ACE epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is how ACE turns task experience into content used on later tasks, and what its checks license. Scope and exclusions are those of the boundary; evidence identities and layers are in its Source register. The main repository supplies implementation and doctrine/design, plus attributed performance reports, not inspected runs or causal experiments. AppWorld is unavailable. No source, test, example or model call was executed.

The assessment distinguishes matching a dataset answer, diagnosing why it matched, proposing a reusable lesson, admitting text, and using that text. These are not one acceptance event. Public claims include RT-CLM-context-improvement and EPI-CLM-context-quality.

Coverage

Direct reads below supplement, rather than silently reproduce, the supplied runtime account. SRC-1 supplies all paths. Searches checked tracked Python entry points, function definitions and packaging against the runtime; they found no packaged CLI script registration in pyproject.toml. The shipped invocation surface is the three run scripts and public Python methods described by the runtime, not a provider tool registry. No system-complete negative is asserted.

Entry point or operation	Source path	Coverage and records / conclusion prevented
Sequential training, run modes, regeneration and online windows	ace/ace.py	Supplied findings: RT-RTE-sample-adaptation, RT-RTE-online-windows. Scheduling and target-derived feedback taken from these records, not independently reread here.
Batch training, grouped proposals, reducer, rollback	ace/ace_batch.py	Supplied findings: RT-RTE-batch-adaptation. Direct reducer prompt and curator parser reads; EPI-RTE-curator-admission.
Finance and both Mind2Web CLI callers, eval_only/test, reload	eval/finance/run.py, eval/mind2web/run.py, eval/mind2web2/run.py, utils.py	Entry-point search plus supplied RT-RTE-evaluate-select, RT-RTE-playbook-reload. No independent caller execution.
Generator/reflector context and bullet selection	ace/core/generator.py, ace/core/reflector.py, ace/prompts/reflector.py, playbook_utils.py	Generator supplied via RT-RTE-sample-adaptation; reflector and bullet extraction directly read. EPI-OBJ-reflection, EPI-RTE-counter-update.
Curator generation, schema, ADD consumer; proposal/reducer policy	ace/core/curator.py, ace/prompts/curator.py, playbook_utils.py	Direct sequential curate, parser/filter, prompt and deterministic application reads; proposal/reducer wiring additionally supplied by runtime. EPI-RTE-curator-admission.
Optional similarity grouping, merge, removal and failure fallback	ace/core/bulletpoint_analyzer.py	Direct merge/analyze read; embedding/grouping mechanics supplied by RT-RTE-bulletpoint-merge and RT-CMP-embedding-model.
Finance and both Mind2Web answer evaluators	eval/finance/data_processor.py, eval/mind2web/data_processor.py, eval/mind2web2/data_processor.py	Direct scoring reads: EPI-RTE-answer-match. Host-interpreter force also RT-RTE-finer-scoring.
Transport, logs, provider weights	llm.py, logger.py, utils.py	Supplied runtime findings. Logs do not feed later curation here. Remote model processing and actual weight identities uninspected; no internal learning conclusion.
Dataset preparation/provenance and extension documentation	eval/mind2web/prepare_data.py, eval/mind2web2/prepare_data.py, EXTENDING_ACE.md, tutorials/ExtendingDatasets.md	Available but not directly assessed. Prevents benchmark-construction validity and completeness of extension-contract conclusions; scoring conclusions are conditional on supplied standardized targets.
AppWorld integration	ace-appworld	Unavailable submodule, as boundary records; prevents full AppWorld feedback, admission and environment-effect conclusions.
Public claims and packaging	README.md, pyproject.toml	Direct relevant passages/package read; EPI-CLM-context-quality. README reports do not supply inspectable comparison evidence.

No supplied execution trace ties an actual candidate to generation, checking or acceptance. Thus implementation supports route capacity, not the semantic classification or acceptance history of a particular generated lesson. No supplied intervention separates reflector, curator, selection and other treatment effects. Missing provider internals prevent weight-update assessment, not assessment of the text-update path. Uninspected preparation remains a local limit, not evidence that targets are defective.

