johnoconnor0's picture
AURORA Stage-1 artefact lift 2026-05-02 — git 329cb110
1f56741 verified
metadata
title: AURORA-Workflow-1  Dataset specification
slug: aurora-workflow-1-spec
type: spec
status: active
dataset_version: v1.0.0
owner: unassigned
created: 2026-04-26T00:00:00.000Z
updated: 2026-04-26T00:00:00.000Z
sources:
  - ../../research/methodology/methods-and-controls.md
  - ../../research/methodology/data-collection.md
  - ../../research/methodology/data-management.md
  - ../../architecture/engineering-spec/schemas/event.schema.json
  - ../../architecture/engineering-spec/schemas/episode.schema.json

AURORA-Workflow-1 — Dataset specification

The primary structured-workflow dataset for Stage 1. Used by H1 (event-vs-token), H2 (sparse routing), H3 (continual learning), H4 (episodic-semantic memory), and several benchmark families (LDA, EUT, ESC, HUB).

Headline numbers

Property Value
Domains 6
Workflows per domain 50
Episodes per workflow 200
Total episodes 60 000 (6 × 50 × 200)
Events per episode (per-domain mean) 80–200 (cross-domain weighted mean ≈ 133; per-domain values are locked in each domains/<slug>/grammar.md)
Adversarial-exception rate 10 %
Temporal rule-replacement rate 20 %
Manual audit sample 5 % per domain
Splits train / validation / calibration / test / OOD / compositional-holdout
Compositional-holdout fraction 20 %
Calibration-split fraction 5 % (added per ADR-0006 for planner-objective-weight calibration)
Seeds 100 (master); per-experiment subset of 20
Estimated raw size ~80 GB

Note on totals. The 60 000-episode total is domains × workflows × episodes-per-workflow. The events-per-episode figure varies by domain (e.g. invoice-triage ≈ 120, lab-sample-routing ≈ 200, household-maintenance ≈ 80) — see each grammar's "Episode generation parameters" section for its exact value.

The six domains

Each domain has a standalone rule grammar at domains/<slug>/grammar.md:

Slug Folder Used by
invoice-triage domains/invoice-triage/grammar.md LDA, EUT, HUB, inventory-business gym
appointment-scheduling domains/appointment-scheduling/grammar.md LDA, HUB, household-scheduling gym
inventory-reorder domains/inventory-reorder/grammar.md LDA, EUT, HUB, inventory-business gym
lab-sample-routing domains/lab-sample-routing/grammar.md LDA, ESC
issue-ticket-escalation domains/issue-ticket-escalation/grammar.md LDA, code-maintenance gym
household-maintenance-planning domains/household-maintenance-planning/grammar.md LDA, LDB, HUB, household-scheduling gym

Generation pipeline

domain rule grammar (vN.0)
   ↓
seeded simulator (per-domain)
   ↓
typed event stream conformant to event.schema.json
   ↓
adversarial-exception injection (10% of episodes; per-domain exception classes)
   ↓
temporal rule replacement (20% of episodes; vN.0 → vN+1.0 over the timeline)
   ↓
manual audit (5% sample per domain; pass/fail recorded)
   ↓
split partitioning (leakage-prevention rules per domain)
   ↓
write to content-addressed object store (ADR-0001) + manifest entry
   ↓
DuckDB catalog row per episode (ADR-0002)

Each stage produces a Provenance Ledger entry; the dataset version is anchored at the final step.

Episode structure

Every episode conforms to episode.schema.json. Required fields populated by the generator:

  • episode_id — UUIDv7 derived from (domain_seed, episode_index).
  • domain — one of the six slugs.
  • actor_role — randomly drawn from the domain's actor pool.
  • event_sequence — list of typed events; mean length 200, distribution per-domain.
  • state_before / state_after — opaque-object snapshots produced by the simulator.
  • available_actions — populated from the domain's action vocabulary.
  • correct_action — set by the simulator; may be a single string, a set of equivalent strings, or null when the only valid response is a clarification question.
  • valid_clarification_questions — populated for episodes where the simulator's hidden state would justify a clarification.
  • risk_level — highest-risk action available in the episode.
  • rule_version — the active rule-grammar version (e.g. invoice-triage@v1.0 or invoice-triage@v2.0 for replaced episodes).
  • exception_flagtrue for the 10 % adversarial-exception episodes.
  • provenance_id — the Provenance Ledger entry where this episode was first recorded.
  • splittrain / validation / calibration / test / ood / compositional_holdout.

