academic-research-skills synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Entering a host-mediated research workflow
Academic-research-skills (ARS) ships research, writing, screening and review workflows, not an enclosing inference runtime. This account covers the whole shipped package at Git revision 060cb3cbffc2ca1bdfb8f7c6d2ffa4883cf09b1a, with an analysis cutoff of 2026-10-17. The evidence is source inspection: executable helpers and adapters supply implementation evidence; model-role Markdown supplies doctrine/design. No target execution, independently established run trace or causal experiment supports this analysis. The host, model providers, scholarly services and user-owned source truth remain external. See SRC-1 in the Source register and the boundary account.
A user starts /ars-full; the host loads instructions and the foreground model orchestrates specialized roles. Pi transforms the command into input text and restores a branch-local activation bit. It does not execute a deterministic stage engine. The host supplies scheduling, tools and inference; the user controls mandatory checkpoints and author choices. Model aliases identify declared routing, not immutable weights or per-call attestations. RT-RTE-command-dispatch, RT-RTE-pipeline-stages, RT-CMP-session-model.
The representative route starts without research materials: research produces a question, methodology, bibliography and synthesis; writing consumes them to produce a draft; integrity checking precedes full review; a Minor/Major decision leads to author-scoped revision, re-review, final integrity and formatting. Screening is separately requested, not silently inserted into this route. This is a supported possible progression, not evidence of a successfully completed session. The workflow account traces the handoffs and delivery conditions; RT-RTE-pipeline-stages carries their protocol status.
Producing and reusing research material
ARS specifies substantive intellectual work. Source-verification roles grade evidence and request cross-source checks; synthesis requests integrated findings, contradiction resolution and interpretive constructs. That is more than a formatting specification. Without actual inputs and outputs, however, the synthesis relation remains indeterminate: the account cannot decide which claims are imported, reshaped, entailed or conjectured, nor establish acceptance of a new scientific finding. Proposed pair resolutions remain pending scholar confirmation. EPI-RTE-source-grading, EPI-RTE-synthesis-content.
Retained corpus records provide another input. Folder imports can retain filename-derived metadata and a file pointer; screening exports retain citation metadata and decision reasons. Neither establishes possession or reading of full text. Later bibliography roles re-screen supplied records under the current research question and merge relevant inclusions with search results. They treat abstracts and notes as untrusted data, cannot mutate the supplied corpus, and fall back to external search on parse failure. A screening inclusion therefore does not force manuscript inclusion. MEM-OBJ-literature-corpus, MEM-RTE-corpus-reuse.
In the separate screening workflow, a confirmed protocol guides independent A/B role calls, conflict adjudication and exclusion QC. The merger preserves incomplete work as pending and gives human overrides priority. A pilot compares automatic decisions with review-team labels; the builder checks the comparison and current context/label hashes before generating out-of-pilot jobs. A recorded override reason permits exceptional generation, and generated jobs reject runtime scope replacement. These are local executable controls, not a guarantee of independent model reasoning, valid human labels or full-dataset recall. Cost approval remains conversational, and manual/edited alternatives do not inherit unchanged-builder protections. RT-RTE-screening-fanout, EPI-RTE-screening-merge, EPI-RTE-pilot-comparison.
Preserving decisions without replaying every turn
The Material Passport and named artifacts carry stage inputs. A separate run ledger retains selected user words, execution receipts and counters. Its helper validates the chain and named files under a cooperating-writer lock before atomic append. Recovery can use valid receipts and conservative counters rather than trust a lossy host summary. Hashes do not authenticate words, recover omitted events or detect restoration of an older valid ledger. Optional reset requests a fresh session and artifact reload; an in-session declaration cannot erase prior tokens, and a stored path is not content already present in the model window. RT-OBJ-run-ledger, RT-RTE-checkpoint-recovery.
Other histories have narrower consumers. Inquiry summaries automatically expose selected live branches and stale-artifact context at named checkpoints, but author adoption controls reliance and the ledger cannot satisfy integrity gates. Criteria bindings check context identity and receipts, not substantive criterion application. Process narratives have a human reader without an established automatic lesson-adoption path. Post-terminal activity history is expressly excluded from model, gate and checkpoint inputs. Persistence alone consequently establishes neither learning nor automatic policy adoption. MEM-RTE-inquiry-summary, MEM-OBJ-review-criteria-binding, MEM-OBJ-process-record, MEM-RTE-activity-history.
