LHTB epistemic report

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

Source-and-claim boundary

The question is what LHTB checks and licenses for reliance, rather than whether a terminal agent can generate plausible explanations. Scope, exclusions, external-service limits and evidence identity are see boundary and its Source register. All added reads below are SRC-1 implementation or explicitly labelled doctrine/design, at the frozen commit. Nothing here is target execution. No independently inspected trajectories or intervention results establish actual repair, summary fidelity, generalization or component benefit.

The runtime baseline is provisional. Direct CLI inspection found exposed check and analyze operations omitted from that baseline. They produce assessments of task quality and reward legitimacy, so a negative about all assessment routes based only on the solver loop would be invalid. Task-local public feedback also differs materially from the hidden binary continuation channel. These are additions, not supersessions of the supplied runtime records.

Entry point or operation	Source path	Coverage and limit
Job/trial invocation, resume, multi-step, verifier disable	harbor/src/harbor/cli/main.py; harbor/src/harbor/trial/trial.py	CLI wiring directly read; progression see RT-RTE-trial-execution, RT-RTE-job-recovery.
Terminal commands, installed agents, dynamic imports, MCP/skills	harbor/src/harbor/agents/terminus_2/terminus_2.py; harbor/src/harbor/agents/factory.py	Supplied RT-RTE-terminal-actions, RT-RTE-agent-dispatch. Other installed-agent/provider internals uninspected: no system-complete absence of additional evaluators.
Filtering, summarization, overflow recovery	harbor/src/harbor/agents/terminus_2/terminus_2.py	Direct filter and handoff-install reads; rest see RT-RTE-context-handoff.
Interim/final checks, cleanup, rejection, separate artifacts, timed checkpoints	harbor/src/harbor/trial/trial.py; harbor/src/harbor/verifier/verifier.py	Supplied RT-RTE-verifier-execution, RT-RTE-verifier-continuation, RT-CLM-grader-isolation. No deployed isolation evidence.
Public feedback, hidden judge, dense scoring	tasks-2.0/acoustic_beamforming_peak__01/environment/app/tools/feedback; tasks-2.0/acoustic_beamforming_peak__01/tests/private/judge.py; tasks-2.0/acoustic_beamforming_peak__01/tests/private/dense_reward.py; tasks-2.0/acoustic_beamforming_peak__01/tests/test.sh	Direct reads: EPI-RTE-public-feedback, EPI-RTE-hidden-semantics, EPI-RTE-dense-reward. One task is not an audit of all 77 oracles.
Task quality check, trial analysis, job aggregation	harbor/src/harbor/cli/analyze.py; harbor/src/harbor/analyze/	Direct caller/backend/checker/analyzer and default rubric reads: EPI-RTE-task-check, EPI-RTE-trial-analysis, EPI-RTE-analysis-cache, EPI-RTE-job-aggregation. Custom prompts and actual SDK behavior unobserved.
Old task check/debug and trial summarize	harbor/src/harbor/cli/tasks.py; harbor/src/harbor/cli/trials.py	Direct dispatch reads: these print removal notices and exit, rather than adding independent evaluators.
END hooks, metrics and saved results	harbor/src/harbor/job.py	Direct completion-hook read: reward updates stats and metrics and writes job result. Recovery see RT-RTE-job-recovery. Configured metric implementations, custom hooks, sweeps, viewer/upload consumers uninspected: no exhaustive ranking or adaptation conclusion.
Patch-maintenance and alternate installed runtime	harbor/skills/apply-lhtb-patches/PATCH.md; harbor/patches/single_step.py.harbor-0.20.0	Supplied RT-RTE-runtime-patching: source-supported alternative policy, not an observed deployment.

