# Proposed design only; no study performed.
title: 'TRACE: source-conflict localization and independently predictive discovery'
publication_version: '0.1'
design_version: '0.1'
prepared: '2026-11-15'
status: Proposed, unimplemented, unregistered and unexecuted
primary_question: Does supplied conflict localization improve candidate generation and sealed prediction beyond
  established contrastive discovery with the same facts and resources?
independent_unit: Disjoint generating episode or source cluster; branches, predictions, seeds and source copies
  are not new independent cases
phases:
  development:
    case_range:
    - 12
    - 20
    purpose: Inspect representation, identifiable targets, controls and comparator fidelity
  exploratory_pilot:
    case_range:
    - 30
    - 60
    purpose: Frozen methods on unseen cases and outcomes; feasibility and paired effect estimation only
    discovery_negative_nonidentifiable_counts: Determine during development and freeze before pilot; report every
      denominator separately
  confirmation:
    sample_size: null
    status: Conditional on pilot; justify power, clustering, worthwhile effect, safeguards and multiplicity independently
  adaptive_H3:
    status: Conditional subsequent study; do not pool with static H2
  transfer:
    status: Conditional; no fixed scenario count or execution grid
hypotheses:
  H2:
    role: primary
    comparison: C minus B
    endpoint: Final source-faithful relationship absent from packet; >=4/5 predictions correct including all prespecified
      diagnostic settings
    novelty_scope: Initial source packet, not science-wide priority
    missing_measurements: No invented values; pending cases score no static discovery success
    target_settings: Lock relationship, parameters and final candidate before revealing prediction contexts; same
      task dictionary, no answer-selective fields; lock predictions before outcomes
    denominator: All evaluator-certified eligible cases fixed before runs, including abstentions and invalid/missing
      predictions; separate control denominators
    prediction_scoring: Point prediction with evaluator-fixed tolerance or single category; intervals/probabilities
      scored separately, never broaden primary acceptance
    adjudication: Two blinded reviewers, third for disagreements; report initial agreement and unresolved cases;
      no feedback to runtime
  H1:
    role: supporting quality measures
    metrics:
    - base correctness
    - nuisance/relevant response
    - coverage
    - false-condition promotion
    - unsupported resolution
  H3:
    role: conditional subsequent investigation
    endpoint: Evidence-supported scoped conclusion within action and cost budget
    required_comparators:
    - Contrastive Discovery adaptation
    - POPPER when implication/conditional-validity assumptions can be met
    - ordinary experimental design
    - random and relevance policies
systems:
  A: Strong iterative hypothesis refinement; simple counterexample-prompting variant
  B: Contrastive Discovery adaptation
  C: B scaffold with generated conflict record and conflict-directed generation; common neutral lineage treatment
  D: C plus explicit dependence-aware weighting and warrant accounting
  B8: Common-extractor ordinary rule/tree/regression learner, exhaustive when inexpensive
comparison_design:
  primary: Static C versus B, with shared source snapshot, model, proposal capacity, row weights and blinded common
    selector; effect of localization record plus instruction
  lineage: D versus C with identical metadata. B/C row weights and minimum rows; D lineage weights, distinct-lineage
    eligibility and warrant accounting. Conditional package effect.
  method_fidelity: Report faithful method-level comparator separately if shared scaffold omits essential published
    method components
  information_control: All have original spans, shared extracted facts and lineage metadata; no true moderator,
    mechanism or eligibility labels
  oracle_localization: Human-corrected records only as separately labeled diagnostic
  end_to_end: Independent extraction/localization, equal source access, errors and costs charged
  candidate_ordering: Fixed pool, vary ranking only
  numerical_ordering: Fixed table/language, contrast order versus seeded permutation and exhaustive search
  factorial: Add full localization x lineage factorial if interaction/main effects become targets
  static_B: Contrast-set generation adapted to fixed evidence; common selector replaces native scoring. No claim
    to reproduce or outperform published adaptive search.
  faithful_B: Native-method comparison with and without generated localization and equal-budget summary control;
    retain scoring/search and report separately. See native_method_transfer.
  common_selector: Hide system identity, conflict IDs and proposal order. Rank source-context coverage without invented
    values, then fewer unsupported assumptions, then canonical ID. Shared numerical training criterion. No sealed
    outcomes.
