MiroShark synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From scenario to running population
MiroShark is an application that builds document-derived agent populations, runs social-media and prediction-market simulations, and exposes retained results through reports, interviews and graph retrieval. This account covers the shipped application, bundled Wonderwall runtime, drivers, clients and graph-memory stack at commit e95d3d8c2a2206b8b6695648e8504cf23436d3de, with an applicability cutoff of 2026-10-15. SRC-1 supplies inspected implementation and attributed design, not observed execution or causal experiments. No target tests, providers or services were run. The whole-system boundary identifies included responsibilities; it does not mean exhaustive inspection of every route. Dependency internals, live remote content and deployed state were not captured (boundary and Source register).
A caller supplies documents and a scenario requirement. An optional question-to-briefing endpoint can supply seed text, but does not itself ingest that text. Graph ingestion extracts entities and relations, resolves identities and retains text, metadata and embeddings. Preparation then selects defined entities with edges, generates personas and parameters, and saves platform profiles and configuration. Software admits generated content without a separate per-persona human approval. Optional research enrichment can use search snippets, or fall back to a model call without snippets. These routes establish context-conditioned population construction, not verified biographies or representative population behavior (EPI-RTE-ask-briefing, MEM-RTE-graph-ingestion, EPI-RTE-web-research, RT-RTE-prepare-population).
The server starts a shipped Python driver. In the synchronized route, code advances rounds and samples active agents; SocialAgent supplies persona, selected context and an environment observation to a model that chooses platform tools. Environments execute those actions against simulated platform state. Previous-round context precedes concurrent platform work, so synchronization does not establish same-round sequential awareness. A round timeout is best effort: synchronous work can block cancellation, and partial platform effects need not roll back. Dedicated single-platform scripts are alternate routes, not proven behavioral equivalents (RT-RTE-synchronized-rounds).
The effect boundary is not universally confined to simulated platforms. Interviews call the model directly rather than using the ordinary action loop; manual actions search the environment's full function list. A default-off MCP bridge can execute registered external tools with inherited host environment. Its synchronized caller grants a registry catalogue to eligible IDs, rather than enforcing the narrower per-persona allowlist offered by a separate helper. Requested results are buffered for later activation; delayed fulfillment does not make them unsolicited memory selection. Process working directories and groups are not sandboxes (RT-RTE-interview-manual, RT-RTE-agent-mcp).
What experience changes
The strongest supported learning-related contribution is implemented trace-fed revision of later agent context. Recorded actions become recent history, compacted narratives and targeted cross-platform digests. Exposure and engagement signals revise numerical stance, confidence and trust; checkpoints can reload those selected fields. These are real architectural update-and-consumer paths, not merely stored logs. Whether they improve capacity for future action remains unestablished: no operation or comparison demonstrates beneficial model dependence on the retained changes (MEM-RTE-round-history, MEM-OBJ-cross-platform-log, MEM-RTE-belief-feedback).
These mechanisms retain different things. Round history is resident narrative memory; older detail is replaced by model summaries or fallback counts, then ancient batches. Individual compaction is background work, but ancient merging can be synchronous and can lose previous narrative on failure. Belief checkpoints retain symbolic scalars, not model weights, underlying reasons or exposure hashes. Reload therefore does not restore identical novelty weighting, conversations or round history. The driver automatically pushes common history and agent-ID-selected stance/digests into later system messages. Delivery requests influence through model interpretation; it does not establish compliance. Numerical confidence tracks heuristic social feedback, not calibrated truth (MEM-OBJ-round-memory, MEM-OBJ-belief-state, MEM-RTE-belief-feedback, RT-BAP-agent-context).
Optional graph updating creates a separate feedback path: rendered activity batches enter extraction and later profile/report retrieval. Batch-level belief/observation labels, skipped actions, limited source identifiers and retry behavior prevent complete per-action provenance or exactly-once conclusions. A default fact label is not a truth acceptance decision. Identity matching can enforce canonicalization; old/new textual adjudication can invalidate older relations. Neither verifies the new assertion against independent evidence (MEM-RTE-graph-ingestion, EPI-RTE-entity-resolution, EPI-RTE-edge-adjudication).