Consequential public claims are EPI-CLM-benchmark-warrant and RT-CLM-grader-isolation. Missing images, opaque provider weights/state and unavailable historical runtime identity prevent deployment assurance. Uninspected remaining graders prevent a uniform claim about every task's criterion. Missing comparison evidence prevents attributing learning or scientific validity to a score.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-task-instruction	Task facts and requirements; imperatives are not themselves truth claims	see record	SRC-1, harbor/src/harbor/agents/terminus_2/terminus_2.py	Imported authored guidance, not verified world knowledge.
EPI-OBJ-reconstruct	Candidate scenario-output mapping and numerical summaries	see record	SRC-1, tasks-2.0/acoustic_beamforming_peak__01/tests/private/judge.py	Implementation alone does not individuate an agent's proposed explanation or actual output.
EPI-OBJ-check-report	Public match results, hidden semantic checks and reward	see record	SRC-1, tasks-2.0/acoustic_beamforming_peak__01/tests/private/dense_reward.py	Program-reference-relative, not a scientific validity certificate.
RT-RTE-context-handoff	Model summary/answers about prior work	see record	SRC-1, harbor/src/harbor/agents/terminus_2/terminus_2.py	Route citation owns handoff findings; semantic preservation unknown.
EPI-OBJ-analysis	Task quality, trial legitimacy, failure explanations, job narrative	see record	SRC-1, harbor/src/harbor/analyze/analyzer.py	Candidate model judgments; schema compliance does not establish truth.
RT-OBJ-trial-result	Recorded rewards, errors and timings	see record	SRC-1, harbor/src/harbor/job.py	Retention and scheduler reliance are not acceptance of explanations.
RT-CMP-terminus-model	Distributed-parametric content not locally individuated	see record	SRC-1, harbor/src/harbor/llms/chat.py	Exact weights, internal reasoning and provider-side updates uninspected.
EPI-CMP-analysis-model	Distributed-parametric content not locally individuated	see record	SRC-1, harbor/src/harbor/analyze/backend.py	Separate model selection and SDK; do not inherit Terminus identity.

Authority-route ledger

Each row separates a consequential function; repeated record citations refer to the same implementation, not new records. All implemented routes remain unobserved in supplied run evidence.