The strongest implemented trace-fed reuse contribution is the resolver cache: automatic reduction of lookup results produces durable decisions that later exact-key verification calls can consume instead of making fresh resolver calls. It is SQLite decision memory, not retained full papers or model reasoning. Version/type/age checks and explicit live bypass constrain reuse. This supports a wired trace-learning route in the write-and-later-consumer sense; the prescribed capture-to-recovery ledger chain supports a separate claimed route. Neither establishes improved future capacity or accuracy, and opaque host summaries remain unresolved. MEM-OBJ-verification-cache, MEM-RTE-cache-verification, RT-OBJ-run-ledger, profile verification.
Revising, checking and delivering
A review roadmap proposes changes; item-level author choices authorize them. The revision protocol gives the writer exact scope and requests a patch plus evidence chain. A rejected patch permits one re-emission under unchanged authority; scope changes return to the author. Re-review commits criteria before exposing revision evidence, judges evidence before revealing persuasive responses, and requires persisted artifacts and checker calls before surfacing a decision. Deferral returns questions to the author; contract mismatch aborts without a decision. These are inspected orchestration obligations, not an independent audit of patch/checker internals or proof of host context isolation. Inline fallback cannot make a reviewer forget earlier material. RT-RTE-revision-review, workflow annotations.
Finalization separates several checks that cannot certify one another:
- Citation lookup derives metadata classifications from resolver outcomes. A match can win despite another miss; manual exemptions can yield unresolvable; anchor-kind presence does not prove claim support. EPI-RTE-citation-lookup.
- Claim audit mediates injected retrieval and model judgment; its corpus argument does not itself implement retrieval. A separate fixed matrix computes refusal recommendations. Unsupported claims can refuse, while ambiguity, paywalls and tool failures can continue with advisories. Non-refusal is not scientific acceptance. EPI-RTE-claim-judgment, EPI-RTE-audit-disposition.
- User reading declarations provide withdrawable coverage eligibility, not demonstrated reading or comprehension. Rescission changes current resolution; the prescribed finalizer must rerun to reflect it in markers. Marking read does not clear source-support failures. MEM-RTE-read-attestation, EPI-RTE-read-withdrawal.
The formatter is instructed to require current evidence/policy stamps and reject terminal markers before removing provenance from clean conversions. Package checking is a further, separately configured policy: advisory by default, with bounded repair under strict conditions and freshness checks for reused reports. Direct shell conversion is not universally intercepted. Final files precede a Stage-5 acceptance checkpoint; optional process-record delivery requires terminal acknowledgement. Missing converters do not count as delivered PDFs. RT-RTE-final-output, workflow account.
Enforcement depends on route and host. A successfully evaluated Claude hook can deny covered structured writes and recognized Bucket A Bash, but main-session Bash, unsupported tools and launcher failure have pass-through paths. Pi does not receive that hook from the wrapper. Optional contained Codex verification requests read-only ephemeral execution and rejects invalid evidence shapes; generic cross-model routes do not inherit that envelope. Neither a valid receipt nor a second model establishes factual accuracy or independence. RT-RTE-tool-scope, RT-RTE-cross-verification.
Assessment by task and property
For a user-mediated manuscript scenario, ARS supplies a detailed allocation of author authority, role outputs, retained provenance and local refusal conditions. It does not establish an unattended, universally gated research process. For screening, it supplies a concrete pilot-to-generation admission path, but no measured recall on unseen records. For scientific warrant, checks concern reported consistency and selected support judgments, not authentic data or performed procedures; live review is explicitly NOT_CALIBRATED. No inspected acceptance-and-integration chain establishes a new scientific claim, and unassessed quantitative and integrity branches prevent a system-wide negative about knowledge production. RT-RTE-pipeline-stages, RT-RTE-screening-fanout, EPI-CLM-integrity-boundary, EPI-CLM-review-calibration, epistemic coverage.
The localized screening-criteria account supports theory-builder condition 1 at afforded status; conditions 2, 3 and 4 are separately claimed. Code supplies the criteria, while content-dependent consumption, claim-directed criticism with blame assignment, and retained criticism shaping another round are not demonstrated by an inspected case. Criteria can be addressed individually and persisted across rounds of one review. Whether criticism improved future action remains unestablished; pilot calibration is a positive capability, not evidence of that benefit. Reflection about ARS itself is unestablished, and autonomous theory building is not established because human labels, confirmations and amendments remain material. RT-RTE-screening-fanout.
Own-system improvement is a separate question. Evaluation tooling can compare selected measured tasks and condition admission of a maintainer-proposed change; it does not generate or install successors, and later causal use of changed ARS machinery is unestablished. Manuscript revision alone changes an external product. Thus exercised self-improvement at this boundary is not established, independently of the positive cache/ledger reuse routes and independently of reflection or autonomy. RT-RTE-evaluation-admission, RT-RTE-revision-review.
