# Meta ARE scenario matrix

## What this pack lets workers do

A collection of nouns and verbs for generating 1,000 dynamic environments, 1,000 decomposition
methods, 1,000 operator personas and 1,000 multi-position stakeholders, evaluating stratified
4D Latin-hypercube sweeps or `10^12` addressable tuples, and digesting safe trajectories into
reusable **Tasks** (verbs) and **Task Objects** (nouns). Inspired by Meta Agents Research
Environments (ARE / Gaia2) and Meta Matrix synthetic multi-agent scaling, workers test how
public-good and technical workflows behave under asynchronous rule updates, OCR noise, lease
preemption and strict privacy regimes before any task is proposed for review.

## Nouns

| Noun (task object) | What it holds |
| --- | --- |
| Guardrails | Block unsplit cloud egress on high-PII or offline airgap data; keep crisis care, eligibility and legal judgment `human_only`; never claim synthetic sweeps are field deployments |
| `spec.four-axis-matrix` | `1,000` environments (`25x10x4`), `1,000` methods (`5x5x40`), `1,000` personas (`5x5x5x8`), `1,000` stakeholders (`10x10x5x2`) and `10^12` bijective tuple indexing |
| `rules.stakeholder-digestion` | Deterministic veto, private/public lane-partitioning and SHA-256 candidate digestion rules |
| `tool.scenario-matrix` | Tested standard-library Python package in [code/](code/README.md), pinned by path, file count and digest |
| `template.matrix-digest` | JSON Schema for stakeholder-gated episode verdicts, block reasons and digested candidate slugs |
| Worked example | Three 4-axis scenario tuples (`example.matrix-episode-bundle`) and recomputed digest (`example.matrix-digest`) |

## Verbs

| Verb (task) | Who | What happens |
| --- | --- | --- |
| `generate-four-axis-catalog` | agent, no model | Materialize all four `1,000`-item axes with unique SHA-256 content hashes and `human_only` boundaries |
| `evaluate-stakeholder-episodes` | agent, no model | Run a `1,000`-cell 4D Latin-hypercube sweep or bounded slice through multi-position stakeholder gates |
| `digest-reusable-nouns-and-verbs` | agent, no model | Rank approved and lane-partitioned trajectories into candidate Tasks and Task Objects for distinct review |

## How to use it

Run `make check` in [code/](code/README.md) to compile with warnings treated as errors and execute
all unit tests including the `1,000`-episode Latin-hypercube sweep and worked-example recomputation.

## Limits

- Quarantined, unverified proposal until reviewed: this pack is not installed, imported, admitted or
  activated, and no worker has claimed or run these tasks through Tessera.
- The `1,000`-episode sweep and `3`-tuple worked example are local deterministic synthetic evaluations (`mode: local_verified`), not field deployments.
- Stakeholder rules enforce structural privacy, offline and review gates, while accountable people still own crisis care, eligibility decisions and final release.