Epistemic-object inventory

Fields owned by a declaring record are see that record. No operative objects were supplied in runtime; the following split content from metadata and current answer from proposed explanation.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-answer	Dataset-relative answer propositions; action descriptions may instead be proposals	see record	SRC-1, eval/finance/data_processor.py	A matching answer does not establish correct reasoning or source truth.
EPI-OBJ-reflection	Candidate error diagnoses, root causes, recommended principles	see record	SRC-1, ace/prompts/reflector.py	Requested fields, not an observed diagnosis or established cause.
EPI-OBJ-playbook	Potential formulas, task regularities and strategies; also non-truth-apt instructions	see record	SRC-1, ace/prompts/curator.py	Actual generated content and source warrant not determinable.
EPI-OBJ-bullet-tags	Model utility judgments plus symbolic counts and identifiers	see record	SRC-1, playbook_utils.py	Counter truth is about processed tags, not independently measured usefulness.

Authority-route ledger

The same container record can support separate functions below; checking, disposition, retention and integration remain separate rows. All rows describe implemented possibilities, not observed events. see defers generic fields to their records. No route below licenses general truth or cross-task transfer merely from successful consumption.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-playbook-reload	content transformation	implemented	EPI-OBJ-playbook	truth-apt transformation: acquisition/import	Caller-supplied text enters playbook	Existing path and truthy content, not truth check	Explicit run initialization	Full text imported	Enforcing delivery	Source warrant unknown; no authentication or endorsement	Initializes context without curator gate	see record and RT-BAP-playbook-guidance	SRC-1, eval/finance/run.py	none	none	Imported content is acquired, not produced by ACE.
RT-RTE-sample-adaptation	content transformation	implemented	EPI-OBJ-answer	truth-apt transformation: indeterminate	Task/context/playbook to answer	Generator model; requested task reasoning	Each generation; see record	Candidate answer/trace and reported IDs	Candidate supply	No warrant before task check	Makes answer available for scoring	Processor consumes extracted answer; enforcing branch input; current sample	SRC-1, ace/core/generator.py	RT-CLM-context-improvement	none	Could copy, derive or conjecture; actual output needed to distinguish.
EPI-RTE-answer-match	check/evidence production	implemented	EPI-OBJ-answer	no content change	Extracted answer versus dataset target	Task processor, restricted matching domain; see record	Training/evaluation scoring	Boolean or aggregate metric	Produces operative comparison	Dataset-relative match only, conditional on target and parser	Feeds next row, not a direct playbook truth veto	Python branch/metric consumer, returned value, enforcing, current round/snapshot	SRC-1, eval/finance/data_processor.py	none	none	FINER truncation, Mind2Web missing target value, skipped evaluation errors limit license.
RT-RTE-sample-adaptation	disposition/acceptance	implemented	EPI-OBJ-answer	no content change	Accept answer for stopping or retry	Processor Boolean; round limit bounds continuation	After scoring	Stop regeneration or request another round	Enforcing	Acceptance for matching this sample target, not reasoning validation	Correct/incorrect feedback and retry selection	Orchestrator, correctness flag, enforcing, sample	SRC-1, ace/ace.py	RT-CLM-context-improvement	Label-free reading contradicted	no_ground_truth still uses oracle; stopping by round limit is not acceptance.
RT-RTE-sample-adaptation	content transformation	implemented	EPI-OBJ-reflection	truth-apt transformation: indeterminate	Trace/feedback/selected bullets to diagnosis	Reflector model, task-local diagnosis	Correct result once, incorrect result per permitted round	Reflection and tags	Advisory interpretation plus counter input	Candidate explanation only	Reflection reaches regenerator/curator	Prompt slots, model interpretation, advisory, current and update rounds	SRC-1, ace/core/reflector.py	EPI-CLM-context-quality	none	Root-cause claim is not entailed by answer mismatch; no actual diagnosis supplied to establish non-entailment of a particular output.
EPI-RTE-counter-update	behavior/policy adaptation	implemented	EPI-OBJ-bullet-tags	non-truth-apt policy/content update: model tags change displayed utility counters	Tagged bullet utility	Reflector judgment; code maps recognized tags to increments	After reflection, before later prompts	helpful/harmful increment or unchanged count	Enforcing metadata write; later advisory use	Counts document accepted tag events, not causal bullet benefit	Changes metadata supplied to future model judgments	Generator/curator prompt text, advisory, later samples and reload	SRC-1, playbook_utils.py	EPI-CLM-context-quality	none	Harmful count is not withdrawal; duplicate IDs/tags are not independent causal trials.
RT-RTE-sample-adaptation, RT-RTE-batch-adaptation	content transformation	implemented	EPI-OBJ-playbook	truth-apt transformation: indeterminate	Reflections/context to proposed lesson; proposals to reducer output	Curator model; reducer novelty/usefulness policy	Scheduled sequential curation or batch group/reduction	ADD proposals or empty operations	Model selection then parser	Proposed guidance, not accepted explanation	Supplies prospective instructions	Curator/reducer prompt, advisory policy then operation output, update stage	SRC-1, ace/prompts/curator.py	EPI-CLM-context-quality	none	Reducer can drop proposals but does not receive complete original traces; no observed semantic preservation/novelty.
EPI-RTE-curator-admission	check/evidence production	implemented	EPI-OBJ-playbook	no content change	JSON outer structure and ADD keys	Python parser; syntax/key-presence domain	After curator response	Structure accepted or exception	Enforcing parser gate	Structural compatibility only; not truth, relevance or transfer	Allows application attempt or failure fallback	Curate/proposer/reducer caller, parsed dict/exception, enforcing, current update	SRC-1, ace/core/curator.py	EPI-CLM-context-quality	none	content/section types not fully checked; unsupported names can pass sequential validation.