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-terminal-actions	content transformation	implemented	see record	truth-apt transformation: indeterminate	Task-solving response	Model; arbitrary terminal tasks	see record	Analysis, commands, completion intent possible	see record	Candidate generation, no truth license from parser	see record	see RT-BAP-task-guidance	SRC-1, harbor/src/harbor/agents/terminus_2/terminus_2.py	none	none	No concrete generated claim proves entailment or ampliation.
RT-RTE-context-handoff	content transformation	implemented	see record	truth-apt transformation: indeterminate	Prior conversation to summary/answers	Model; historical task state	see record	Replacement context	see record	Fidelity unverified; intended compression is not proven non-ampliative reshaping	see record	see RT-BAP-task-guidance	SRC-1, harbor/src/harbor/agents/terminus_2/terminus_2.py	none	none	Omission or invented inference cannot be ruled out.
RT-RTE-context-handoff	operational admission/selection/consumption	implemented	see record	no content change	Install handoff as next context	Code; successful model return, not semantic test	see record	Automatic context replacement	see record	No evidence-consuming acceptance	Later solver receives selected history by value	see RT-BAP-task-guidance	SRC-1, harbor/src/harbor/agents/terminus_2/terminus_2.py	none	none	No raw-history reload established.
EPI-RTE-public-feedback	check/evidence production	implemented	see EPI-OBJ-reconstruct	truth-apt transformation: entailed derivation	Public output/signature match	see record	see record	Public checks and diffs	Advisory	Exact predicate over supplied reference; no hidden-scenario license	Feedback can guide shell-driven repair, not stop trial	Solver via terminal stdout; advisory; current trial	SRC-1, tasks-2.0/acoustic_beamforming_peak__01/environment/app/tools/feedback	none	none	Signature alternative can pass despite unequal payloads.
EPI-RTE-hidden-semantics	check/evidence production	implemented	see EPI-OBJ-reconstruct	truth-apt transformation: entailed derivation	Hidden numerical outputs and artifact gates	see record	see record	Pass/fail report	Input to score, not direct command veto	Toleranced agreement on finite hidden scenarios only	Grader report reaches score converter	Judge JSON to dense grader; enforcing scoring input; one verification	SRC-1, tasks-2.0/acoustic_beamforming_peak__01/tests/private/judge.py	EPI-CLM-benchmark-warrant	Scope narrower than real-tool reconstruction	Oracle correctness and source encoding not certified.
EPI-RTE-dense-reward	content transformation	implemented	see EPI-OBJ-check-report	truth-apt transformation: entailed derivation	Judge report to numeric score	see record	see record	Graded reward or binary fallback	Scoring	Ladder-defined arithmetic, not proportion of all requirements met	Reward becomes continuation/metric input	Reward file to Verifier and trial; enforcing continuation input; trial	SRC-1, tasks-2.0/acoustic_beamforming_peak__01/tests/private/dense_reward.py	none	none	Dense partial credit differs from acceptance.
RT-RTE-verifier-continuation	disposition/acceptance	implemented	see RT-OBJ-trial-result	no content change	Whether numeric full pass ends continuation	see record	see record	Full pass or rejection/another phase	see record	Acceptance of tested task result for benchmark stopping, not a general explanatory claim	see record	Trial controller via reward; enforcing; remaining budget	SRC-1, harbor/src/harbor/trial/trial.py	EPI-CLM-benchmark-warrant	Binary default not universal	Timeout or disabled grading is not acceptance.
RT-RTE-verifier-continuation	behavior/policy adaptation	implemented	see record	non-truth-apt policy/content update: another repair phase	Continue task after rejection	see record	see record	Another model turn/run	see record	Rejection identifies no blamed claim	Current workspace can be revised; not rolled back	see RT-BAP-task-guidance	SRC-1, harbor/src/harbor/trial/trial.py	none	none	Adapted continuation does not demonstrate learning.
EPI-RTE-task-check	check/evidence production	implemented	see EPI-OBJ-analysis	truth-apt transformation: indeterminate	Task quality under rubric	see record	see record	Model check outcomes/explanations	Advisory	Candidate rubric assessment, not certified quality	Printed or optionally saved, not a launch veto	CLI/operator via table/JSON; advisory; later human use unobserved	SRC-1, harbor/src/harbor/analyze/checker.py	none	none	No outcome-fidelity check shown.
EPI-RTE-trial-analysis	check/evidence production	implemented	see EPI-OBJ-analysis	truth-apt transformation: indeterminate	Reward legitimacy and task specification	see record	see record	Trial summary/checks	Advisory	Candidate explanation against available files	Can inform operator; does not withdraw trial reward	CLI and aggregation via structured assessment; advisory; post-run	SRC-1, harbor/src/harbor/analyze/analyzer.py	none	none	Possible unsupported inference from missing task files.
EPI-RTE-analysis-cache	operational admission/selection/consumption	implemented	see EPI-OBJ-analysis	no content change	Structured output admissibility	Pydantic schema, not truth evaluator	see record	Reject invalid structure; admit valid result	Enforcing structure	No epistemic acceptance of explanation	Valid results can be printed/persisted/aggregated	Analyzer via model validation; enforcing; assessment invocation	SRC-1, harbor/src/harbor/analyze/analyzer.py	none	none	A valid conclusion can be false.
EPI-RTE-analysis-cache	retention	implemented	see EPI-OBJ-analysis	no content change	Save and reuse analysis JSON	see record	see record	Cached assessment returned	Permissive reuse	No new evidence check on cache hit	Reanalysis skipped unless overwrite	Later Analyzer via file/cache flag; enforcing reuse; subsequent invocations	SRC-1, harbor/src/harbor/analyze/analyzer.py	none	none	Existence is not freshness or endorsement.
EPI-RTE-job-aggregation	content transformation	implemented	see EPI-OBJ-analysis	truth-apt transformation: indeterminate	Per-trial assessments to job narrative	see record	see record	Saved job summary	Advisory	Inherits unverified inputs; no new warrant from aggregation	Narrative reaches CLI/operator and cache	Summary text/JSON; advisory; later analysis use	SRC-1, harbor/src/harbor/analyze/analyzer.py	none	none	No independent source recheck in aggregation call.
RT-RTE-job-recovery	retention	implemented	see RT-OBJ-trial-result	no content change	Completed results retained/reloaded	see record	see record	Completed trial skipped, stats reused	see record	Saved outcome metadata, not accepted theory	see record	Scheduler via saved files; enforcing; job restart	SRC-1, harbor/src/harbor/job.py	none	none	Does not restore solver conversation.

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
EPI-CLM-benchmark-warrant	see record	SRC-1, README.md, doctrine/design and reported operation	see record	RT-RTE-verifier-continuation, EPI-RTE-hidden-semantics	No inspected run trace	No independently varied comparison; finite tests and environment contracts bound scores	Self-reported completion alone does not set full-pass reward; task tests and numeric threshold do	Binary-only default claim does not describe every bundled restarted path. Reproducing oracle outputs is narrower than establishing the mechanism or real-tool validity.
RT-CLM-grader-isolation	see record	SRC-1, harbor/skills/apply-lhtb-patches/PATCH.md, doctrine/design	see record	RT-RTE-verifier-execution, RT-RTE-runtime-patching	see record	see record	Binary prompt construction and separation mechanisms are inspectable	Best-effort cleanup, alternate runtimes, mounted logs and provider contracts prevent universal confidentiality.

