Synapptic synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From coding sessions to reusable guidance
Synapptic implements a durable trace-to-guidance pathway: recorded coding conversations feed observations, weighted profiles and narrative instructions; retained profiles also shape later extraction. A separate benchmark can change which guards qualify for future synthesis. These are implemented contributions, not demonstrated improvements in future coding work. The analysis covers the shipped repository at commit bcde026770daa5cd9a8ca18ca7182b0aeaa87d1f, with an applicability cutoff of 2026-10-15. Its source, SRC-1, supplies inspected implementation and first-party doctrine, but no observed-run or causal-experiment evidence. No target code, tests or model calls were executed. The whole-system boundary includes the optional relay and operator browser; it excludes external assistants' instruction-loading behavior and model-service internals. See the boundary and evidence and Source register.
The claimed work is learning preferences and assistant weaknesses from sessions, then supplying preventive rules to later sessions. Architecturally this is a memory and context-processing system, not the coding assistant's enclosing reasoning-and-tool runtime. Python selects work, sequences stages and writes files. Models propose observations and narrative wording. External assistants retain ownership of coding decisions, tools and instruction precedence. The representative source-supported workflow is operator synapptic ingest, with Claude Code publication as one concrete branch; it is a reconstructed implementation path, not a successful execution trace. See RT-RTE-ingest-profile, RT-RTE-publish-archetypes and the workflow account.
Ingest discovers sessions not represented by observation filenames, optionally restricts the project, and processes them in descending file-size order. Filtering retains user and assistant text, prioritizes correction/preference/reaction turns and recent content, and omits tool-result bodies and thinking. This creates a separate extraction context rather than compacting the live assistant window. The character-based budget is approximate; template and profile overhead are added afterward. Omitted evidence remains in the original file if available, but the extractor receives no retrieval tool to reopen it. Internal model calls disable tools or omit tool definitions; this is a provider-contract restriction, not a sandbox for Synapptic's local file, subprocess or network effects. See RT-RTE-ingest-profile and the runtime guarantees account.
Existing profile text can enter extraction as already-established context. The ingest path activates this only when the global profile has dimensions; a project-only profile does not suffice. Selection uses weight at least 0.7 and the first eight entries per dimension after global-first combination, not task-semantic retrieval or a globally re-sorted top eight. Exclusions and model verdicts do not filter this summary. Models therefore receive selected retained claims without a new truth check. See RT-RTE-ingest-profile and its memory annotation.
Admission, retention and publication have different force
The observation parser rejects unusable JSON shapes, unknown dimensions and missing observation text. It does not require or verify supporting evidence, calibrate confidence, or independently adjudicate a failure diagnosis. A user correction establishes that a correction was uttered, not that its technical explanation is correct. The model's semantic transformation remains indeterminate without generated instances: the evidence does not distinguish faithful reshaping, entailed diagnosis and extrapolation in actual outputs. This is operational admission of candidates, not evidence-consuming acceptance for factual reliance. See EPI-RTE-extract-candidates and EPI-RTE-observation-admission.
Per-session JSON may retain a model-supplied evidence string. Newly constructed profile entries retain observation text, weights, counts and coarse session lineage, but omit that separate evidence field. Synthesis reads profiles rather than reopening those evidence files. Similarity matching reinforces entries and can replace wording with longer text; decay and pruning reduce availability; cross-project grouping can promote entries to global scope. These operations change future reliance without establishing semantic agreement, independent corroboration or contradiction resolution. Counts can increase twice from one session, and capped source lists are not a permanent once-only merge ledger. See MEM-OBJ-observations, MEM-OBJ-profile-observations, MEM-OBJ-profile-metadata and EPI-RTE-profile-curation.
Synthesis selects guards and assistant-failure entries at weight 0.5; ordinary dimensions use weight 0.3 and evidence count one. Global exclusions apply, while model-specific redundant/backfire verdicts apply only when an explicit target model is supplied. The model is asked to produce a user archetype, imperative guards, reasons and known weaknesses. Nonempty prose is accepted without a grounding, heading or word-count gate. Requested failure-grounded reasons are therefore stronger than the storage chain guarantees. See EPI-RTE-narrative-inputs, MEM-OBJ-archetype and EPI-CLM-failure-grounding.
Integration reloads saved global/project narratives, combines them, and dispatches configured writers to discovered destinations. Claude Code receives user_archetype.md and a reference in an existing MEMORY.md; a missing index is not created. Other writers use always-apply rules, marked instruction sections or whole-file outputs. Local publication is wired. Later host loading and instruction influence are afforded, not observed or enforced by Synapptic. Integration computes model-filter counts but writes the saved narrative unchanged; explicit target-model synthesis is the path that actually filters its source entries. See RT-BAP-archetype-host and RT-CLM-target-filtering.