EPI-RTE-curator-admission	disposition/acceptance	implemented	EPI-OBJ-playbook	no content change	Admit supported ADD; reject malformed response or empty result	Python application plus upstream model choice	After schema gate	Applied ADD or prior playbook	Enforcing text admission	Operational admission, not evidence-consuming acceptance of lesson truth	New text can govern next generation without a per-lesson accuracy veto	Operation interpreter, ADD list, enforcing write, subsequent calls	SRC-1, playbook_utils.py	EPI-CLM-context-quality	No semantic acceptance established	No human per-addition approval; failure fallback subject to logging success.
EPI-RTE-curator-admission	content transformation	implemented	EPI-OBJ-playbook	truth-apt transformation: non-ampliative reshaping	Fixed single-line content placed under section with fresh ID	Deterministic application for well-formed ADD	Application stage	Content retained verbatim with zero counters and section placement	Enforcing formatting	No extra warrant from reshaping	Addressable bullet enters playbook	String construction to generator/curator, advisory downstream, later tasks	SRC-1, playbook_utils.py	none	none	Bounded deduction in record; arbitrary multiline/non-string content excluded from this preservation conclusion.
RT-RTE-playbook-reload	retention	implemented	EPI-OBJ-playbook, EPI-OBJ-bullet-tags	no content change	Runtime playbook to files and later explicit reload	Save schedule/path choice	In-run snapshots/end; later invocation	Retained text	Enforcing write/read	Retention is not endorsement	Makes prior guidance available across runs	see record	SRC-1, ace/ace.py	none	none	No complete provenance ledger or restored ID history.
RT-RTE-evaluate-select	disposition/acceptance	implemented	EPI-OBJ-playbook	no content change	Whole snapshot for final offline testing	Strict validation-accuracy improvement; successful scored subset	Periodic offline validation	best_playbook replaced or kept	Enforcing selection	Empirical selection for this validation criterion, not per-claim acceptance or causal explanation	Chooses final-test context; current training state not rolled back	Python snapshot selector, metric, ranking then permissive, offline run	SRC-1, ace/ace.py	EPI-CLM-context-quality	none	Exceptions skipped; no observed comparison or transfer test here.
RT-RTE-bulletpoint-merge	check/evidence production	implemented	EPI-OBJ-playbook	no content change	Candidate redundancy groups	Embedding dot-product threshold, not logical equivalence	Optional analyzer after curation	Similarity groups	Ranking/grouping	Numerical similarity only	Selects groups for rewrite/removal	Analyzer group indices, threshold, permissive, current maintenance	SRC-1, ace/core/bulletpoint_analyzer.py	EPI-CLM-context-quality	none	High similarity does not establish same truth conditions.
RT-RTE-bulletpoint-merge	content transformation	implemented	EPI-OBJ-playbook	truth-apt transformation: indeterminate	Similar group to replacement or first-member-only result	Merge model plus regex; merge=False keeps first	After grouping	Group replaced, other members removed	Enforcing reconstruction	No checked meaning-preservation license	Can withdraw guidance, including on merge failure	Analyzer to orchestrator playbook, enforcing write then advisory use, later calls	SRC-1, ace/core/bulletpoint_analyzer.py	EPI-CLM-context-quality	Preservation claim limited	Loss may be real distinct content, not proven duplicate removal. No analyzer-specific undo.
RT-BAP-playbook-guidance	operational admission/selection/consumption	implemented	EPI-OBJ-playbook	no content change	Retained text delivered for task solving and further additions	Model interpretation; no enforced per-bullet obedience	Regeneration, later samples/windows and explicit reload	Full playbook delivered	Advisory content, enforcing delivery	Operational reliance allowed without lesson truth acceptance	Can influence answers and new guidance	see record	SRC-1, ace/prompts/generator.py	RT-CLM-context-improvement	none	Delivery is not observed dependence or benefit.
RT-RTE-sample-adaptation	lifecycle integration	no route found within boundary	EPI-OBJ-playbook	no content change	Accepted individual truth claim connected to evidence/use	No such phase established by inspected curate/apply or supplied loop	No established activation	Operational retention exists instead	No established epistemic integration force	No individual accepted ampliative claim established	Do not confuse use before acceptance with integration	No such consumer established within these inspected routes	SRC-1, ace/core/curator.py	none	none	Bounded to main-repository routes above; AppWorld and actual candidate histories unresolved.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
RT-CLM-context-improvement	see record	SRC-1, README.md; doctrine/design and reported operation	Three model roles, retained context, no-GT prompts	RT-RTE-sample-adaptation, RT-RTE-batch-adaptation, RT-BAP-playbook-guidance	None supplied	No intervention; AppWorld absent	Improvement-directed context-change capacity is wired	Label-free interpretation does not hold for inspected target-scored drivers. Exercised self-improvement and learning benefit unestablished.
EPI-CLM-context-quality	see record	SRC-1, README.md; doctrine/design and reported operation	Reflection asks for diagnoses; curator asks for new useful content; reducer asks to discard low-value proposals	EPI-RTE-answer-match, EPI-RTE-curator-admission, RT-RTE-bulletpoint-merge, RT-RTE-evaluate-select	None supplied; README gains are attributed reports	No inspectable comparison design or independent component variation; similarity is not equivalence	Incremental ADD formatting and optional redundancy-management machinery exist; snapshot ranking uses scored outcomes	Preservation is consumer-local; merging is model-mediated and can delete on fallback. Quality, latency and accuracy gains not verified here.