Bounded conclusion

LHTB imports task instructions and terminal observations, generates candidate actions, checks task products and lets reward alter continuation. These are separate authorities: command parsing permits effects; numeric grading licenses stopping; natural-language feedback guides the next model turn. An adverse reward does not undo earlier workspace changes. See RT-RTE-terminal-actions and RT-RTE-verifier-continuation.

The inspected reconstruction evaluator can derive bounded propositions about agreement with a shipped reference computation. Its acoustics reference computes an exponentially decaying sinusoid and convolution; the judge checks length, energy and peak with tolerances. Those facts do not warrant that a candidate recovered a physical mechanism, independently reproduces a real tool across its domain, or generalizes beyond hidden scenarios. Public feedback is an additional, detailed advisory channel, not the hidden rejection channel; it even permits signature equality as an alternative to payload agreement. See EPI-RTE-public-feedback, EPI-RTE-hidden-semantics.

Checks, arithmetic and stopping are not interchangeable. A parsed score is evidence for the task's declared acceptance criterion, not an accepted ampliative explanation. No concrete model-produced proposition in the accessible run evidence establishes production, evidence-consuming acceptance and subsequent lifecycle integration of an ampliative claim. This is a bounded evidential limit, not a system-complete architectural negative: opaque agents, remaining tasks and uninspected hooks remain relevant. See EPI-RTE-dense-reward.

Model-produced handoffs retain selected historical content for later solver invocation without a semantic acceptance gate. The intended compression cannot be classified as faithful reshaping from source wiring alone. The deterministic filter is different: it preserves selected byte segments and explicitly omits the interior; an observation can therefore disappear without being rejected as false. See RT-RTE-context-handoff.

Post-run check and analyze provide a second evaluator family. They request file-informed model judgments about task coverage and reward legitimacy, reject malformed structure, save assessments, reuse them and summarize across trials. Within the inspected callers, their verdicts are advisory: no consumer withdraws a scored reward or blocks a later benchmark launch because an analysis says fail. Cache reuse is operational reliance before any demonstrated epistemic acceptance. See EPI-RTE-task-check, EPI-RTE-trial-analysis, EPI-RTE-analysis-cache, EPI-RTE-job-aggregation.

Continuation, retained summaries and post-run diagnosis independently establish implemented behavior-changing routes, not improved future capacity. Model parameter changes are not exposed in the inspected call paths, and provider internals remain uninspected. No supplied experiment establishes learning, benefits of summarization, causal effects of hardening or exercised self-improvement of the harness. Runtime patch adoption remains a separate afforded machinery-change path, see RT-RTE-runtime-patching. None of these limits collapses the independent positive findings about checking, retention or operational control.

Shared records

Components

EPI-CMP-analysis-model

Label: Analysis SDK model

Distinct from RT-CMP-terminus-model: selected through Claude Agent SDK in SRC-1, harbor/src/harbor/analyze/backend.py, with CLI defaults sonnet for check and haiku for analyze in harbor/src/harbor/cli/analyze.py. Distributed-parametric weights live outside the repository. No parameter-update operation is exposed in this backend; provider-side change is uninspected. The aliases do not pin exact weights/version; external resolution can change. Prompts and returned judgments are natural language, not the parameters themselves.

model=normalize_model_name(model), --- harbor/src/harbor/analyze/backend.py

Operative objects

EPI-OBJ-reconstruct

Label: reconstruct.py scenario behavior

Distinct from RT-OBJ-task-instruction: this is the task product, not its specification. Source-native /app/workspace/reconstruct.py is symbolic source on sandbox disk, executed on scenario JSON to produce numerical/symbolic summaries. Those outputs are truth-apt relative to a named scenario/reference; actual candidate contents are not supplied. SRC-1, tasks-2.0/acoustic_beamforming_peak__01/tests/private/judge.py.

reconstruct = args.workspace / "reconstruct.py" --- tasks-2.0/acoustic_beamforming_peak__01/tests/private/judge.py

