o-o synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From a readable file to a research request
o-o is a workflow for maintaining living research articles, not an implementation of an agent's internal model/tool loop. Each local .o-o.html file is both browser-readable HTML and a Bash entry point. Its shipped index creates documents, refreshes a library and rebuilds navigation. This account covers the whole shipped repository at commit b0d5063e37b4eafdbc23c1899a31f1836168b989, with an analysis cutoff of 2026-10-15. SRC-1 supplies inspected implementation, declared procedures and attributed retained content. Neither target execution nor an observed update trace or causal comparison was supplied. Claude Code, providers and research websites are outside this frozen boundary. See the boundary and evidence and Source register.
The README's “self-updating” claim means that running a document can launch an update; it does not establish an internal scheduler. The local shell derives the target path, parses freshness metadata and normally skips a fresh article. --force bypasses that gate. For eligible work, it extracts a monetary budget and starts claude -p with the path, instructions and requested Bash, Read, Edit, WebSearch and WebFetch tools. The complete article is not automatically placed in the initial prompt. The external agent must read the file to obtain the document-specific contract and prior research. Exact model version, effective permissions, provider billing enforcement and internal context handling remain uninspected. See RT-CLM-living-document, RT-RTE-document-update and RT-CMP-claude. Source basis: SRC-1 README.md, example/index.o-o.html.
The contract supplies subject, audience, research intents, required sections and source preferences. It delegates query selection, evidence judgments, writing and edits to the same external agent. Ordinary documents request incremental enrichment when little news exists. Sentinel documents instead request concise non-event reports, expanding coverage only when the model judges that the specified event occurred. These are interpreted instructions, not separate deterministic event-detection branches. An AGI-declaration check concerns a credible declaration, not proof that AGI exists. See MEM-OBJ-contract, EPI-RTE-web-research, EPI-RTE-sentinel-check and EPI-RTE-sentinel-selection. Source basis: SRC-1 example/anthropic-leadership.o-o.html, example/agi-declared.o-o.html, example/trump-incarceration.o-o.html.
Research becomes retained context without a separate acceptance gate
The principal cross-invocation contribution is a persistent document-local handoff: article, manifest, prior sources, factual summaries and change summaries remain available, and the next update is explicitly asked to build on them. This supports a reuse pathway, not demonstrated learning. Which facts actually reach the later agent, change its decisions or improve its capacity remains unresolved. The file path selects a document; it does not establish semantic fact retrieval. Whole-file reads may include generator literals and encoded media, and o-o supplies no stated token quota or selective compaction policy. See MEM-BAP-cache-reuse, MEM-OBJ-facts and EPI-RTE-cache-reuse. Source basis: SRC-1 example/index.o-o.html, example/anthropic-leadership.o-o.html.
The retained parts have different consequences. Manifest dates and intervals control local dispatch; the model is asked to use as_of to focus research. Cached facts supply prior context, not independently certified truth. Source records retain pointers rather than page snapshots. The ordinary example uses coarse strings, while sentinel examples contain richer URL/title/tier/access-date objects. Neither a uniform provenance schema nor per-fact acceptance follows. Change summaries report edits without authenticating their execution or preserving restorable prior files. See MEM-OBJ-manifest, MEM-OBJ-source-cache and MEM-OBJ-changelog. Source basis: SRC-1 example/anthropic-leadership.o-o.html, example/agi-declared.o-o.html, example/trump-incarceration.o-o.html.
The agent is asked to reassess source validity and preserve still-valid sources. Image instructions request format/size checks and reduction or rejection. These checks have intended consequences, but their requested execution is not observed. They do not establish factual acceptance of article claims or captions. Article synthesis remains indeterminate between reshaping, derivation and conjecture because linked inputs and outputs are unavailable. No evidence-consuming claim acceptance followed by lifecycle integration is established in this boundary. That is narrower than claiming the external agent cannot evaluate evidence. See EPI-RTE-source-check, EPI-RTE-source-selection, EPI-RTE-image-check, EPI-RTE-image-selection and EPI-RTE-web-research. Source basis: SRC-1 example/index.o-o.html, example/anthropic-leadership.o-o.html.
Changes enter the original file directly, without per-edit human approval or an o-o post-edit acceptance test. Successful CLI return ends the shell step; it does not certify citations, JSON, accuracy or preservation of machinery. The section-preservation instruction is best effort: the same delegate has requested Bash access and no section-aware write filter is interposed. A failure may leave partial edits, with no shipped automatic restore in this route. See RT-RTE-document-update, RT-BAP-update-prompt and EPI-RTE-document-retention. Source basis: SRC-1 example/index.o-o.html.
From per-document work to local delivery
The representative connected workflow is --update-all: the index discovers eligible documents, invokes their Bash paths sequentially and rebuilds navigation after processing at least one document. Each child receives its own target path, not a combined research context. The parent does not forward its model override. Batch selection uses age greater than the interval, whereas direct invocation skips age less than the interval; their exact-threshold behavior differs. A child failure can interrupt the batch before rebuilding, leaving earlier article edits intact but navigation unrefreshed. Recovery belongs to the operator, not a shipped repair worker. See the workflow account and RT-RTE-batch-create. Source basis: SRC-1 example/index.o-o.html.
Delivery is local availability followed by a separate browser action. Browsers consume article HTML, metadata and references; the index retains a lossy projection. Its eight prominent cards are date-ranked, and excerpts strip citation markers and truncate at 120 characters. That can remove essential qualifications; visibility is not research warrant. Recursive discovery also does not ensure nested links work, because links use basenames. Browser theme replay is separate UI memory, not research reuse. See RT-RTE-library-read, MEM-OBJ-index-content, EPI-RTE-index-excerpt and MEM-OBJ-oo-theme. Source basis: SRC-1 example/index.o-o.html, example/anthropic-leadership.o-o.html.