Bounded conclusion

ACE can import text of unknown warrant, generate answers against it, compare extracted answers with dataset targets, ask a model to diagnose errors, and let another model propose retained guidance. It stores natural-language playbook content with symbolic identifiers/counters, not updated provider weights. The separate remote and embedding components' mutability limits are in RT-CMP-generator, RT-CMP-reflector, RT-CMP-curator and RT-CMP-embedding-model.

Answer matching has consequential but narrow authority: it controls retry/stop decisions and validation-snapshot ranking. FINER field matching, formula numeric/string matching and Mind2Web action-description matching do not validate the producing reasoning, causal diagnosis or transferable strategy. The FINER interpreter path adds operational host effects before comparison, not extra warrant. Successful outcomes therefore cannot establish reflector explanations or individual bullet utility (EPI-RTE-answer-match, RT-RTE-finer-scoring).

The deterministic ADD wrapper is non-ampliative for the inspected well-formed single-line case. No warranted-premise derivation of new task knowledge is established. Model-generated answers, diagnoses, proposals and merges remain semantically indeterminate without concrete outputs; some might copy, derive, conjecture or merely instruct. The architecture requests generalization, but that is not evidence of a produced accepted ampliative claim. Schema admission licenses application attempts, not truth. Counter increments represent model judgments; harmful tags do not automatically revoke guidance (EPI-RTE-curator-admission, EPI-RTE-counter-update).

