PROMPT 1 - P33-v2.1a - AUTONOMOUS RESEARCH PORTFOLIO / PARAL...

Prompt

PROMPT 1 - P33-v2.1a - AUTONOMOUS RESEARCH PORTFOLIO / PARALLEL COMPUTE / RESOURCE ALLOCATION / RECURSIVE OPTIMIZATION MODE: EXACT-PAYLOAD / SEARCH-OFF / P-ONLY / DESIGN-AUDIT INPUT BOUNDARY Use only this complete payload. Do not use Web/Search, prior Prompt 1 history, M/Z, sibling P files, provider identity or hidden platform state. Mechanisms below are design text, not runtime proof. UNKNOWN != ABSENT; PARTIAL != TOTAL; DESIGN-SOUND != IMPLEMENTED. DELIVERY / ARTIFACT CONTRACT This family is one standalone exact UTF-8 TXT dispatch artifact. Controlled external run = fresh clean user-controlled session containing only this family payload: no sibling P, M/Z, prior result, extra preamble or model-specific follow-up. Aggregate Px.docx is optional noncanonical review/archival only and never substitutes for or changes TXT identity. If serialization or hidden provider state is not observable, equivalence is ENVIRONMENT-UNVERIFIED rather than inferred. Visible Search/tool use despite SEARCH-OFF is an environment mismatch. EVALUATION RULES - Prefer REUSE -> MERGE -> EXTEND -> NEW. NEW requires a distinct surviving failure mechanism. - Exact applicable counterexample + decision consequence outrank vote count, confidence and prose volume. - Reject claimed defects closed by explicit text. Retain only compliant bypass, material ambiguity, decision-incomplete state or harmful false positive. - Positive identity/currentness/completeness/independence/applicability/authority claims require positive evidence. - Candidate-generation evidence cannot validate its own candidate. Correlated parallelism/replication increases coverage but not independent truth credit. - Missing evidence is not zero; labels/audit outputs never create action, treatment, deployment or release authority. - Blocking is dependency-scoped; unrelated safe/read-only/protective work remains available unless independently blocked. - Preserve benign paths and identify over-blocking as a defect. - This run is one evaluator slot; it is not independent confirmation of itself. REQUIRED OUTPUT - EXACT ORDER 1. EXECUTIVE VERDICT - 3-6 sentences; exactly one dominant NEXT ACTION. 2. AUDIT TRACE / REJECTED CANDIDATES / UNCERTAINTIES. 3. SECTION RATINGS - DESIGN-SOUND / DESIGN-DEFECT / UNVERIFIED / N/A, one sentence each. 4. STRESS-TEST MATRIX - every A-T -> BLOCKED / BYPASS / UNRESOLVED / ALLOWED-BENIGN; cite exact clause. 5. TOP DOMINANT FINDINGS - max 5; claim, exact clause, impact, path, minimal repair, risk/complexity, falsifiable validation, disposition. 6. REDUNDANCY / MERGE CANDIDATES + CONCISE RESIDUAL REGISTER. 7. MISSING-CONTROL CANDIDATES - only if not encoded or directly mergeable. 8. RECOMMENDATION SET - max 5; REUSE/MERGE/EXTEND/NEW/REJECT/DEFER + falsifiable validation. 9. FINAL SCOPE STATEMENT - exact-P design audit only; implementation effectiveness UNVERIFIED; no action/deployment/release authority. AUDIT TARGET A. SOURCE-PORTFOLIO SEARCH A1. Material methodology research is a portfolio across source roles: current primary/authoritative guidance; foundational peer-reviewed methods; recent frontier work; contradiction/negative/failure/replication evidence; independent reproduction where available; and mature adjacent disciplines carrying distinct mechanisms. A2. Maintain SOURCE-PORTFOLIO fields: source role, discipline family, date/version, authority/maturity, directness, contradiction state, applicability boundary and dependency. Many sources from one lineage do not become independent perspectives. A3. For load-bearing conclusions, run a contradiction/failure lane and seek at least one source outside the originating method lineage when feasible. Currentness-sensitive claims use current primary evidence. A4. Search stops on decision sufficiency or diminishing marginal information value, not a fixed link/pass count. Rare high-impact lanes cannot be retired solely because average yield is low. A5. Source diversity and query diversity are separate. Multiple paraphrased searches hitting the same evidence lineage do not count as independent breadth. B. IMPROVEMENT-DAG / PARALLELISM B1. Substantial `pracuj` work is represented as an IMPROVEMENT-DAG. Node fields: inputs, outputs/state, mutability, authority, cost estimate, expected information value, safety/decision impact, dependencies and merge contract. B2. Approximate WORK and CRITICAL PATH/SPAN. Parallelize genuinely dependency-independent nodes; synchronize at barriers before consuming unfrozen shared outputs. B3. Idle capacity may pull the highest-value ready node, but no dependent node may execute early merely to increase utilization. A blocked critical-path node is not fixed by adding parallel workers to unrelated nodes. B4. Parallel outputs are normalized, fingerprinted, deduplicated and reconciled against the same frozen baseline. Parallel/correlated lanes earn throughput/coverage, not automatic independent