Graph retrieval is requested selection by query, filters and ranking. Default edge search excludes invalidated relations, but historical requests, node hits and an unfiltered fallback have different rules. Profile/report adapters can strip kind and provenance before supplying fact strings to a model. Community summaries add sampled overviews, not complete evidence: wording uses at most fifteen member descriptions, while stored membership can be broader. Missing-community browsing can build them automatically; replacement is not atomic and missing-only construction is not continuous refresh (MEM-RTE-graph-retrieval, MEM-RTE-community-rebuild, RT-RTE-storage-mcp).
These positive adaptation findings do not settle theory building, reflection, autonomy or self-improvement. The runtime does not supply a separate assessment of theory-builder conditions 1–4, so this synthesis assigns no system-wide verdict on them. In particular, retained assertions do not prove decision dependence on a formulated theory; social feedback is not formulated content-directed criticism with blame allocation. Whether criticism of a consumed theory improved future capacity is unestablished. The automatic scheduling and updates are supported, but an autonomous theory-builder conclusion and a reflective-system conclusion are not established by those facts. For self-improvement, the supported contribution remains evidence-responsive revision of behavior-shaping context; an independently specified improvement objective and subsequent operation causally depending on the change are not established. This prevents an exercised self-improvement claim without denying the wired adaptation routes (MEM-RTE-belief-feedback, RT-RTE-synchronized-rounds, epistemic bounded conclusion).
From retained simulation to delivered account
Report generation from an existing simulation is the representative multi-step workflow. The API checks prerequisites, can reuse a completed report, or starts a background task. A model proposes an outline; planning failure supplies a default outline. Code delegates sections to parallel workers with local histories and no sibling drafts. Each worker can request graph, action, trajectory or interview evidence. Only the first tool call per response executes. The normal completion branch requires two calls, but eight-iteration exhaustion forces generation without that same minimum. Local checks regulate activity and response shape, not factual sufficiency (RT-RTE-report-sections, WF-RTE-report-delivery).
The coordinator retains returned section text, including error placeholders. Optional cross-section synthesis sees only the first 3,000 characters per section and cannot veto or revise the originals; synthesis failure need not stop assembly. Delivery means saved report data and Markdown reachable through retrieval/download endpoints, not the initial start response or user receipt. Completed status can coexist with section errors and omitted synthesis. Optional graph traces expose emitted report thoughts and truncated observations to explicit inspection, not private provider reasoning or an established automatic prior-report learning route (WF-RTE-report-delivery, MEM-OBJ-reasoning-trace).
For exploring a caller-defined scenario, the implementation supports a population/action/context/report progression. Its report guidance explicitly frames results as plausible dynamics under assumptions, not calibrated forecasts. Stronger downstream labels require separate warrant. Quality badges measure recorded participation and diversity, not predictive accuracy. Binary resolution compares one selected market consensus with an administrator-supplied outcome. Public outcome annotation admits a label and optional URL without fetching the URL's evidence; verified-only filtering consumes annotation presence. Trading adapters map aggregate stance and health into probability-, confidence- and risk-shaped fields. Their arithmetic is inspectable, but does not establish event calibration or measured financial risk (EPI-CLM-scenario-caveat, EPI-RTE-quality-diagnostics, EPI-RTE-prediction-resolution, EPI-RTE-outcome-annotation, EPI-RTE-signal-derivation).