The main loop implements operational retention/use before any established per-lesson evidence-consuming acceptance. Validation chooses a whole playbook for final testing under a named accuracy criterion; it neither accepts each lesson nor rolls back current training state. Similarity maintenance can remove content, including when merging fails, without truth-directed withdrawal or a preservation check. Explicit reload permits cross-run reuse without provenance authentication (RT-RTE-evaluate-select, RT-RTE-bulletpoint-merge, RT-RTE-playbook-reload).

The behavioral-authority path is full playbook text in generator and curator prompts: delivery is enforced by code, interpretation is advisory, and the horizon spans immediate regeneration, later samples/windows and explicit cross-run reload. This wires an improvement-directed self-change pathway relative to the independent answer-accuracy objective; subsequent behavior causally depending on a retained change is not evidenced. Thus exercised self-improvement and successful improvement remain separately unestablished. Reflection has an implemented feedback channel, but both causal directions of actual reflective behavior remain unobserved. Computation performs the bounded driver's internal proposal/decision/application roles; no actual localized theory, content-dependent theory consumption or claim/result/blame-directed criticism was supplied. The theory-builder conditions and learning limits remain separate as assessed in RT-RTE-sample-adaptation, not a combined negative verdict.

Shared records

Operative objects

EPI-OBJ-answer

Label: Extracted answer

Source-native identity: generator answer extracted for DataProcessor.answer_is_correct; natural-language or numeric answer strings in Python memory, with task-specific symbolic parsing. Dataset targets are external answer references, not ACE-produced facts. Closest supplied RT-RTE-sample-adaptation and RT-RTE-finer-scoring are routes, not an object counterpart; this object spans their input/output interfaces. SRC-1, eval/finance/data_processor.py, eval/mind2web/data_processor.py, eval/mind2web2/data_processor.py.

if self.task_name == "finer": return self._finer_answer_is_correct(predicted, ground_truth) --- eval/finance/data_processor.py

EPI-OBJ-reflection

Label: Reflection response

Source-native identity: reflection_content returned as reflector response text; natural-language diagnosis and proposed principles mixed with JSON tags, held in Python strings and model-call logs. Closest supplied RT-RTE-sample-adaptation is the encompassing process, not this object; batch also produces reflections. Requested fields do not establish an actual truth-apt theory or a correct diagnosis. SRC-1, ace/core/reflector.py, ace/prompts/reflector.py.

• error_identification: what specifically went wrong in the reasoning?
• root_cause_analysis: why did this error occur? What concept was misunderstood?
• correct_approach: what should the model have done instead?
• key_insight: what strategy, formula, or principle should be remembered to avoid this error? --- ace/prompts/reflector.py:20-23

EPI-OBJ-playbook

Label: Playbook content

Source-native identity: current_playbook / self.playbook, specifically bullet content and section organization, excluding the separately assessed tags/counters. Natural language in Python strings and saved text files is interpreted by generator and curator models. It may express truth-apt formulas/regularities or prescriptions; no concrete produced instance is supplied. Closest supplied RT-RTE-playbook-reload and RT-BAP-playbook-guidance describe transport and consumers rather than declare this object. SRC-1, ace/prompts/curator.py, playbook_utils.py.

• The playbook you created will be used to help answering similar questions. --- ace/prompts/curator.py:9-9

EPI-OBJ-bullet-tags

Label: Bullet tags and counters

Source-native identity: reflector bullet_tags, stored line helpful/harmful counters and bullet IDs. JSON/list values and integer fields have symbolic update rules; model utility judgments are semantically interpreted before their encoding. Substrates are Python values, playbook strings and saved text. Distinct from EPI-OBJ-playbook: counters can change while content remains fixed. Closest supplied RT-RTE-sample-adaptation spans their production/consumption, not an object counterpart. SRC-1, playbook_utils.py.

return f"[{bullet_id}] helpful={helpful} harmful={harmful} :: {content}" --- playbook_utils.py

Routes

EPI-RTE-answer-match

Label: Dataset answer matching

Endpoints/progression/owner: task processor receives extracted prediction and dataset target, parses according to task and returns correctness or aggregate accuracy. Computational comparison owns the decision; human/data authors supply reference labels, not per-prediction approval. It overlaps the check portions of RT-RTE-sample-adaptation, RT-RTE-evaluate-select and RT-RTE-finer-scoring; it is a cross-route scoring part, not a replacement for their orchestration or host-effect findings. Mode is bounded labelled training/evaluation. SRC-1 implementation: eval/finance/data_processor.py, eval/mind2web/data_processor.py, eval/mind2web2/data_processor.py.

