BRAN synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From a request to bounded repository evidence
BRAN is a repository context system, not the enclosing agent responsible for deciding and implementing a task. Its ordinary contribution is to locate, rank and cite evidence for that agent. This analysis covers the shipped public repository at 7bc428b2ccdfad38e05caa0c0cba2bb19041993a, with an applicability cutoff of 2026-11-08. The whole-system boundary includes CLI/TUI drivers, document admission, persistence, maintenance, shipped consumption instructions and source-defined evaluators. The evidence is implementation and doctrine/design from SRC-1; no observed run or causal experiment is registered. External hosts, providers, compressors, index production and live deployment remain outside the frozen source. See the boundary and evidence, Source register and task-ownership claim RT-CLM-evidence-location.
A caller supplies a repository root and request to query or packet. Rust code scans files, parses metadata, constructs graph input, matches request terms against paths, metadata and bodies, orders results, compiles a view and expands bounded dependencies. Packet assembly adds item, byte and estimated-token limits. Ordinary evidence can be omitted; required evidence that cannot fit causes an error. The result separates natural-language evidence from symbolic selection accounting, locators, digests, warnings and truncation information. This is deterministic lexical and metadata selection, not embedding retrieval or a model choosing the next action. RT-RTE-query-packet, RT-OBJ-context-packet, RT-OBJ-packet-receipt.
Native documents contribute separately admitted anchors and optional excerpts. Embedded matches are score-ordered; OCR-derived positives follow without the same byte-derived ranking. Excerpt budgets can further reduce what reaches the reader. An existing SCIP index can supply symbol navigation, filtered to retained sources; stale indexes supply no symbol evidence, and unverified freshness remains partial. These are distinct evidence paths, not proof that all repository information reaches the agent. MEM-RTE-document-selection, EPI-RTE-document-admit, MEM-RTE-symbol-reload.
Context control happens before invocation. BRAN selects a bounded by-value payload rather than compacting an accumulated conversation or demonstrating an isolated subagent frame. Omitted files remain on disk for an explicit follow-up query; their locators do not automatically reload them. The shipped skill affords an outer-agent consumption route, while connected execution wires selected evidence into a provider request. Neither delivery establishes actual reliance. RT-BAP-evidence-supply, memory read-back account.
For evidence-location work, this architecture provides inspectable selection reasons, bounded output and explicit reload opportunities. It does not establish that the selected files are the right or complete evidence for a task. The README's “authoritative files” and “declared authority” language describes selection claims, not independently verified truth. Even byte-preserving excerpts can omit a necessary qualification: the epistemic report derives a prefix-truncation case in which an access restriction disappears. That is a source-code deduction, not an observed execution. EPI-CLM-authoritative-files, EPI-CLM-declared-authority, bounded epistemic conclusion.
Optional connected work: delegation, veto and delivery
The representative multi-step workflow is production connected bran -p: it joins request admission, repository selection, external delegation, answer checks and retained-result delivery in one shipped call chain. Connected project settings and explicit root trust are prerequisites. The CLI constructs a task-plus-packet prompt and a grounding contract containing admitted locators, source digests and preservation anchors. Space is reserved for the prompt prefix and requested output before evidence assembly. Empty grounding does not silently become an ungrounded connected request. Code owns progression; the host proposes an answer. RT-RTE-connected-request, workflow account.
Per attempt, BRAN invokes an external provider host through a preflight/execute/result protocol. The production grant set permits read/search, denies mutation and supplies depth controls. Executable hashing, framed I/O, cleared environment, timeout and cancellation constrain the local protocol; they are not an OS filesystem/network sandbox. Echoed model, session and budget fields require an honest host and do not independently identify model weights or effective tool confinement. There is no inspected BRAN outer planning/tool-use loop behind this single invocation per attempt. RT-RTE-connected-request, RT-CMP-agent-model, runtime capability account.
Grounded completion has an enforcing but narrow veto. It requires admitted citations, answer reconstruction from claims, material claim/support equality, fresh-file digest checks, support occurrence in current source bytes, protocol/surface checks and successful storage. Output compression cannot change the grounded answer. This supports source-byte correspondence at completion, not source truth, complete propositions, explanatory adequacy or a correct outer-task answer. Support occurrence is a substring check. The local reference is the admitted repository, not an independent expected solution. RT-RTE-connected-request, EPI-OBJ-provider-claims.