Assessment at the shipped workflow boundary
For an operator-invoked local research-library refresh, the source supports a connected selection, delegation, persistence and reader-entry path. For a scenario requiring independently accepted facts, bounded model context, enforced content-only editing or transactional library recovery, the inspected workflow does not establish those guarantees. Actual quality, cost and successful incremental reuse remain unobserved rather than disproved. These distinctions are supported by RT-RTE-document-update, MEM-BAP-cache-reuse and RT-RTE-library-read, from SRC-1 example/index.o-o.html and example/anthropic-leadership.o-o.html.
Operator configuration edits and explicit shared-code synchronization can change future instructions and executable machinery. This is a concrete maintenance capacity, not evidence-responsive autonomous improvement. No recorded criticism-to-revision-to-later-use chain establishes improved capacity for future action. Agent-owned research decisions establish delegated discretion, not autonomous theory building. The accepted runtime supplies no separate theory-builder conditions 1–4 assessment; this synthesis therefore makes no condition-by-condition determination. Reflection and exercised self-improvement likewise remain unestablished, not inferred absent from external-model opacity. Retained research could affect future behavior, but no observed evidence-shaped organizational update and subsequent causally dependent operation closes the self-improvement chain. See RT-RTE-contract-edit, RT-RTE-code-sync, RT-CMP-claude and the epistemic bounded conclusion. Source basis: SRC-1 example/index.o-o.html, README.md.
An update trace linking fetched evidence, edits and later cache consumption would change the activation assessment. A comparison varying retained-cache access could test reuse benefit at that treatment level. A post-edit validator, enforced write boundary or snapshot/restore mechanism would change the corresponding source-level guarantees. Evidence of criticism shaping retained guidance and later operation would be needed for stronger learning or self-improvement attribution; successful article refresh alone would not suffice.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed update, recovery or causal comparison | SRC-1; RT-RTE-document-update, RT-RTE-batch-create, MEM-BAP-cache-reuse | Shipped shell, contracts and retained examples; no target execution | Actual cost, reliable delivery, successful recovery, reuse activation or benefit | Inspectable update/recovery traces and appropriately controlled reuse comparisons |
| External runtime and model controls uninspected | SRC-1; RT-CMP-claude, RT-BAP-update-prompt | o-o CLI dispatch, not installed Claude/provider internals | Effective permissions, isolation, budget enforcement, pinned model, parameter updates or internal evaluation | Versioned host/provider contracts, deployment configuration and relevant traces |
| Incorrect positional contract guidance, carried from report verification | SRC-1; MEM-OBJ-contract, RT-BAP-update-prompt; declared limit | Ordinary and sentinel operative contracts precede window.stop(), contrary to guidance |
Guarantee that positional guidance identifies or delivers the operative contract | Correct block-location guidance and a trace showing operative-contract reading |
| Requested whole-file reads do not establish complete context delivery | SRC-1; MEM-BAP-cache-reuse, RT-OBJ-shared-shell, RT-CMP-claude | Path-first prompt, duplicate template literals, possible media and external read handling | Which facts were read or influential; bounded context and reliable template/state disambiguation | Tool-read and model-context traces, including omission/compaction and actual selected blocks |
| Research warrant and production provenance unresolved | SRC-1; MEM-OBJ-source-cache, MEM-OBJ-facts, EPI-RTE-web-research | Selected article material and retained caches; external pages not acquired | Factual accuracy, complete lineage, checked derivation or accepted novel claims | Source contents, claim-linked inputs/outputs and evidence-consuming acceptance records |
| Schema variation and summaries do not supply an audit checkpoint | SRC-1; MEM-OBJ-source-cache, MEM-OBJ-changelog, MEM-OBJ-manifest | Ordinary versus sentinel retained structures and generated defaults | Uniform provenance schema, authenticated change history or restorable prior versions | Validated instance schemas, execution-linked histories and prior-state snapshots |
| Media semantics and provenance uninspected | SRC-1; MEM-OBJ-images, EPI-RTE-image-check | Located encoded payloads, readable captions and requested format/size checks | Decoded-media classification, authenticity, caption truth or demonstrated resizing compliance | Decoded media inspection, provenance and executed check results |
| Direct admission and operator recovery bound safety claims | SRC-1; RT-RTE-document-update, RT-RTE-contract-edit, RT-RTE-code-sync | Complete shipped dispatch, regex configuration edits and marker-based replacement | Content-only enforcement, semantic configuration validation, behavioral correctness of replacements or automatic rollback | Enforced write checks, semantic validators, tested replacements and restore evidence |
| Reader projection is not a faithful evidence-bearing mirror | SRC-1; EPI-RTE-index-excerpt, RT-RTE-library-read | Citation-stripping/truncating excerpts, date-ranked cards, basename links and static browser code | Guaranteed semantic preservation, nested navigation correctness or demonstrated rendering | Scope-preserving excerpts with citations, path-preserving links and browser traces |
| Trace-fed learning and stronger improvement properties remain unresolved | SRC-1; MEM-OBJ-changelog, MEM-OBJ-facts, RT-CMP-claude, RT-RTE-code-sync | Product summaries, reuse instructions, explicit operator maintenance and opaque host context processing | Positive or bounded-negative trace-learning classification, trace-source inventory, demonstrated learning or exercised self-improvement | Qualifying trace-to-retained-update-to-consumer evidence; for improvement, objective-linked comparisons and later causal dependence |
The supplied reconciliation identifies no unresolved cross-report conflict. Report verification declares the positional-guidance limit carried above; profile verification declares no additional limit. No source reread or new evidence was used for this synthesis.