EPI-OBJ-check-report

Label: Judge report and dense reward

Part of: RT-RTE-verifier-execution

This object is material intermediate state inside the containing verification route, not a replacement for that route. Symbolic JSON check outcomes/details and numeric text reward are written to verifier files. Public feedback is stdout JSON consumed interpretively by the solver. Private results enter Verifier and trial control. SRC-1, tasks-2.0/acoustic_beamforming_peak__01/tests/test.sh.

cat /tmp/dense_reward.out > /logs/verifier/reward.txt --- tasks-2.0/acoustic_beamforming_peak__01/tests/test.sh

EPI-OBJ-analysis

Label: Analysis judgments and summaries

No counterpart among supplied runtime objects: these are post-run task/trial assessments, not TrialResult or active solver context. Natural-language summaries and explanations coexist with symbolic check outcomes in structured model responses and disk analysis.json/analysis.md. JSON fields control cache/formatting; prose guides operator interpretation and job summarization. SRC-1, harbor/src/harbor/analyze/analyzer.py and harbor/src/harbor/analyze/checker.py.

return QualityCheckResult(checks=parsed.model_dump()) --- harbor/src/harbor/analyze/checker.py

Routes

EPI-RTE-public-feedback

Label: Public scenario feedback

No counterpart in supplied runtime routes: a task-local, solver-invoked evaluator, not hidden verification. SRC-1 implementation, tasks-2.0/acoustic_beamforming_peak__01/environment/app/tools/feedback: code runs reconstruct.py, reads public reference products, compares signature or summaries, prints results. Candidate comes from solver; fixed code decides matches; no human veto or rollback is required. Mode is bounded public-scenario testing, triggered by invoking the tool. It admits no workspace change; it supplies advisory repair evidence.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: JSON public checks, field diffs and pass count to stdout.
• Later read-back: inapplicable — this tool does not persist feedback for a later invocation; terminal history handling belongs to the agent.
• Delegated visibility: Candidate receives public scenario JSON; solver receives public expected values/diffs and signatures.
• Selection predicate: Public manifest scenarios and workspace reconstruct.py presence.
• Invalidation or expiry: Rerunning recomputes checks against current candidate; no stored verdict is withdrawn.
• Activation or effect: Feedback is delivered through terminal output if invoked; repair compliance unobserved.
• Evidence limits: Signature-match disjunction and public-only domain prevent a semantic or hidden-generalization guarantee.

passed = got_signature == expected_signature or got_summary == expected_summary or got == expected_output --- tasks-2.0/acoustic_beamforming_peak__01/environment/app/tools/feedback

The disjunction is a source-supported counterexample to interpreting public pass as necessarily equal numerical output: a dict with the expected signature and wrong summary can satisfy its first term. This is a deduction from inspected code, not an executed exploit.

EPI-RTE-hidden-semantics

Label: Hidden oracle semantic checks

Part of: RT-RTE-verifier-execution

The task-specific evaluator inside that containing harness route. SRC-1 implementation, tasks-2.0/acoustic_beamforming_peak__01/tests/private/judge.py; family selection confirmed in tasks-2.0/acoustic_beamforming_peak__01/tests/private/variant.json. Task author supplies scenario/reference/criteria, solver supplies candidate, fixed code judges. Target is artifact gates and finite hidden scenario summaries, not the candidate's explanation. Acoustics checks rir_len with absolute tolerance 6/relative .02, energy relative .25, peak relative .20 and positive energy. The reference below is executable arithmetic, not a proof of physical encoding fidelity. Tests can veto pass but do not roll back the product. Mode is bounded benchmark verification.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: JSON report and success/failure exit code.
• Later read-back: Report reaches dense scoring in the same verification; retained host rewards reach later job consumers through supplied runtime routes.
• Delegated visibility: Candidate receives scenario; private evaluator holds expected output and checks.
• Selection predicate: Artifact/config gates precede hidden loop; private oracle selected by family; loop stops on first failure.
• Invalidation or expiry: New verification recomputes against current workspace; prior saved results are not automatically erased.
• Activation or effect: Check report drives task score conversion.
• Evidence limits: Scientific source truth, images, package behavior and all-task coverage unverified; finite tolerances bound the license.

