case One effect, accepted in each case 1 November 2026 11 Autonomous software synthesis Research paper The keyed provider returned its previous receipt for an identical payload and rejected reusing its key for different content. Without destination support, the hold policy preserved safety at the expense of automatic completion. These observations concern process loss and fixed ordering on one machine. They do not measure network partitions, provider key expiry, machine power loss or Byzantine ac- tors. A production adapter must document how long its destination honors a key and what evidence permits reconciliation after expiry [S9–S11]. 6.6 Additional domain checks The structural model has 168 inclusion-minimal profile/support pairs. Each selected root set derives its profile. Removing any one member breaks that derivation under that set. There are 24 alternatives each for the web and distributed profiles and 12 for each of the other ten. The number counts inclusion- minimal witnesses in this catalog. It does not give the minimum number of products one must buy, the unique architecture of a foundry, or the financial cost of a build. Three directly authored local programs tested additional domain-specific oracles. Six SVG jobs pro- duced 8,564 pixels matching an independently specified raster expectation after Inkscape rendering. Six stereo WAV-to-FLAC jobs preserved all 37,448 decoded PCM sample values, sample rates, channel counts and frame counts when read through libsndfile after FFmpeg encoding. Four STEP plates with asymmetric through-bores re-imported as single solids with seven faces each. The CAD trial passed 20 occupancy checks and reached a maximum relative volume error of about 2.84 × 10 −15 against an- alytic geometry. The exporter and importer shared the OCCT kernel. The independent measurement here was the analytic reference. These tests exercise real graphics, media and CAD tools. Their produc- tion programs were authored for the study. The APCC did not autonomously reconstruct Inkscape, FFmpeg or a CAD kernel. 7. Interpretation, engineering implications and limitations The experiment with the hidden branch identifies the central reliability issue. The selection procedure obtained one surviving candidate after one query in every binary trial, including the failed one. Con- vergence of a search procedure answers a question about the learner’s own hypothesis family. It is the comparison against an independently exercised reference boundary that exposed the missing rule. An architecture that equates synthesis convergence with admission will eventually ship an