Delivery also has stage-specific failure boundaries. Empty extraction results can create processed markers that ordinary ingest subsequently skips. Failed or empty synthesis can leave an older narrative available for integration. Integration's five-element update tuples conflict with ingest's three-name unpacking, so a nonempty update list can produce an exception after output writes. These are inspected conditional defects, not observed incidents. Per-file atomic replacement does not make the pipeline transactional. Global rollback, reset and uninstall do not jointly retract every saved narrative or published platform copy. See the workflow terminal-state account and MEM-RTE-state-maintenance.
The installed SessionEnd hook is a different route: on selected exit reasons it detaches extraction, merge and synthesis, without integration or success chaining. A live PID can cause an event to be skipped without scheduled retry. Automatic internal maintenance is therefore supported, but automatic refreshed external guidance after every session is not. See RT-RTE-session-hook and EPI-CLM-background-learning.
Feedback changes eligibility, not established warrant
The benchmark is consequential: models generate scenarios, respond under alternative profile conditions, and judge textual rule compliance; code derives categories and writes per-model verdicts before optional human confirmation of global exclusions. Later explicit-target synthesis can consume those verdicts. Re-inclusion removes global exclusion flags, not model verdicts. Detailed result reasons are retained separately rather than automatically consumed by synthesis. See RT-RTE-benchmark-feedback, EPI-RTE-benchmark-disposition and MEM-OBJ-benchmark-results.
The comparison's scope matters. Per-guard mode heuristically removes a matching narrative block or adds a missing guard; default batch mode contrasts the whole archetype with no archetype. Batch per-rule labels cannot isolate individual guard effects. Judges see response snippets—2,000 characters per-guard and 300 in batch—not coding outcomes. Same-model judging, resumed condition histories, mode-dependent cache freshness, respondent instructions that disclaim following embedded profile instructions, and possible positional shifts after missing batch verdicts further restrict interpretation. Categories such as effective or backfire follow scoring conventions; they do not establish prevention or regression in real tasks. See EPI-RTE-benchmark-judgment, MEM-OBJ-benchmark-cases, MEM-OBJ-claude-session and EPI-CLM-individual-impact.
The optional relay separately forwards external host requests and retains SQLite metrics, derived request displays, transcript indexes and session bookkeeping for operator read-back. It does not inherit internal inference's tool restriction. Browser search is requested retrieval over indexed titles/first-user snippets, not full-conversation semantic search or demonstrated automatic profile learning. Raw transport, stored display and reduced browser view are different representations. An OpenAI session-return mismatch prevents assuming provider-handler equivalence. See RT-RTE-relay-forwarding, MEM-OBJ-request-log, MEM-OBJ-session-index and MEM-RTE-browser-readback.
Independent capacity assessments and what would change them
For repeated-session preference accumulation, the supported result is automatic durable trace-fed updating with later prompt consumers. For verified failure diagnosis, isolated guard efficacy or fewer recurrent coding failures, the present evidence is insufficient. Retention, heuristic eligibility and publication do not establish accepted ampliative knowledge or post-acceptance integration in the traced routes. This is not a whole-repository negative about every possible truth-apt transformation. See the epistemic bounded conclusion.
The runtime separately assesses theory-builder conditions: 1, localized guidance: afforded; 2, decisions depending on that guidance: afforded; 3, content-directed criticism naming claim, result and blame: uninspected; 4, criticism shaping another round: afforded. None is inferred from another. Editable profile entries and whole narratives provide localized guidance surfaces, but no supplied generated criticism closes the criticism/revision cycle. Benchmark scores changing reliance do not themselves supply that criticism. Consequently improved future capacity from criticism of a consumed theory is not established. See RT-RTE-publish-archetypes.
Reflection has a representational return path across the host-plus-Synapptic boundary—host failures can inform guidance returned to the host—but representation-mediated behavioral change is unobserved. Automatic stage sequencing is wired; autonomous theory building is not established role by role, independently of whether the workflow needs human per-item approval. Self-improvement toward fewer recurrent coding failures across future host sessions remains dispositional at that wider boundary: trace-fed updates exist, but later live behavioral dependence and successful improvement are unestablished. Synapptic alone does not own the external host's behavior. These conclusions do not negate the wired trace-fed update mechanisms. See RT-RTE-publish-archetypes and RT-RTE-benchmark-feedback.