if not passed: checks["candidate_matches_hidden_semantics"] = False break --- tasks-2.0/acoustic_beamforming_peak__01/tests/private/judge.py

n = 256 + stable_int(params.get("rir_length"), 256) decay = stable_float(params.get("absorption_level"), 0.02, 0.005) rir = np.exp(-np.arange(n) * decay) * np.sin(np.arange(n) / 4) src = np.zeros(n); src[0] = 1; src[50] = 0.5 y = signal.fftconvolve(src, rir)[:n] --- tasks-2.0/acoustic_beamforming_peak__01/tests/private/offcode20_oracle.py

EPI-RTE-dense-reward

Label: Judge report to dense reward

Part of: RT-RTE-verifier-execution

This material score conversion differs from the hidden check and later stopping disposition. SRC-1 implementation, tasks-2.0/acoustic_beamforming_peak__01/tests/private/dense_reward.py and tasks-2.0/acoustic_beamforming_peak__01/tests/test.sh. Fixed ladder requires prior gates; semantic refinement adds within-stage credit. Launcher falls back to judge exit-code reward when conversion fails. No human selects a score per candidate; criterion is arithmetic, intended use benchmark credit, not explanation acceptance.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Numeric score text to launcher/reward file, or parser error causing binary fallback.
• Later read-back: Reward is retained in trial result and consumed by continuation/job stats; see supplied runtime routes.
• Delegated visibility: Converter sees private judge JSON; ordinary same-conversation solver receives binary rejection instead.
• Selection predicate: Loadable report with checks; sequential ladder and semantic detail flags.
• Invalidation or expiry: Later verification can overwrite current reward output; saved historical outcomes require separate retention policy.
• Activation or effect: Numeric output becomes the verifier result input for pass/continue.
• Evidence limits: Arithmetic cannot authenticate report provenance or establish oracle validity.

graded = stage / MAX_STAGE if stage < MAX_STAGE: graded += refine / MAX_STAGE --- tasks-2.0/acoustic_beamforming_peak__01/tests/private/dense_reward.py

Concrete deduction: {"checks":{"workspace_nonempty":true,"workspace_has_editable_artifact":true},"details":[]} clears stage 1 only; refinement is zero, so grade is round(1/6,4) = 0.1667. With no passed flag, CLI chooses that grade, not 1.0. This is checked implementation arithmetic, not observed operation or a newly validated world claim.

EPI-RTE-task-check

Label: Task quality rubric assessment

No counterpart among supplied routes: explicit CLI check, not solver command parsing or hidden grading. SRC-1 implementation, harbor/src/harbor/analyze/checker.py, harbor/src/harbor/cli/analyze.py; rubric doctrine/design, harbor/src/harbor/analyze/prompts/check-rubric.toml. Caller selects model/prompt/rubric; model judges task coverage, schemas, dependencies and cheating risks. Program validates structure, not truth. Model is the evaluator, with no independent expected-answer oracle. Mode is operator-requested assessment; optional output is saved automatically, no per-judgment human admission gate. It does not edit/reject the task itself or provide rollback.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Structured quality checks, printed table and optional caller-selected JSON output.
• Later read-back: uninspected — optional saved output has no later consumer established in these callers.
• Delegated visibility: SDK gets task working directory, file tree, rubric guidance and Read/Glob/Grep tools.
• Selection predicate: Existing valid task directory and chosen/default rubric/prompt.
• Invalidation or expiry: Reinvocation recomputes; no automatic semantic withdrawal found in inspected caller.
• Activation or effect: Advisory results reach CLI/operator; not a benchmark launch veto here.
• Evidence limits: Actual file reads, model compliance and rubric judgment accuracy unobserved; source assumptions about hidden tests are not deployment facts.

tools=["Read", "Glob", "Grep"], output_schema=response_model.model_json_schema(), --- harbor/src/harbor/analyze/checker.py

EPI-RTE-trial-analysis

Label: Trial legitimacy and specification assessment

No counterpart among supplied runtime routes: post-run diagnosis is distinct from binary in-trial rejection. SRC-1 implementation, harbor/src/harbor/analyze/analyzer.py; rubric doctrine/design, harbor/src/harbor/analyze/prompts/analyze-rubric.toml. Model receives task access when locally available and trial directory access, proposes summary and rubric judgments. The default rubric asks whether reward came from legitimate solving and whether instructions suffice. These are substantive truth-apt assessments, but request and schema do not guarantee evidence consumption or accuracy. No human per-result veto; invalid structure is rejected by code. No reward withdrawal or automatic machinery change occurs in this inspected path. Mode is requested post-run analysis.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Structured trial summary and check judgments.
• Later read-back: Results persist/reload through EPI-RTE-analysis-cache, and feed job aggregation.
• Delegated visibility: SDK gets trial cwd, read tools and optional task directory; when absent, prompt requests inference from trajectory/test output.
• Selection predicate: Selected trial directory, cache miss/overwrite; optional passing/failing filters at CLI/job selection.
• Invalidation or expiry: Overwrite requests reanalysis; unchanged cache otherwise reuses prior judgment.
• Activation or effect: Judgment enters reports and aggregation, not scored reward mutation.
• Evidence limits: Diagnosis is not causal identification; absent task source weakens specification assessments; no observed SDK trace.