  preprocessing_control: Equally budgeted source-summary call for B; C uses its call for localization; all costs
    included
  partition_semantics: Residual explanation and unresolved epistemic state are distinct; log overlapping/incomplete
    partitions
pilot_budgets:
  input_tokens_per_case: 16000
  output_tokens_per_case: 4000
  scope: All static extraction, preprocessing, generation, repair and common selection; faithful method-level competitors
    can allocate calls differently
  external_actions: 0
  primary_models: 1
  primary_seeds: 1
  usage: Meter actual calls, tokens, CPU time, latency and costs; no forced spending
later_action_ceiling:
  actions: 3
  cost_units: 12
  status: Engineering starting setting for conditional H3, not guaranteed optimal
pilot_analysis:
  status: Exploratory, no superiority claim
  report:
  - paired root-level effects and sampling-appropriate uncertainty
  - actual positive/control denominators
  - all predictions and invalid outputs
  - candidate equivalence with blinded adjudication
  - fidelity, coverage, false promotion and cost
  - source-cluster dependence and exclusions
  limitations:
  - No equivalence from nonsignificance
  - No 5% safety certification from small error-free control stratum
  - No post hoc pooling of adaptive/static or seeds as independent cases
  confirmatory_fields_pending:
  - estimand and sampling frame
  - minimum worthwhile effect and quality margins
  - paired discordance and cluster-aware power
  - sample size
  - primary comparison and multiplicity
  - valid interval/test and handling of invalid outputs
  generation_gate: If candidate yield is unchanged but selection improves, test selection. Inconclusive small-pilot
    results do not establish equivalence.
gates:
- representation without invented values
- different measurable candidates versus B
- sealed predictive value without quality loss
- cost and fresh case construction
stop_or_reduce:
- No useful localization effect
- Unsupported conditions dominate
- Gain fully explained by extraction or deduplication
- Prior work already covers operational distinction
families:
- F1 confounding
- F2 effect modification
- F3 selection/missingness
- F4 measurement/representation
- F5 regime change
- F6 dependent reporting
- F7 rare/nonlinear interactions
- F8 observational equivalence
family_use: Coverage taxonomy; no equal-size allocation. Certify target/view identifiability without runtime labels.
freeze_requirements:
- exact models and prompts
- all source packets/licenses and disjoint splits
- candidate and repair limits
- comparator fidelity and adaptation details
- field-level provenance and localization cost
- five prediction settings, tolerances and novelty rubric
- independent units and denominators
- quality margins and analysis plan appropriate to study phase
- sealed evaluator and immutable commitments
prior_art_audit: https://gtcode.com/papers/trace/prior-art-audit.md
contribution_matrix: https://gtcode.com/papers/trace/contribution-matrix.md
method:
  specification: paper_sections_4_1_through_4_6
  stages:
  - alignment
  - lineage_accounting
  - rival_generation
  - investigation_and_warrant
  record_schema: https://gtcode.com/papers/trace/trace-records.schema.json
  variable_provenance:
  - observed_shared
  - observed_asymmetric
  - inferred_unmeasured
  extraction_calls: 1
  generation_calls_per_round: 1
  schema_repair_calls_per_round: 1
  proposals_per_round: 8
  active_candidates: 4
  static_rounds: 1
  adaptive_rounds: 2
  candidate_forms:
  - interpreted_expression
  - conditional_rule
  - action_specific_commitment
  - pending_prediction
  prediction_expression_operators:
  - constant
  - variable
  - add
  - subtract
  - multiply
  - divide
  - lt
  - le
  - eq
  - and
  - if
  variable_growth: Allowed with source or conjecture provenance and explicit measurement dependencies; never impute
    missing observations from explanatory convenience.
  prioritization:
  - localized_incompatibility
  - supported_context_coverage
  - fewer_new_assumptions
  - canonical_candidate_id
  prompts: Paper section 4.3; comparator adaptations in section 8.2; serialized templates frozen before exploratory
    pilot
  semantic_adjudication: Blinded frozen fidelity and novelty rubric; never feedback to runtime.
  independence: Distinct origins are not a statistical independence certificate.
  priors: Conflict localization is a proposed generation aid, not a calibrated truth prior
  localization_calls: 1
  localization_instruction: Cite observed incompatibilities and missing context; preserve unknowns, propose no explanatory
    condition
  comparison_override: Controlled B/C use shared selector and row-based compiler; full TRACE ranking and lineage
    accounting tested separately
contracts:
  integrity: Restricted representation/evidence-integrity checks with validated preconditions; semantic cases can
    remain unknown.