evidence credit. B5. Speculative parallel branches are allowed only when their side effects are isolated/reversible and the merge rule is fixed before results are seen. Losers cannot mutate canonical state. C. MULTI-FIDELITY RESOURCE ALLOCATION C1. Begin broad with heterogeneous candidates and cheap discriminators: semantic dedup, explicit-clause closure, deterministic invariants, cheap counterexamples and source-quality checks. C2. Escalate compute to high-impact, high-uncertainty, novel, contradictory, hard-to-falsify or boundary-adjacent candidates. Expensive Web research, fresh-context adjudication, mutation/metamorphic suites, model checking or external P slots are allocated when expected information gain justifies them. C3. Cheap proxies may prune low-value/redundant candidates but cannot eliminate catastrophic/high-impact safety candidates without a strongest-disconfirmation path or qualified defer state. C4. Allocation is multi-objective: expected validity/safety/robustness/portability gain, uncertainty reduction, novelty/coverage, reversibility, complexity, time/token/tool cost, blast radius, independence deficit and monitoring burden. Pareto alternatives remain visible; any scalar queue key is operational only and its components remain recoverable. C5. Adaptive test-time compute gives more effort to difficult/risky/uncertain tasks and less to settled/easy tasks. Stop when marginal probability of changing the decision is low or additional same-lineage runs add mostly correlated evidence. C6. Resource policy itself is evaluated: predicted information value, realized decision changes, wasted duplicate compute, latency/critical-path impact and missed high-value candidates are logged and used to recalibrate allocation. D. AUTONOMOUS / EVOLUTIONARY IMPROVEMENT D1. Candidate generators include failure-driven repair, cross-domain analogy, simplification/retirement, OPRO-like textual optimization, mutation/recombination, prompt/program optimization, self-critique, external/user misses and quality-diversity exploration. D2. Maintain bounded semantic diversity, e.g. islands/novelty clusters, so one early lineage does not monopolize compute. Periodically inject fresh external candidates; preserve useful negative results and failed-but-informative branches. D3. Generator/optimizer cannot change its own objective, acceptance boundary, protected invariants, hidden/held-out set, authority model or primary evaluator within the same optimization episode. D4. A material change to generator, evaluator, scheduler or meta-optimizer creates a successor meta-epoch with frozen acceptance criteria, shadow validation, regression/portability checks and rollback. Same-process self-approval earns no independent credit. D5. Significant user/external improvements that AIG reasonably should have found become EXTERNAL-MISS regression fixtures and update discovery coverage. D6. Recursive self-improvement has an explicit stopping condition: no further meta-iteration merely because self-modification is possible. Continue only when expected decision value exceeds complexity/risk/cost and oversight remains qualified. E. SCHEDULER HEALTH / CONTINUOUS APPLICATION E1. Track predicted vs realized information value, useful-finding rate, independent-source diversity, duplicated compute, critical-path delay, parallel utilization, correlation, tool failure and merge conflicts. E2. Downweight/retire chronically low-yield redundant lanes, but protect low-frequency catastrophic-hazard lanes from average-yield optimization. E3. At end of a material state transition, if another autonomous `pracuj` bundle has positive expected decision value and no different user decision is required, explicitly invite exact command `pracuj`. Do not repeat on unchanged state or when only an external dependency remains. E4. One `pracuj` executes the largest safely compatible chain; do not stop at artificial phase boundaries when search, tests, integration and next-P materialization are safe in the same run. E5. When a lane repeatedly predicts high value but produces none, investigate calibration, source access, query formulation and hidden dependencies before either retaining or retiring it. STRESS TESTS A-T A. Ten recent AI papers from one lab support a new mechanism; no foundational, contradiction or adjacent-domain source is checked; result is called broadly corroborated. B. One regulator/standard and one independent failure report contradict five same-lineage frontier papers; the five-paper majority wins automatically. C. Current safety/currentness claim relies only on an old foundational paper despite current primary guidance being available. D. Contradiction search repeatedly yields rare but severe failures; lane is retired because average hit-rate is lower than mainstream search. E. Current+foundational+failure+adjacent-domain search converges on same mechanism with dependencies/conflicts recorded. F. Four research/test tasks share no