The guarantee is route-specific. Generic AgentRuntime callers can omit grounding and token ceilings. The separate structural ConnectedCoordinator checks packet identity and coverage rather than fresh-file exact support; rejected grounding empties accepted returned bytes but preserves the raw response in the outcome. Neither library alternative demonstrates a bypass of the production CLI builder. Optional SQZ checks required substrings, DLP and bounds, then selects byte reduction—not semantic equivalence or task benefit. Default deterministic packet selection needs no provider, but a configured external SQZ executable prevents a universal no-network guarantee. RT-RTE-runtime-port, RT-RTE-structural-connected, EPI-RTE-sqz-check, RT-RTE-query-packet.
Recovery is specific, not a general answer-revision critic. An incomplete grounding failure with input-SQZ missing anchors can halve the source limit only when the missing list is nonempty and entirely source-prefixed, the current limit exceeds one, and rebuilding changes the prompt. This trades evidence breadth for a smaller compression input; no supplied run establishes successful recovery or preserved answer quality. After runtime completion and storage, an excessive response citation count can still produce a CLI error. Runtime completion, command delivery and outer-task success are different terminals. workflow account.
What persists, and what feedback can change
Accepted connected answers, citations/support and opaque artifacts are retained separately in local result batches. The filesystem store has entry, byte, per-item and age bounds, with integrity checks and eviction. Explicit get decodes an inline result by content identity; it does not revalidate current cited files or automatically insert old answers into a new prompt. Opaque artifacts have immediate integrity read-back, but no dedicated later agent consumer is established. Crucially, --no-session is a host-session protocol request/echo, not suppression of BRAN-local result publication. RT-CLM-session-retention, RT-RTE-connected-request, MEM-OBJ-result-answer, MEM-OBJ-result-support, MEM-OBJ-lossless-artifact.
BRAN's strongest supported feedback contribution is retained query-use diagnostics. Opt-in recording keeps the request and selected outcome/ranking descriptors. Later summarize/propose commands compare history with current inventory; replay reruns recorded requests and reports differences. These are wired retrospective consumers across invocations. They do not establish improved ranking or knowledge: proposals leave replacement and authority null, replay does not overwrite the baseline, and a changed result need not be better. Document matches can affect coverage without being preserved as a complete anchor-level history. MEM-RTE-evidence-feedback, MEM-OBJ-evidence-rankings, EPI-RTE-evidence-summary, EPI-RTE-evidence-replay.
Two other routes change retained behavior-relevant material. TUI Apply admits operator-edited effective configuration, consumed by later packet/connected setup. Maintenance can replace source bytes, but CLI apply requires the exact fixture marker, digest, asserted authority and strict revalidation; core rollback can itself fail on I/O. Asserted authority is not authenticated human approval. A separate derived-only helper can construct automatic authority and publish deterministic outputs through repair, but its production caller and later consumer remain unresolved. Hash indexes, raw snapshots, truncated reports and validator-framed bytes retain different information; a GeneratedValidator label does not establish learned executable checks. Generic state storage, database extraction and immediate Obsidian export/reparse likewise do not by themselves establish an integrated learning loop. RT-RTE-settings-apply, RT-BAP-settings-consumption, RT-RTE-maintain-apply, MEM-RTE-derived-rebuild, MEM-OBJ-file-state, MEM-CMP-database-evidence, MEM-CMP-obsidian-export.
The supported result is diagnostic feedback plus configuration/source-change capacity, not demonstrated improvement of future action through criticism of a consumed theory. No comparison establishes that benefit. The accepted records do not establish a complete formulated-theory criticism/iteration account; conditions 1–4 are therefore not assigned a system-wide yes or no here. Reflective representation with both causal directions, autonomous theory-building roles and exercised self-improvement remain separately unestablished, not bundled into an absence claim. Retention and revision alone close none of those questions. The verified memory assessment likewise leaves trace learning and its source not determinable rather than asserting either a positive or a whole-system negative. epistemic conclusion, memory limitations, report verification, profile verification.
