Autocontext synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From a goal to evaluated attempts
Autocontext is an improvement plane: it coordinates attempts against evaluation, retains material from them and supplies selected material to later work. This account covers the shipped Python, TypeScript and Pi-extension surfaces at commit f72c154152397f54d55407508b153d28861888a5, with an applicability cutoff of 2026-10-15. Its sole source, SRC-1, supplies inspected implementation and attributed design, not inspected execution traces or causal experiments. The whole-system boundary includes shipped callers, writers and consumers; it does not imply exhaustive inspection. External model providers, enclosing hosts and hardware implementations are not frozen. See the evidence boundary and Source register.
The strongest supported contribution is a connected architecture for evaluated attempts, durable feedback-derived guidance and later consumption—not merely a directory of reports. The broader product claim that useful lessons improve agents remains untested in the inspected evidence: RT-CLM-agent-improvement.
The representative workflow is a bounded Python saved-scenario run. The operator supplies a scenario and iteration bound; runner code schedules generations and loads retained context. A competitor produces text, a translator turns it into the strategy mapping, and support roles produce analysis, guidance and tooling proposals. The coach receives analyst text through its prompt. Crucially, standard support roles run before the current tournament; they do not retrospectively analyze its completed results. The tournament executes the strategy and supplies scenario-specific scores and validation. The basic gate compares score improvement with a configured minimum and advances, retries or rolls back. A normal retry invokes competitor and translator again, not all support roles. These ordering and recovery findings are attached to RT-RTE-scenario-generation in the workflow annotation.
Runner code owns continuation and failure recovery. Rollback preserves incumbent score/rating state; it does not reverse every executed effect. Exceptions record failed progress and re-raise, while reentry can resume pending context evaluation. Normal delivery is a returned summary, terminal events and durable SQLite/file artifacts. A completed generation may contain a rollback result, so completion does not mean every proposal was accepted. Terminal receipts count durable completed generations; the summary counts those completed by this invocation. This is a source-supported delivery contract, not an observed successful run: RT-RTE-scenario-generation, workflow account.
What changes before the next attempt
Strategy selection and context admission target different objects. When a Python context candidate and coordinator exist, matched candidate/incumbent evaluation can admit an immutable bundle or preserve the incumbent and defer progress. Promotion must activate the evaluated digest. This is a scoped protocol, not proof that each instruction is true or that every future prompt change was evaluated. Bundle-level comparison also does not isolate a component effect; the inspected attribution branch requires an eligible single-component difference and heldout trials. See RT-RTE-context-promotion and the attribution qualification in EPI-RTE-judge-evaluation.
The bundle's parts have different consumers. Playbook, hints, prompt fragments, context-policy text, completion checklists and tool instructions provide interpreted guidance. Routing JSON changes supported model/settings fields. Reference helper source is rendered explicitly as noncallable context. Executable validators instead reach a conditional veto: when two-tier gating and harness validators are enabled, the loader selects the active digest's source, and validity checking can trigger revision or prevent tournament progress. A checklist string is not that executable predicate, and source materialization alone is not activation. See MEM-OBJ-context-bundle, MEM-OBJ-routing-config, MEM-OBJ-tool-spec, MEM-OBJ-completion-check and MEM-OBJ-harness-validator.
A separate durable feedback path writes operational lessons to SKILL files. Under its conditions, advance retains coach lessons; nonadvance can produce an avoidance warning from the current strategy, score and replay excerpts. Curator consolidation can replace accumulated bullets. Later setup reads selected lessons independently of the active bundle. Thus current-result-to-warning-to-later-prompt wiring is supported without pretending the coach saw the future tournament. It also demonstrates why matched bundle admission is not a universal gate over retained advice: MEM-RTE-lesson-maintenance.
The npm generation route uses another publication policy: structured coach output passes a guard and optional curator, then is written live or staged under configured approval. Its inspected support-role call follows lifecycle persistence without a local advance predicate. Direct writers remain alternative paths. Starting either runtime does not establish human approval of each edit: RT-RTE-npm-generation.
Portable packages install guidance, source and attributed history under conflict rules, not fresh performance evaluation. Imported scores can populate synthetic accepted/completed history without becoming local measurements. Existing active-bundle playbook/hints can supersede mutable imported files, so a successful import need not immediately change serving context. Stale marking and deletion likewise affect particular readers, not all bundles, snapshots or already-delivered prompts. See MEM-RTE-package-import, EPI-OBJ-package-history and MEM-RTE-lesson-lifecycle.
What the later consumer actually receives
Generation setup automatically supplies scenario/run/role-selected retained material; MCP and Pi tools instead fulfill requests for artifacts or inspection. A documented file-consumer role is afforded, not observed host discovery. Python compaction selects sections, bullets, keywords and trajectory rows deterministically, then truncates; a later budget can remove more while protecting hints and dead ends. Original files can survive outside the window, but automatic recovery of every omitted passage is not established. The compaction ledger's shorter audit summary is neither the full delivered prompt nor a demonstrated canonical continuation checkpoint. Separate role calls isolate selected work only to a degree: persistent host sessions prevent assuming every call frame is discarded. See RT-RTE-context-assembly, MEM-RTE-component-compaction, MEM-OBJ-compaction-entry and MEM-RTE-requested-readback.
Evaluation grants local operational authority, not a universal truth license. Scenario scores concern scenario behavior. Task revision uses rubric judgment and optional references; an external verifier can veto, whereas a failed checkpoint only logs a warning. It may return a best output below threshold. Session consolidators return extracted candidates without an established durable promotion chain; TypeScript prefix slicing can omit a scope qualification. None of these labels establishes accepted knowledge. No identified produced ampliative claim was shown accepted and integrated with its evidence, but unassessed routes prevent turning that finding into a whole-system absence. See RT-RTE-task-revision, EPI-RTE-judge-evaluation, MEM-RTE-consolidation-preview and EPI-RTE-report-consolidation.