An inspectable end-to-end trace showing exact dispatches, user decisions, helper outcomes, rejected artifacts and final acknowledgement would change the operational assessment. Actual source-to-synthesis comparisons and recorded evidence-consuming acceptance would change the scientific-claim assessment. A screening case retaining the challenged criterion, failure attribution, amendment and next-round use would strengthen the theory account; a matched comparison would be needed for benefit attribution. For own-system improvement, evidence must connect objective-relevant evaluation to an installed behavior-determining change and subsequent operation causally dependent on it. These are specific missing links, not requests to infer quality from more stages or more agents. The cited production, screening and evaluation records delimit each comparison.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Source inspection without established execution or causal results | SRC-1; RT-RTE-pipeline-stages, RT-RTE-revision-review, RT-RTE-final-output | Shipped workflow instructions and selected executable paths | Successful delivery, reliable compliance, accuracy, reviewer independence and causal benefit | Inspectable runs with dispatch/input/output provenance; appropriately controlled outcome comparisons |
| Passport selector inventory omits two retained authority mechanisms | SRC-1; MEM-RTE-corpus-reuse, RT-RTE-pipeline-stages, RT-RTE-checkpoint-recovery; report-verification limit | Verified orchestrator doctrine retains confirmed standing constraints by applicable stage with withdrawal, and excludes integrity NOT_FOUND references from later writer inputs until user restoration; persistence implementation and operation are not audited | Complete passport-selection coverage or exhaustive user-constraint propagation and negative source-admission comparisons; existing named routes remain supported | Separate accepted records tracing retention, withdrawal/restoration, selection and writer consumption for both selectors |
| Profile units group independently operative mechanisms | SRC-1; MEM-OBJ-review-criteria-binding, MEM-OBJ-activity-store, EPI-OBJ-screening-decisions, RT-RTE-evaluation-admission; profile-verification limits | Partial local-file, fixed-rule-field, automatic-admission and screening/evaluation-authority classifications | Counts of independently classified objects/mechanisms and complete per-consumer comparisons; supported existence findings remain valid | Finer units with separately recorded object, consumer and admission boundaries |
| Optional claim-audit verdict reuse lacks individual write/read classification | SRC-1; EPI-RTE-claim-judgment; profile-verification limits | API permits cached verdict reuse sensitive to model/prompt/config; broader retention/production wiring remains qualified | Exhaustive reusable-decision-store inventory or complete audit-cache versus resolver-cache route/provenance comparison | Accepted records and classification of the audit cache's producer, persisted parts, provenance, invalidation and later consumers |
| Host context and provider behavior remain external or opaque | SRC-1; RT-RTE-command-dispatch, RT-RTE-checkpoint-recovery, RT-CMP-session-model, RT-CMP-cross-model, RT-RTE-tool-scope | Adapter interfaces, requested reset/dispatch controls and conditional hook logic | Universal isolation/write mediation, exact weight fixity, provider training conclusions or complete host-summary memory classification | Host/provider contracts and inspected implementations, served-model attestations and deployed control traces |
| Optional helper and semantic branches remain incompletely inspected | SRC-1; MEM-OBJ-review-criteria-binding, MEM-RTE-inquiry-summary, MEM-RTE-corpus-reuse, RT-RTE-revision-review, RT-RTE-final-output; epistemic coverage | Named criteria consumers, inquiry replay/publication, non-folder adapters, enrichment/retraction/audit paths, patch/re-review/package internals and quantitative synthesis | Uniform transaction, rollback, semantic-preservation or trust guarantees; system-complete negative about derivation or knowledge production | Route-specific source audits and supported records for consequential transformations and consumers |
| Input warrant, human control and acceptance remain local | SRC-1; MEM-RTE-cache-verification, MEM-RTE-read-attestation, MEM-OBJ-review-criteria-binding, MEM-RTE-activity-history, EPI-RTE-synthesis-content | Generic injected cache dictionaries, initial declaration control, attestation resolution, deletion tokens and synthesis doctrine | Uniform human curation/authentication, actual reading, source truth or accepted novel scientific claims | Provenance/admission evidence for each branch, authenticated decisions where required, actual source/output comparisons and recorded acceptance criteria/use/scope |
| Learning benefit and exercised own-system change are not established | SRC-1; MEM-RTE-cache-verification, RT-OBJ-run-ledger, RT-RTE-screening-fanout, RT-RTE-evaluation-admission | Wired resolver reuse, claimed recovery chain, prescribed criterion amendments and selected metric admission | Improved future capacity from criticism, autonomous/reflective theory building or exercised ARS self-improvement | Retained criticism-to-next-round evidence; objective-specific comparisons; installation and subsequent causal use of an evidence-responsive own-system update |
Source-read log: none. This synthesis uses the supplied accepted members and their record passages; it adds no source evidence.