• Immediate return: Boolean correctness or task-level aggregate accuracy; FINER can return matched/total field counts.
• Later read-back: inapplicable — this scorer does not retain updated content for later invocations; snapshot selection and retained playbooks are separate routes.
• Delegated visibility: Full extracted answer and target strings reach the host processor; it does not see or validate the full generating reasoning.
• Selection predicate: FINER task selects lowercased comma-fragment matching; formula selects float equality with string fallback; Mind2Web selects parsed element/operation/value matching.
• Invalidation or expiry: inapplicable — each comparison evaluates current inputs, not a retained cache.
• Activation or effect: Returned correctness controls retries and contributes to validation ranking through supplied runtime consumers; actual effects unobserved.
• Evidence limits: Targets are assumed references, not independently validated facts. FINER can execute prediction expressions after successful target eval; aggregate evaluation can skip errors per supplied runtime. No target execution performed.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Concrete deterministic trace, by code inspection: FINER target 1, prediction 1,999 produces labels ["1"] and predictions ["1","999"]; length normalization truncates the latter to ["1"]. Both retained fragments evaluate to integer 1, count is 1 and score is 1, so the checker returns true. This is a deduction about the implementation, not a run or an invented observed candidate. It refutes an exact-whole-answer interpretation while preserving the narrower field-match license. Formula float equality also checks values, not derivations. In both Mind2Web processors a matching CLICK needs only element and operation, and a TYPE/SELECT with no parsed target value can pass without value verification. Neither checks live browser success.

if len(pred) != len(label): if len(pred) > len(label): pred = pred[:len(label)] --- eval/finance/data_processor.py

score = count / len(pred) if pred else 0 if return_counts: return count, len(pred) return score == 1 --- eval/finance/data_processor.py

if truth["value"] is None: return True # No value to check --- eval/mind2web/data_processor.py

if truth["value"] is None: return True # No value to check --- eval/mind2web2/data_processor.py

EPI-RTE-counter-update

Label: Counter update

Endpoints/progression/owner: reflector response supplies tags; update_bullet_counts creates an ID lookup and increments matching parsed lines. Python owns the write; reflector supplies utility judgments. This spans counter-update portions of RT-RTE-sample-adaptation and RT-RTE-batch-adaptation, not a new adaptation container. The target is bullet usefulness for the current task according to model interpretation, not independently isolated causal contribution. SRC-1 implementation, ace/core/reflector.py, playbook_utils.py; prompt policy, ace/prompts/reflector.py.

• Immediate return: Playbook text with matching helpful/harmful counters incremented, or original text if no valid tag map.
• Later read-back: Updated counters travel in full playbook prompts to generator/curator and may persist through files and explicit reload.
• Delegated visibility: Reflector receives the response trace and only parsable bullets selected by generator-reported IDs; this does not prove actual bullet use. Future consumers see counters without independently validated utility evidence.
• Selection predicate: Dict tag entries with id or bullet key; matching parsed IDs; only helpful and harmful values increment. Last tag for an ID wins within the lookup.
• Invalidation or expiry: Neutral leaves counts unchanged; harmful increments do not remove content or automatically expire it.
• Activation or effect: Metadata is rewritten and delivered for possible model interpretation; causal behavioral dependence unobserved.
• Evidence limits: Code counts tags, not successful interventions on bullet presence. Malformed/unmatched tags may not change counters. No independent causal utility oracle or per-content human approval is wired here.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Admission is symbolic tag recognition, with no content-specific human veto. The guidance asks helpful/harmful/neutral judgments. There is no truth-directed rollback in this utility; later caller replacement/analyzer rewriting are distinct mechanisms. Content persists unchanged by this utility.

if tag == 'helpful': parsed['helpful'] += 1 elif tag == 'harmful': parsed['harmful'] += 1 --- playbook_utils.py

EPI-RTE-curator-admission

Label: Curator structural admission

Endpoints/progression/owner: curator response goes through _extract_and_validate_operations, then deterministic apply_curator_operations; batch additionally filters proposer/reducer operations to ADD. This cross-cutting parser/application part overlaps RT-RTE-sample-adaptation and RT-RTE-batch-adaptation, but does not replace either. Models propose/select content; Python checks outer structure and writes supported additions. No per-content human approval or dataset answer oracle evaluates each proposed lesson. Context is scheduled training curation, not a universal knowledge-acceptance gate. SRC-1 implementation, ace/core/curator.py, playbook_utils.py; proposal guidance, ace/prompts/curator.py.