An observed production trace could establish exercised checks, recovery and later consumption. A feedback-to-revision trace with an explicit consumer and an outcome comparison could change the improvement assessment; component attribution would require independent variation of that component. Completing the alternative API/evaluator consumer inventory could change coverage without invalidating the inspected production route. Full parser/index audits and real host/deployment evidence would address different questions: fidelity, currency and effective enforcement. None is supplied by command success or a source-defined evaluator's presence.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Native conversion and index fidelity are incompletely assessed; carried from report verification. | SRC-1; MEM-RTE-document-selection, MEM-RTE-symbol-reload, EPI-RTE-document-admit; verification limits | Document admission/selection and CLI existing-index consumption, not comprehensive parser conformance or source-to-index derivation. | Complete format-fidelity and live index-currency comparisons. Transport, budgets and stale refusal remain supported. | Parser/conformance inspection and bounded source/index provenance and freshness evidence. |
| API and evaluator inventory is partial; carried from report verification. | SRC-1; MEM-RTE-derived-rebuild, MEM-OBJ-file-state, MEM-CMP-database-evidence, epistemic coverage; verification limits | Inspected APIs, CLI and bounded consumer searches; downstream/evaluator routes remain unresolved. | Complete comparison of shipped evaluation/adoption routes or system-wide absence of learning and knowledge acceptance. | Recorded material callers, evaluator decisions, payload origins and later consumers, with route-specific admission effects. |
| No observed operation or causal comparison is registered. | SRC-1; Source register, RT-RTE-connected-request, MEM-RTE-evidence-feedback | Source implementation and doctrine only; no target execution. | Exercised reliability, actual reliance, successful recovery, retrieval benefit or causal improvement. | Inspectable execution traces; outcome comparisons and interventions appropriate to the claimed attribution. |
| External execution and deployment contracts are not independently attested. | SRC-1; RT-CMP-agent-model, RT-RTE-connected-request, EPI-RTE-sqz-check, boundary exclusions | BRAN adapters, protocol and shipped deployment configuration; host/provider/compressor/index producer internals and live infrastructure excluded. | Actual model version, host confinement, external session erasure, universal no-network operation or deployed isolation effectiveness. | Frozen external implementations and operational attestations/traces covering the specific claim. |
| Byte and structural checks have narrower domains than semantic warrant. | SRC-1; EPI-OBJ-provider-claims, RT-RTE-structural-connected, EPI-RTE-sqz-check, epistemic conclusion | Exact-support substrings, structural associations, preservation anchors and bounded prefixes. | Source truth, complete propositions, semantic compression equivalence, exhaustive relevance or accepted novel knowledge from those checks alone. | Evidence-consuming semantic evaluation against a named criterion, intended use and scope; checks preserving load-bearing conditions. |
| Historical results and opaque payloads have unresolved later authority. | SRC-1; MEM-OBJ-result-support, MEM-OBJ-lossless-artifact, MEM-OBJ-generated-validator, RT-RTE-runtime-port | Local canonical retrieval/integrity and generic byte/port APIs, not every later semantic consumer. | Current-source re-attestation by get, automatic recurrent prompt use, complete payload/form/authority inventory or validator learning from names alone. | Specific later-consumer wiring and payload provenance; fresh-source checks at retrieval for current-support claims. |
| Diagnostic feedback does not establish an improvement or theory-building loop. | SRC-1; MEM-RTE-evidence-feedback, EPI-RTE-evidence-summary, RT-RTE-settings-apply, RT-RTE-maintain-apply | Retrospective diagnostics, configuration admission and bounded source replacement; no traced criticism-to-adoption-to-benefit chain. | Improved future capacity from theory criticism, exercised self-improvement, or complete affirmative/negative reflection, autonomy and theory-builder assessments. | Separate evidence for formulated criticism and iteration, reflective causal links and autonomous roles; objective-responsive organizational update, later exercised dependence and independent outcome evidence for successful improvement. |
No source checkout files were read for this synthesis; the account uses the supplied accepted members. Profile verification declares no additional limits, and reconciliation establishes no unresolved report conflict.