Retained transcript-to-observation-to-profile-to-narrative histories would permit fidelity and rationale assessment. Traces linking published versions to actual host context and later decisions would address activation. Matched coding-outcome comparisons with independently varied guards would address benefit and component attribution. Preserved claim/result/blame/revision artifacts consumed in another round would address theory criticism. Implementation changes that propagate target filtering into published text, bind narratives to profile versions and repair retry/delivery contracts would change the source-level assessment of those paths; they would still need separate execution and outcome evidence for reliability or improvement. The report verification supports the bounded implementation account; the profile verification adds no publication limit.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied operation or outcome evidence | SRC-1; RT-RTE-ingest-profile, RT-RTE-publish-archetypes, RT-RTE-benchmark-feedback | Frozen source implementation and doctrine; no executed target work | Successful end-to-end operation, actual profile accuracy, benefit, cost or long-term reliability | Retained real execution histories and appropriately controlled outcome comparisons |
| External host activation and causal dependence are unobserved | SRC-1; RT-BAP-archetype-host, RT-RTE-publish-archetypes | Local publication, not external host internals | Actual instruction loading, precedence, compliance, exercised reflection or self-improvement | Version-linked host context/decision traces and subsequent operations causally depending on the update |
| Generated content and retained reasons are not demonstrated | SRC-1; MEM-OBJ-observations, MEM-OBJ-profile-observations, MEM-OBJ-archetype, EPI-RTE-extract-candidates | Shape admission, evidence-dropping profile constructor and profile-only synthesis input | Guaranteed grounding, semantic preservation, novel accepted claims or completed content-directed criticism | Linked generated artifacts with original evidence, explicit checks and criticism/revision records |
| Synthetic comparisons do not establish isolated guard efficacy | SRC-1; EPI-RTE-benchmark-judgment, EPI-RTE-benchmark-disposition, MEM-OBJ-benchmark-cases | Whole-profile batch and heuristic per-guard contrasts, truncated judgments, cache and positional failure paths | Individual causal effects, reliable attribution of every batch label, or transfer to live coding | Identity-preserving complete trial traces, validated judgments, fresh matched treatments and external task outcomes |
| Opaque inference and resumed evaluator state | SRC-1; RT-CMP-call-llm, MEM-OBJ-claude-session | Unpinned external services and resume handles | Exact weights, full retained payload/form/provenance, checkpoint maintenance, fresh isolated trials or complete hidden write/consumer classification | Pinned service contracts and inspectable retained-state/admission histories |
| Freshness, recovery and withdrawal differ by stage | SRC-1; RT-RTE-ingest-profile, RT-RTE-session-hook, RT-CLM-target-filtering, MEM-RTE-state-maintenance | Processed markers, detached maintenance, saved-text integration and local maintenance | Guaranteed retry, every-session publication, effective target-specific output filtering, transactional rollback or all-consumer retraction | Corrected source contracts plus stage/version-linked retry, publication and withdrawal traces |
| Privacy and isolation are not comprehensive guarantees | SRC-1; RT-RTE-ingest-profile, RT-RTE-relay-forwarding, MEM-OBJ-request-log | Transcript scrubbing and text-only internal calls versus unscreened model-output persistence and raw relay transport | Complete confidential-data exclusion, a system sandbox or equivalent security across routes | Audited end-to-end data-flow and deployed permission/retention evidence covering each route |
| Batched evaluator debug traces are omitted from the reports' inventory | SRC-1; MEM-OBJ-benchmark-results, EPI-OBJ-benchmark-results, RT-RTE-benchmark-feedback; report-verification limit | src/synapptic/benchmark.py retains raw respondent/judge returns and truncated prompt displays under benchmarks/traces |
Complete evaluator retention, deletion and confidentiality inventory; omission establishes no additional behavior-shaping consumer | Records covering those files' production, retention/deletion and actual later consumers |
| Relay and truth-apt coverage are not exhaustive | SRC-1; MEM-RTE-browser-readback, MEM-OBJ-session-index, RT-RTE-relay-forwarding; epistemic boundary | Selected forwarding, persistence and browser routes; incomplete live UI, metrics/parser and full test characterization | Universal display fidelity, equivalent OpenAI persistence, complete memory-route enumeration or system-wide absence of epistemic routes | Further bounded source records for material live UI/measurement alternatives, plus deployment traces where operation is at issue |
| Synthetic feedback's controlled trace-origin class remains unresolved | SRC-1; RT-RTE-benchmark-feedback, MEM-OBJ-benchmark-results; profile verification | Automatic activity-fed benchmark eligibility updates; session-log origin is resolved only for conversation extraction | Complete trace-source classification; API wrappers alone do not establish tool-trace origin | Accepted records locating original evaluator input content within the relevant source classes |