• Immediate return: Parsed operations, then updated playbook/next ID; handled parse/application errors or empty-response sentinel return previous text and empty operations where failure logging succeeds.
• Later read-back: Admitted text enters post-curator generation, later samples/windows, future curator inputs and persisted snapshots/reload.
• Delegated visibility: Sequential curator gets current playbook, recent reflection, question context and stats; batch reducer gets base playbook and operation groups rather than complete raw traces.
• Selection predicate: Reasoning must be a string, operations a list, each operation a dict with type, and ADD must contain type/section/content. Batch drops non-ADD names; deterministic application implements ADD only.
• Invalidation or expiry: This application consumer preserves existing lines; no implemented DELETE/UPDATE here. Failure fallback and the separate analyzer/reload routes qualify whole-system preservation.
• Activation or effect: Supported additions are written without per-addition accuracy veto and delivered downstream; actual interpretation and gain unobserved.
• Evidence limits: Structural checks do not validate truth, semantic novelty or full content types. No actual proposed/accepted lesson supplied. Batch exception-recovery limitations remain as supplied by runtime.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Guidance asks new, nonredundant, concise strategies; batch reducer may drop already-covered/vague/low-value proposals. That is model-mediated disposition, not checked semantic equivalence or evidence-consuming acceptance against a proposition-specific truth criterion. Empty operations can reject all proposals; no score threshold vetoes every admitted lesson. Curator reasoning is logged, but only admitted bullet text/counters form the reconstructed playbook.

Concrete deterministic reshaping deduction: with initial text ## OTHERS, next ID 1, and one well-formed ADD to others whose single-line content is Revenue equals price times quantity., application normalizes the section, obtains a section slug, allocates a fresh suffixed ID, initializes both counters to zero, and places exactly that content after the header. The wrapper adds no proposition about revenue. It preserves the given proposition without checking it. This case does not prove that a model-produced proposal was an entailed derivation, nor that arbitrary multiline content is safely represented.

required_fields = {"type", "section", "content"} missing_fields = required_fields - set(op.keys()) if missing_fields: raise ValueError(f"ADD operation {i} missing fields: {list(missing_fields)}") --- ace/core/curator.py

content = op.get('content', '')

        new_line = format_playbook_line(new_id, 0, 0, content)
        bullets_to_add.append((section, new_line))

--- playbook_utils.py

• Drop proposals that are already covered by the current playbook.
• Drop vague, redundant, or low-value proposals. --- ace/prompts/curator.py

Claims

EPI-CLM-context-quality

Label: Context quality and preservation claims

Source-native claims: README says incremental updates preserve prior knowledge, reflection improves context quality/downstream performance, curator uses deterministic merging/de-duplication/pruning, and reports performance gains. Closest supplied RT-CLM-context-improvement concerns evolving contexts and label-free self-supervision; this record separately addresses preservation, quality mechanism and quantitative outcome claims, not that record's identity. SRC-1, README.md, doctrine/design and reported operation, not observed or causal evidence in this run.

• 📈 Incremental Delta Updates: Localized edits that preserve prior knowledge while accumulating new insights --- README.md

• Reflector separates evaluation and insight extraction from curation, improving context quality

• Curator converts lessons into structured delta updates with helpful/harmful counters, using deterministic merging with de-duplication and pruning --- README.md

ACE consistently outperforms strong baselines, achieving average gains of +10.6% on agent tasks and +8.6% on domain-specific benchmarks, across both offline and online adaptation settings. --- README.md

The ADD consumer preserves old lines; the optional analyzer can remove them and its replacement is model-generated. Thus the deterministic description does not establish deterministic semantic preservation across all paths. Benchmark and efficiency reports remain attributed, without supplied experimental design/trace sufficient for independent causal attribution.