Improvement-directed self-change and its unresolved effects
Beyond advice, training revises executable training source, selects checkpoints and can auto-activate the terminal accepted model. Deterministic promotion uses summary-derived trial input; adaptive promotion requires a trial executor. Ambient activation applies separate charter, real-generation, drift and anchor conditions. Kernel evolution selects campaign champions against benchmark policy, verifies source/receipt identity and supplies primary feedback rather than detailed confirmation evidence to later proposals. These are distinct improvement-directed change paths, not interchangeable score loops: RT-RTE-training-experiments, RT-RTE-ambient-promotion, RT-RTE-kernel-evolution.
For later generations, runs or campaign experiments, retained lesson and source-update routes wire experience into changes to behavior-shaping organization. Their objectives—scenario performance or configured training/benchmark criteria—are separable from the proposed updates. This supports dispositional improvement-directed self-change. The unresolved self-improvement link is subsequent operation causally depending on the installed change; exercised self-improvement and beneficial outcomes remain unestablished. Bounded task-answer revision alone changes an external product, not necessarily the harness. These distinctions follow the runtime routes above and RT-RTE-npm-generation.
The independent property assessments remain separate:
- Theory-builder condition 1, localized theory content: uninspected for actual generated artifacts.
- Condition 2, decisions depending on that content: uninspected; delivery is insufficient.
- Condition 3, formulated content-directed criticism: uninspected; scores selecting variants are insufficient.
- Condition 4, retained criticism shaping another round: uninspected; iterative storage alone is insufficient.
- Learning as improved future capacity: unestablished. Automatic trace-fed durable updates are wired, but they do not demonstrate improved capacity or that criticism of a consumed theory produced it.
- Reflection: diagnosis and harness proposals afford reflection-like work; content-specific self-representations and both causal directions are unobserved.
- Autonomy: bounded computational scheduling and internal admission are implemented on named routes. Autonomous theory building is unestablished; this is not a claim that every maintenance or approval branch is computationally controlled.
These are the bounded findings in RT-RTE-npm-generation, RT-RTE-training-experiments and the epistemic conclusion, not a single system grade.
For repeated scenario attempts, the architecture can carry evaluated experience forward through durable guidance and configuration. For a causal assessment of a particular lesson, matched or independently varied lesson interventions and linked later behavior would change the conclusion; aggregate bundle success alone would not. For training and kernel campaigns, comparable execution results, immutable artifact identities and downstream use would strengthen the separate performance and self-improvement assessments. Concrete generated theories, their criticisms and consequent revisions would be needed to decide the four theory-builder conditions. The accepted report verification and profile verification report no additional publication limits; they do not supply missing execution evidence.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Static wiring without inspected run or intervention evidence | SRC-1; RT-RTE-scenario-generation, MEM-RTE-lesson-maintenance, RT-RTE-training-experiments, RT-RTE-kernel-evolution | Selected shipped implementations at the frozen commit; archived results not inspected | Successful delivery, exercised retained-change dependence, improved capacity, causal benefit or transfer | Linked attempt/update/installation/later-use traces and comparable interventions at the claimed treatment level |
| Generated explanatory content and its criticism were not inspected | SRC-1; EPI-RTE-role-proposals, RT-RTE-npm-generation | Prompt requests and iterative plumbing, not actual produced theories | Any of theory-builder conditions 1–4, content-directed criticism improving future capacity, reflective or autonomous theory building | Concrete theory artifacts, criticism with blame placement, retained revisions and evidence of their consequential use |
| Selective whole-system inspection leaves included alternatives unresolved | SRC-1; RT-RTE-context-assembly, RT-RTE-runtime-dispatch, RT-RTE-ambient-promotion | Notebook/report/hint writers, evidence processing, npm maintenance/memory packs, datasets/production traces, session/mission reload and host state incompletely traced | Complete memory/authority inventories, full provenance, parity, universal admission or system-wide negative findings | Part-specific producer, admission, storage and later-consumer records for the unresolved alternatives |
| Mutable files, immutable serving state and external copies are not universally synchronized | SRC-1; MEM-RTE-package-import, MEM-RTE-lesson-lifecycle, MEM-OBJ-compaction-entry, MEM-RTE-consolidation-preview | Import/restore and reader precedence, local withdrawal, ledger inspection and extraction functions | Immediate import activation, all-store revocation, complete restoration from summaries or durable session promotion | Administrative refresh/rebase and recovery traces, copy-revocation contracts and persistent extraction-result consumers |
| External execution contracts and exact runtime models are not frozen | SRC-1; RT-RTE-runtime-dispatch, RT-CMP-role-models, RT-CMP-training-checkpoints, MEM-OBJ-harness-validator | Adapter and conditional validator code, not host/provider internals, deployed grants, immutable weights or hardware operation | Universal isolation, termination/security guarantees, exact parameter identity, correct generated validators or demonstrated serving behavior | Pinned model/checkpoint contents, deployed permission evidence, independent execution/security assessment and linked serving results |
| Local evaluation does not authenticate imported claims or explanations | SRC-1; EPI-OBJ-package-history, EPI-OBJ-package-harness, EPI-RTE-judge-evaluation, RT-RTE-context-promotion | Import conflict rules, rubric/reference judgments and scoped bundle comparison | Fresh local performance from imported status, source truth, complete import-to-validator enforcement, or component/mechanism warrant from aggregate outcomes | Authenticated provenance and comparable measurements, traced imported-source consumers, checked references and independently varied components |