Splits

Split Fraction Purpose Leakage rules
train 55 % Hyperparameter search + apprenticeship All rule combinations from the train-pool grammar; entity names from the train pool only.
validation 9 % HP-budget selection (50 random configs) Rules and entity names disjoint from test/OOD/calibration.
calibration 5 % Planner-objective-weight calibration (ADR-0006) and per-term scale-factor estimation Rules and entity names disjoint from test/OOD; never used for HP-budget selection.
OOD 10 % Distribution-shift evaluation Reserves a per-domain OOD-grammar variant: unseen rule version, unseen entity-name pool, intensified exception rate.
test 15 % Final hypothesis evaluation Held until pre-registration is filed; never seen during HP search or calibration.
compositional-holdout 6 % Compositional generalisation Reserves rule-pair combinations not present in any other split (per methods-and-controls.md §4).

Test, OOD, calibration, and compositional-holdout splits are SHA-256-checksummed and signed; the test/OOD/compositional-holdout content is not exposed to any system before pre-registration is filed. The calibration split is exposed to the planner-weight search after pre-registration of the search procedure but before pre-registration freeze.

Sum: 55 + 9 + 5 + 10 + 15 + 6 = 100 %.

Adversarial-exception injection (10 %)

The 10 % rate is fixed across domains; the kinds of exceptions are per-domain (each grammar lists 5–8 exception classes). Injection procedure:

  1. Sample 10 % of episodes uniformly across the train + OOD splits.
  2. For each sampled episode, choose one exception class from the domain's grammar.
  3. Generate the exception-modified event stream while leaving correct_action consistent with the exception (i.e. the correct action under the exception, not the rule's default).
  4. Set exception_flag=true.

Exception episodes are critical for the H4 exception-recall metric and the M3 exception-preservation gate.

Temporal rule replacement (20 %)

For each domain, rule grammar vN.0 is in effect for the first 60 % of the timeline; a transition window from 60 %–80 % phases in vN+1.0 (probability of v1 vs v2 ramps linearly); the final 20 % uses vN+1.0 exclusively.

Half of the OOD split episodes use yet a third grammar vN+2.0 to test adaptation to genuinely unseen rule versions.

The grammar's rule-versioning policy section per-domain documents the v1 → v2 substantive changes.

Reproducibility manifest

Every published dataset version ships:

  • seed_manifest.json — master seed + per-domain seeds + per-episode seed-derivation rule.
  • grammar_versions.json — grammar version per episode (matches the rule_version field).
  • commit_pin.json — git commit SHA of the generator at dataset build time.
  • checksums.json — SHA-256 of every artefact file.
  • audit_log.csv — manual-audit pass/fail per audited episode.
  • provenance_anchor.json — Merkle root from the Provenance Ledger at dataset build time.

These files live alongside the dataset under the content-addressed object store (per ADR-0001).

Manual audit (5 %)

Per-domain auditors run through 5 % of the episodes against the audit checklist in each grammar file. Failures trigger:

  • For systematic failures (>1 % pass rate fall): generator bug; halt dataset build, fix, re-run.
  • For isolated failures: log to audit_log.csv; episode is excluded from the build.

The audit must complete and pass before any test-set evaluation is permitted.

Versioning

v1.0.0 is the first publishable version. Bump policy:

  • Patch (v1.0.1): bug fix in the generator; no semantic change to episodes; existing dataset is repaired in place if possible, or republished with the same content hash if not.
  • Minor (v1.1.0): new domain added or new exception class; existing episodes unchanged; new episodes appended.
  • Major (v2.0.0): change to event schema, episode schema, or episode-distribution that invalidates pre-registration. Triggers re-pre-registration.

Privacy

All synthetic. No PII. Vendor / customer / patient names drawn from a hardcoded synthetic-name pool. The dataset is privacy_tier='public' by default; downstream events derived from interaction with the dataset inherit higher tiers per event.schema.json rules.

Cross-references

Open questions

  • Whether 5 % is enough for manual audit at 60 000 episodes (= 3 000 episodes auditable). Currently fixed; revisit after first build.
  • Whether the OOD split should also include an "exception-rate spike" condition (e.g. 25 % exceptions) in addition to unseen rule versions. Currently no; exception_flag=true rate stays at 10 % across splits except where explicitly noted in a domain grammar.
  • Whether the compositional holdout should be expanded for H1's compositional-generalisation claim. Currently 5 %; revisit if pre-registration negotiation increases sample-size demand.
  • Whether dataset versions should be cross-referenced from the Memex wiki entity page for the project, or only from this spec. Currently only from this spec; the wiki cross-link is a one-line addition.