A captured end-to-end run would change conclusions about exercised routes, not automatically predictive validity. Summary-to-input checks could establish fidelity for sampled transformations. Independently authenticated outcomes and calibration comparisons could change the forecast assessment. A controlled intervention on retained context, with later consumer traces and an independently specified objective, could establish a benefit or component effect at the comparison's actual scope. Evidence-consuming report acceptance and provenance-preserving delivery would change the present limits on factual reliance; universal withdrawal would additionally require coverage of fallback, derived and copied content.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed execution or causal comparison | SRC-1; RT-RTE-synchronized-rounds, WF-RTE-report-delivery | Frozen application implementation and design; no live providers, database or user artifacts | Successful operation, actual uptake, reliability, cost/latency, improved capacity or causal benefit | Provenance-linked execution artifacts; objective-specific comparisons with appropriate intervention controls |
| Market/media feedback is omitted from the detailed memory inventory; carried report-verification limit | SRC-1; report-verification limits, MEM-RTE-round-history, MEM-RTE-belief-feedback, RT-RTE-synchronized-rounds | Verifier inspected backend/scripts/market_media_bridge.py and synchronized caller: database-price snapshots and stance/action-derived sentiment feed later prompts |
Exhaustive context-transformation, cross-platform aggregation and recovery-equivalence comparisons; narratives, digests and beliefs are not the only feedback | Canonical records tracing market/media retained parts, updates, prompt consumers and recovery alternatives |
| Auxiliary and dependency alternatives remain partial | SRC-1; memory coverage, epistemic coverage, RT-RTE-agent-mcp | Dedicated-driver parity, full analyzer/initialization control, acquisition and auxiliary publication/export/cache routes, opaque host/provider state not completely assessed | Complete memory/authority inventory, universal controls, full recovery parity, source-complete grounding or system-wide absence of evaluation/learning | Scoped inspection and records for the unresolved routes; captured remote contracts and payload lineage where needed |
| Lossy context and incomplete restoration | SRC-1; MEM-RTE-round-history, MEM-OBJ-belief-state, MEM-CLM-observation-equivalence | Compaction, fallback, checkpoint and current-observation selection | Faithful preservation of all experience, semantic equivalence of filtering or behaviorally identical resume | Source/output fidelity checks, failure traces and restoration comparisons covering omitted fields and context |
| Graph categories and withdrawal have consumer-specific force | SRC-1; MEM-RTE-graph-ingestion, MEM-RTE-graph-retrieval, MEM-RTE-community-rebuild, EPI-RTE-report-cleanup | Ingestion, default/historical/node/fallback reads, sampled communities and file deletion | Verified ground truth, precise exactly-once provenance, atomic community replacement, universal freshness/retraction or all-store erasure | Per-claim provenance and acceptance evidence; transactional/replay checks; consumer-wide invalidation and deletion coverage |
| Report completion and trace retention do not establish factual acceptance | SRC-1; RT-RTE-report-sections, WF-RTE-report-delivery, MEM-OBJ-reasoning-trace | Local tool-count gate and bypass, placeholders, truncated synthesis, optional recorder | Sufficient evidence in every completed report, full trace retention or automatic learning from prior traces | Actual report/tool artifacts and evidence-consuming acceptance criteria; recorder coverage and a separately traced later learning consumer |
| Appraisal and trading semantics exceed their demonstrated warrant | SRC-1; EPI-RTE-quality-diagnostics, EPI-RTE-prediction-resolution, EPI-RTE-outcome-annotation, EPI-RTE-signal-derivation | Activity-health cache, caller reference/annotation and conditional adapter arithmetic; summary producer incompletely traced | Independent outcome verification, calibrated event probabilities, measured financial risk, actual trading effects or mechanism attribution from score success | Authenticated aligned event references, complete producer traces, calibration/risk studies and observed downstream consumption |
| Independent higher-order properties remain unestablished | SRC-1; MEM-RTE-belief-feedback, RT-BAP-agent-context, epistemic bounded conclusion | Wired scalar/context adaptation without a runtime theory-condition assessment or exercised-benefit evidence | Theory-builder conditions 1–4 as a system verdict, criticism-led improved capacity, reflective/autonomous theory building or exercised self-improvement | Separate condition-specific evidence; self-representation and both causal directions for reflection; role-level autonomy evidence; objective-bearing update and later causal-dependence evidence for self-improvement |