dependencies but are serialized while compute is idle. G. Integration node consumes search output before contradiction/provenance lanes freeze; later conflict is overwritten. H. Independent ready node is pulled by idle worker while critical-path node waits on a real dependency. I. Ten parallel copies of the same judge on the same evidence are counted as ten independent confirmations. J. Duplicate semantic candidates receive full expensive evaluation before fingerprint/dedup. K. Cheap proxy score is low for a catastrophic safety candidate; candidate is discarded without disconfirmation. L. Low-impact candidate consumes fresh-context multi-judge testing while a high-impact unresolved boundary case remains under-tested. M. Easy settled cases each receive five passes; one difficult high-impact case receives one pass solely for uniformity. N. Additional same-lineage passes have near-zero chance of changing disposition; system keeps spending budget. O. Candidate search uses multiple objectives; one high-validity/high-cost and one lower-cost/reversible alternative remain Pareto-visible. P. A single opaque scalar hides that chosen candidate has high blast radius and low reversibility. Q. Evolutionary search keeps only current best lineage; novelty clusters are discarded and search stagnates. R. Optimizer edits acceptance threshold and evaluator so its own candidate passes. S. New scheduler/meta-optimizer successor is shadow-tested under frozen criteria; regression/portability checks pass and rollback exists. T. Post-result `pracuj` finishes search but asks for another `pracuj` before compatible internal tests/integration/P materialization. SUPPLEMENTAL DISCRIMINATORS U1. Search lane discovers a formal-method invariant already semantically covered by an existing control. Adding a new control would duplicate it; expected disposition MERGE/REUSE. U2. Frontier vendor claim reports a large speedup but no independent reproduction; allocator routes it to VENDOR-EARLY evidence and a bounded independent-validation candidate, not immediate promotion. U3. Adjacent discipline offers a mechanism with incompatible assumptions; transfer is rejected with explicit applicability boundary rather than adopted for novelty. U4. Two parallel lanes share the same hidden benchmark and rubric; scheduler records common-cause dependence before merge. U5. One lane returns a negative result that falsifies a favored candidate; optimizer cannot suppress it to preserve population fitness. U6. Cheap screen deterministically closes a wording-only duplicate; expensive evaluation is skipped and compute reassigned. U7. High-impact candidate remains uncertain after cheap screen; successive-fidelity escalation spends additional compute until boundary resolves or DEFER is justified. U8. Resource scarcity forces choice between many medium-value lanes and one catastrophic unresolved lane; catastrophic lane retains protected floor allocation. U10. `pracuj` has two independent research lanes and three independent test-generation lanes; they run in parallel and synchronize before integration. ADDITIONAL PROCESS AUDIT V1. Scheduler must distinguish compute saturation from dependency blockage: more workers cannot shorten a true critical path. V3. Candidate novelty is mechanism-level; a new acronym without a new invariant, detector, control boundary or efficiency gain is overlap. V4. If two allocation policies disagree materially, retain both as Pareto/dispute candidates until a decision-relevant criterion or fresh test resolves tradeoff. V5. An improvement that reduces tokens but increases correlated evaluator dependence is not automatically efficient; efficiency includes evidence quality. V6. Search breadth should not create unlimited scope: each new domain must name a distinct threat/control/measurement mechanism and bounded support-domain expansion. V7. Population diversity should be measured by semantic mechanism/failure family, not by wording distance alone. V8. A candidate that performs well only on generator-known fixtures but fails fresh adversarial cases is overfit and cannot promote. V9. Compute allocation after an outcome is known must not retroactively alter the acceptance criterion used to select the outcome. AUDIT QUESTIONS - Does the design distinguish coverage from independence for parallel lanes? - Can the scheduler identify critical-path blocking vs mere lack of utilization? - Is a candidate ever pruned by a cheap fidelity whose false-negative cost exceeds compute savings? - Are source-role diversity and discipline diversity measured separately? - Can population diversity become uncontrolled scope growth, and is there bounded retirement? - Does the meta-optimizer have a protected objective/evaluator/holdout boundary? - Is there an explicit stop rule against recursive self-improvement with diminishing value? - Are realized compute-value metrics used to improve future scheduling without letting the scheduler validate itself?