research-package RP-2026-07-30-NHE-INVENTORY

Non-Human Experimental Inventory

Non-Human Experimental Inventory

Scope and search method

This inventory covers repository evidence through commit aff6ec9 and the pre-change validation run on 2026-07-30. It includes computational experiments, designed probes, preflights, failed or blocked tests, and design studies. It excludes cited external studies as repository experiments, human-subject proposals, ordinary repository validation not framed as research, and candidate tasks that were never instantiated.

The corpus was located by reading governance and the REP specification, searching all Markdown, JSON, Python, Git history, registries, cycles, journals, fixtures, outputs, and validation logs for experiment/test/run/probe language, and then checking the implemented evaluator against the records. Missing means not recorded; it does not mean zero.

Master inventory

ID What was actually done Type Executed? Unit/sample Raw data Reconstruction Capability evidence?
EX-E001 Repository artifacts and history were mapped to 16 candidate task families and eight constructs Design/observational mapping Yes One repository snapshot; no trials Narrative registry/cycle record Substantial; search protocol and item-level coding are not preserved No
EX-E002 ET-004-P01 contract and four authored outcome probes were designer-audited Task-integrity pilot Yes 1 task, 4 probes Versioned fixture files Complete for supplied artifacts No
EX-E003 ET-014-P01 contract and three authored outcome probes were designer-audited Task-defect pilot Yes 1 task, 3 probes Versioned fixture files Complete for supplied artifacts No
EX-E004 One deterministic evaluator classified the seven probes against predeclared labels In-sample grader calibration Yes 7 dependent authored probes, one execution recorded Inputs, expected labels, code; console result reproducible Complete except environment snapshot and immutable raw run record No
EX-E005 Repeated contained agent baseline Stochastic capability baseline No; blocked Planned ≥3 runs per valid pilot None Protocol only No
EX-E006 Compare contamination conditions (Git/web/gold access) Controlled comparison No; planned Not specified None Question only; design incomplete No
EX-E007 Export each pilot into an agent-readable directory and scan exported filenames for gold/probe/grader indicators Preflight/security check Yes 2 exports; narrow detector Exporter code and recorded summary; disposable exports not retained Substantial; output file hashes and raw export report are absent No

Reconstruction fields

Field EX-E001 EX-E002 EX-E003 EX-E004 EX-E007
Research question Recorded Recorded Recorded Recorded Recorded
Hypothesis Linked, not one-to-one HY-E002/HY-E006 HY-E002/HY-E004/HY-E006 HY-E003 HY-E006/HY-E009
Subject Repository corpus ET-004 contract/probes ET-014 contract/probes Evaluator + 7 probes Exporter + 2 fixtures
Date 2026-07-23 2026-07-28 2026-07-28 2026-07-28, reproduced 2026-07-30 2026-07-28
Executing agent/model/provider Missing Missing Missing No model invoked; researcher identity missing No model invoked; researcher identity missing
Intervention/comparison Candidate family coding Outcome class variations Defect diagnosis variations Expected versus actual class Source versus exported contents
Baseline/control None Base fixture Task materials Predeclared labels Source fixture contents
Controlled variables Not documented Fixture and grader Fixture and grader Code/fixtures versioned; environment partly missing Export code/fixture versioned
Software/repository version Git 0205032 for inspection Commit not recorded in experiment Commit not recorded in experiment Commit not recorded; reconstructed by Git history Commit not recorded
Seed/randomization Not applicable None; probes purposive None; probes purposive None; deterministic None; deterministic
Trials/runs One mapping exercise 4 authored cases 3 authored cases One recorded aggregate run; exact rerun succeeds 2 exports
Evaluation Researcher categorization Designer integrity audit Designer integrity audit Exact class equality Filename substring detector
Statistics None Counts only Counts only 7/7 overall, 4/4 accepted, 3/3 rejected 2/2 export; 0 flags
Confidence claim Design cycle medium Provisional Provisional High only for probes High only for current exporter
Key limitation No coding representativeness Same designer authored task/probes/grader Same designer; semantic space narrow Complete circularity risk; no unseen cases Detector ignores many content/semantic leakage routes
Reproducibility Partial High locally High locally High locally High locally if a new directory is supplied

Stated–implemented–measured–claimed mismatches

Mismatch Affected IDs Finding
“Independent review” appears in pilot independent variables but no reviewer participated EX-E002, EX-E003 The implemented activity is designer review plus probe construction, not an independent-review experiment
“Blind” describes export content, not a blinded evaluator or trial EX-E007 The check detects certain filenames only; no agent received the exports and no adversarial leakage test occurred
Seven matches can resemble seven replications EX-E002–EX-E004 They are seven purposive cases reused by one calibration run, not independent replications
Contract validity can resemble task validity EX-E002–EX-E004 Required JSON fields and known probe separation do not establish fairness, representativeness, or unseen grading validity
Design changes are used as support for audit value EX-E002, EX-E003 This is credible process evidence that audit found issues, but not a controlled estimate of prevented false conclusions
Registry status “completed” can resemble empirical completion EX-E001 EX-E001 is a design/observational mapping exercise, not a controlled experiment

Exclusions and discovery limits

  • ET-001–ET-016 are task families, not 16 experiments.
  • Repository integrity validator tests are software tests, not part of the historical evaluation research question.
  • External papers EV-E001–EV-E013 inform hypotheses but are not executions in this repository.
  • Course notes and handbook assertions contain no reconstructable trials.
  • No raw model outputs, reasoning traces, agent runs, stochastic seeds, cost logs, intervention logs, or independent ratings were located.

Coverage confidence is high for versioned artifacts and lower for uncommitted or discarded work that left no durable record. Git history is short and commit messages are coarse, so silent failed attempts cannot be ruled out.