"The task directory is not available locally. " "Use the trajectory and test output to infer what the task required." --- harbor/src/harbor/analyze/analyzer.py

EPI-RTE-analysis-cache

Label: Analysis validation persistence and reuse

No counterpart to RT-RTE-job-recovery: different files, consumer and consequence; this reuses model assessments rather than trial outcomes. SRC-1 implementation, harbor/src/harbor/analyze/analyzer.py. Program rejects invalid structured response, writes valid assessment, and can return cached JSON without querying the model. This is structural admission and retention, not evidence-consuming epistemic acceptance. Caller can choose overwrite; no human review gate or automatic truth-based rollback. Existing file policy controls replacement.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Validated AnalyzeResult, from fresh output or cached JSON.
• Later read-back: Later analyze invocation can reuse trial analysis; job aggregation can consume cached trial assessments and cached job summary.
• Delegated visibility: Later consumer receives parsed assessment, not necessarily original raw evidence.
• Selection predicate: Cache exists and overwrite is false; otherwise model validation and disk writes.
• Invalidation or expiry: Explicit overwrite; no evidence-change digest check in inspected cache branch.
• Activation or effect: Cache can skip model evaluation and supply prior judgment downstream.
• Evidence limits: Schema validity does not certify factual correctness, freshness or execution provenance.

if not overwrite and cached.exists(): return AnalyzeResult.model_validate_json(cached.read_text()) --- harbor/src/harbor/analyze/analyzer.py

parsed = self._config.response_model.model_validate(raw) --- harbor/src/harbor/analyze/analyzer.py

EPI-RTE-job-aggregation

Label: Trial assessment to job narrative

No counterpart among supplied runtime routes: natural-language aggregation of diagnoses is distinct from numeric job metrics. SRC-1 implementation, harbor/src/harbor/analyze/analyzer.py and harbor/src/harbor/analyze/backend.py. Code builds text from trial name, summary and checks; model returns job summary without file tools. No independent oracle or semantic acceptance gate; summary is automatically retained/returned. Model proposes narrative, code admits it; human authors rubric/prompt or later judges use, not per-content approval in this path. Mode is bounded post-run job analysis; failed trial analyses are reported separately, not silently converted into success.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: JobAnalyzeResult with generated summary and constituent assessments.
• Later read-back: Job analysis JSON can be reused when trial analyses were all cached and overwrite is false.
• Delegated visibility: Aggregation model gets summaries and checks in prompt, not raw trajectories through its tool-less call.
• Selection predicate: Successful selected per-trial analyses; passing/failing filter and cache policy.
• Invalidation or expiry: Overwrite or uncached trial analysis triggers new aggregation.
• Activation or effect: Job narrative is written and displayed; causal benefit or later operator action unobserved.
• Evidence limits: Summarization may omit or infer; failed analyses and selected subset bound coverage; retention does not establish lifecycle integration after acceptance.

return await query_agent( prompt=prompt, model=model, cwd=".", tools=[], --- harbor/src/harbor/analyze/backend.py

Claims

EPI-CLM-benchmark-warrant

Label: Hidden verification rather than self-report

SRC-1, README.md, doctrine/design, claims useful-work measurement with hidden rebuild-from-artifact verification, reconstruction of scientific/engineering tools and binary rejection until timeout/pass. These are benchmark claims, not a claim to produce accepted scientific theories. Closest supplied RT-CLM-grader-isolation concerns hardening; this claim covers what the benchmark measures and licenses. Reported leaderboard results do not supply causal evidence of generalization or installed-policy identity.

Unlike short-horizon coding benchmarks where an agent writes one artifact and stops, LHTB drops the agent into a stateful environment and grades it with hidden, rebuild-from-artifact verifiers — self-reported progress does not count. --- README.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member