# FieldHash: authority across a changing multi-agent workflow

Author: **Aaron Martinez, FieldHash, Inc.**

Public analysis bundle, version 2. Study collected September 12–13, 2026.

This bundle provides all 444 scheduled workflow rows in sanitized form, the matched case identifiers needed to reproduce the prespecified comparisons, an offline verifier, and six derived tables. It supports **exported-data and arithmetic consistency** checks. It does not authenticate the publisher and does not replay or independently classify the original evidence.

## Reproduce the disclosed analysis

Python 3.9 or later is sufficient. No third-party packages, credentials, network access or provider calls are needed.

```sh
python3 -I -B verify.py
python3 -I -B -O verify.py
python3 -I -B verify.py --json
python3 -I -B verify.py --write-derived ../recomputed-fieldhash-tables
```

Run these commands from the extracted bundle directory. The output directory in the last command must be new and outside the bundle. The verifier checks the full inventory, derives U/C/J from the exported component counts and closure flag, validates matched cases and ablation references, recomputes every result table, and compares them with the supplied tables and frozen disclosed totals. Its checks remain enabled under optimized Python.

The archive's adjacent SHA-256 digest checks that the archive matches the distributed digest. Neither that digest nor the internal manifest is a signature or a trusted publisher identity. Rewriting the data and the verification program together can defeat local checks. A party seeking assurance about the original classifications requires access to the retained evidence and an appropriate independent review.

## Contents

| File | Purpose |
| --- | --- |
| `workflows.csv` | All 444 sanitized workflow observations, without original workflow identifiers or local paths. |
| `DATA-DICTIONARY.md` | Field definitions, values, matching and outcome rules. |
| `RESULTS.md` | Main results, primary comparisons, ablations and principal limitations. |
| `METHODS.md` | Study design, controls, analysis, construction history and verification scope. |
| `verify.py` | Standard-library-only offline consistency and reproduction program. |
| `expected/main-results.csv` | All five main controls with separate completion and safe completion. |
| `expected/primary-comparisons.csv` | Six exact family-weighted contrasts, separate by model and track. |
| `expected/ablation-results.csv` | Three ablations with reused protected main baselines. |
| `expected/lifecycle-results.csv` | Separate clean and recovery cases. |
| `expected/exposure-results.csv` | Target exposure, attempt, denial and violation counts. |
| `expected/per-action-violations.csv` | All seven known per-action violations and their shared-budget flag. |
| `manifest.json` | Explicit inventory, sizes and SHA-256 hashes for every other bundle file. |

The public export omits prompts, provider responses, reasoning traces, authority credentials, internal principals, original identifiers, private paths, implementation code and raw replay material. It includes recorded model-profile names, coarse failure categories and anonymous matching references. Exported component classifications remain assertions from the retained study record; this analysis does not independently establish their truth.