  rival_response: Hypothesis-indexed response prediction, not a truth/admissibility oracle.
  evaluator: Sealed gold target/response relation used only after output commitment.
  evidence_deletion: No universal monotonic confidence rule; preserve support and defeat paths.
acquisition_policy:
  type: proposed deterministic normalized pairwise set separation per positive cost
  prediction_region: Comparable outcome region after frozen tolerance/noise expansion exactly once
  candidate_set: Deduplicated with frozen normalization; residual duplicates audited
  eligibility: At least two live comparable candidates, all providing admissible predictions for the action
  missing_prediction: Action unscorable; logged, not silently dropped from pair denominator
  zero_cost: Common initial lookup evidence, excluded from positive-cost acquisition scoring
  tie_break: Stable action ID assigned independently of hidden mechanism
  no_positive_score: no_separating_action; never implies resolution
  limitations: Not optimal information gain or logical refutation of stochastic models; may miss useful overlapping/repeated/multivariate
    observations
  action_menu: Public and equal; audit names/costs/outcomes for answer cues, including menu-only controls
  prerequisite_measurements: When all-candidate separation is unscorable because of missing dependencies, choose
    the lowest-cost permitted measurement supplying an explicit dependency; ties by candidate priority then stable
    action ID. Same total action budget; log purpose as data completion, not information gain.
candidate_failure_policy:
  all_inadequate: Reopen only within fixed candidate/round/action budget; otherwise unresolved/model-inadequate.
    Never promote least-bad failed candidate.
  adequacy: Noise-aware thresholds and sequential/stopping rules pending freeze; an interval tail event is not automatic
    falsification.
  warrant: Dependency-aware support/defeat graph, not a scalar weakest-link ladder.
numerical_compiler:
  scope: One Stage-C candidate compiler, not the full architecture
  rule_language: constant_or_two_constants_split_by_one_or_two_predicates
  minimum_measurement_lineages_per_leaf: 2
  numeric_thresholds: observed_midpoints
  fit: equal_lineage_weighted_least_squares
  target_scale: declared_positive_per_question
  predicate_penalty_initial: 0.01
  penalty_status: development_justification_and_freeze_required
  small_table_policy: exhaustive_search_when_affordable
  missing_cells: requires_measurement_no_silent_complete_case_selection
  search_order_diagnostic_prefix: 64
  prefix_scope: high_dimensional_distractor_cases_with_budget_curves_not_a_B8_cap
  outputs_outside_supported_domain: unknown
  comparison_weights: B/C equal unique-row weights and minimum rows; D lineage weights and minimum distinct lineages;
    exact duplicate records collapsed for all
claims_not_made:
- experimental validation
- completed TRACE implementation
- empirical superiority
- guaranteed truth
- universal OOD generalization
- exhaustive novelty
- preregistered study
- current vendor pricing
readiness: Conceptual publication only; comparator implementations, data and statistical choices remain unexecuted
canonical_protocol_page: https://gtcode.com/guides/trace/evaluation/
native_method_transfer:
  purpose: Test whether generated conflict localization improves an existing capable discovery method, separately
    from controlled static C-B.
  conditions:
  - Native discovery method with original evidence
  - Same native method plus generated conflict records and usage instruction
  - Same native method plus equally budgeted source summary
  preserve:
  - hypothesis generation including conditional explanations
  - native belief scoring and search
  - revision and stopping rules
  controls: Same source facts and analysis access; no TRACE selector/compiler substitution; no sealed outcomes accessible
    to tools; charge all preprocessing, elicitation, failed attempts and analyses.
  budget: Set a common ceiling during development that accommodates native operation; do not automatically impose
    the static pilot token ceiling on adaptive search.
  endpoint: Lock native selected explanation at budget boundary; same outcome-blind prediction interface and sealed
    scoring. Report native resolution separately.
  implementations:
    Contrastive_Discovery: Author code listed forthcoming; any paper-based implementation labeled reproduction attempt
      with fidelity gaps.
    AutoDiscovery: Released implementation at 30e080f4ecf50912bb0771abb4accbf1e67bb736; document differences from
      paper defaults.
  reporting: Pin versions and adapters; distinguish fixed-evidence analysis from adaptive acquisition. Static-only
    gains do not establish transfer; inconclusive results do not establish absence.
