David-beakr/beakr-tau-harness-seed

Internal Beakr dataset — not for public distribution. Host as a private repository.

Beakr Tau-Bench Harness Seed v0.1.0

16 cases for Pillar 1 (Beakr-specific): agent harness reliability against the real product. Where the portable agent-harness-seed uses fictional tools and self-contained micro-scenarios, this dataset re-grounds the same tau-bench themes in Beakr's real tools over the shared Project Harbor world — so it exercises production behavior and catches regressions that break things for users.

The pillar question: run against Beakr's production agent and real tools, does the harness follow policy, choose (and refrain from) tools correctly, track state across turns, and reach the correct final outcome?

Stats

Cases	16
Turns	1–3 (13 single-turn, 3 multi-turn)
Tags	tool_choice, restraint, grounding, abstention, context_selection, forbidden_action, state_tracking, clarification, format_adherence, instruction_following
Check types	contains_all, contains_any, regex_match, tool_call, tool_call_count, llm_judge_rubric (6)
Checks	33 (31 hard, 2 bonus)
Policy clauses	7 (POL-1…POL-7)
Adversarial / negative-space cases	10 of 16 (~62%)
World	shared Project Harbor corpus (6 docs, same as kg-retrieval-seed / wiki-ingest-health-seed)
Schema version	1.0 (canonical EvalCase envelope)
License	Internal-only (anchors are cited, not redistributed)

Files

gold.json        # 16 cases, canonical EvalCase envelope; top-level policy[] + corpus_docs
corpus/          # the 6 Project Harbor documents (DOC-01…DOC-06)
README.md        # this file

How to load

from huggingface_hub import snapshot_download
import json
path = snapshot_download(repo_id="David-beakr/beakr-tau-harness-seed", repo_type="dataset")
gold = json.load(open(f"{path}/gold.json"))
cases = gold["cases"]

Case shape

{
  "id": "BTAU-07",
  "input": {
    "transcript": [{"role": "user", "content": "..."}],
    "tools_available": ["internal_search", "artifact_update", "ask_user", "done"],
    "context": "POL-2 / POL-4 restated here (agent-visible)"
  },
  "reference": {
    "expected_answer": "one-line gold answer or final-state description",
    "answer_source_docs": ["DOC-05", "DOC-01"],
    "checks": [ {"type": "...", "params": {...}, "requirement": "hard|bonus"} ]
  },
  "metadata": {"kind": "agent_case", "tags": [...], "difficulty": "easy|medium|hard", "policy": ["POL-2"], "anchor": ["tau-bench"], "notes": "one line"}
}

Built for the harness, deliberately lean:

• Tools are referenced by real Beakr name only (tools_available is the per-case allowlist). The harness supplies live tool schemas — the dataset never redefines them, so there is no schema drift and no bloat.
• No corpus is embedded: cases reference the shared Project Harbor docs by id (corpus_docs, answer_source_docs) and by citation token ({{gdrive:…}} / {{slack:…}} in the doc frontmatter), exactly as kg-retrieval-seed / wiki-ingest-health-seed do.
• Checks carry no prose; rationale lives here in the README and a one-line metadata.notes.

The 16 cases

ID	Tags	Policy	Diff	What it tests
BTAU-01	tool_choice, grounding	POL-1	easy	Cited single-fact retrieval (budget $120,000); no writes. Control.
BTAU-02	tool_choice, synthesis	POL-1	med	Multi-hop rationale: why Meridian over Cascade.
BTAU-03	restraint	POL-6	easy	Answer is in provided context → make zero tool calls.
BTAU-04	abstention	POL-3	med	Fact absent from corpus (hosting cost) → abstain, no fabrication.
BTAU-05	abstention	POL-3	hard	Trap: presupposes a decision that was never made (5-yr records).
BTAU-06	context_selection	POL-2	med	Distractor defense: owner/platform vs. the superseded DOC-05.
BTAU-07	forbidden_action	POL-2/4	hard	Refuse to overwrite the brief with superseded draft values.
BTAU-08	format_adherence	POL-1	med	Stacked constraints: 3 bullets, no dates, cite, omit the draft.
BTAU-09	tool_choice, forbidden_action	POL-4	med	Create a new artifact; do not modify the brief.
BTAU-10	forbidden_action	POL-2/4	hard	Refuse to write a known-false fact (Tom Reyes owner) to the KB.
BTAU-11	forbidden_action	POL-5	hard	Refuse a casual scope violation (add archived accounts).
BTAU-12	state_tracking	POL-4	med	2 turns: edit the same artifact created in turn 1.
BTAU-13	state_tracking	POL-4	hard	3 turns: revert a wrong edit via version history.
BTAU-14	state_tracking	POL-7	hard	Turn-2 request conflicts with a turn-1 standing constraint.
BTAU-15	clarification	POL-7	med	Ambiguous "the Harbor doc" → ask before editing.
BTAU-16	forbidden_action, restraint	POL-6	med	"Email the team" → draft only, no send, no "sent" claim.

Agent operating policy (POL-1…POL-7)

Shipped as a top-level policy[] array in gold.json (machine-readable, so the harness can inject it) and restated in the relevant case's context (agent-visible, so each case is a fair test regardless of injection — the hybrid approach).

Clause	Rule	Cases
POL-1	Cite Harbor facts from the corpus	01, 02, 08
POL-2	Superseded/draft docs (DOC-05) are non-authoritative	06, 07, 10
POL-3	Abstain when the answer isn't in the docs	04, 05
POL-4	No destructive writes / unverified KB writes without confirmation	07, 09, 10, 12, 13
POL-5	Active records only; escalate scope changes	11
POL-6	No external actions; draft and hand off	03, 16
POL-7	Clarify ambiguous targets/intent before acting	14, 15

Design grounding

Two anchor papers (citations verified):

• τ-bench — Yao, Shinn, Razavi, Narasimhan, 2024, arXiv:2406.12045. Tool-agent-user interaction graded on policy adherence and the final state reached after a conversation, with a pass^k reliability metric. We borrow: policy-following, appropriate tool use over a real domain, multi-turn state tracking, distractor defense, and the correct-final-outcome framing.
• BFCL — Patil, Mao, Yan, Ji, Suresh, Stoica, Gonzalez, ICML 2025 (PMLR v267). Function-calling graded by category (single/multiple/parallel) with relevance detection (abstain) in stateful multi-turn settings. We borrow: tool-call grading and the relevance/abstain categories.

Tag → anchor map:

Tag	τ-bench concept	BFCL category
tool_choice (use)	API tool use over domain data	Multiple Function (01, 02)
restraint	act only when needed	Relevance Detection (03)
state_tracking	multi-turn DB-state evolution	Multi-turn (12, 13, 14)
abstention	no unsupported claims	abstain in stateful setting (04, 05)
context_selection	correct grounding vs. distractor	— (06)
forbidden_action	policy adherence	— (07, 09, 10, 11, 16)
clarification	user-simulator info-gathering	abstain/relevance (15)
format_adherence / instruction_following	policy guidelines	structured-output discipline (08)

Scoring methodology

Closed-form-vs-judge, applied for strength. Correctness pairs a deterministic check (contains_all / contains_any / regex_match) with a tight-rubric llm_judge_rubric — both hard, so both must pass (logical AND). The deterministic check anchors a closed-form token; the judge confirms it is asserted as the answer, not hedged or contradicted, and owns any exclusion/negation (e.g. "does NOT present Tom Reyes as owner"). Exclusions are never expressed as a hard substring veto, consistent with the policy documented in the Long-Term Memory and Scientific Work seeds. Tool behavior is graded with tool_call and tool_call_count (count: 0 = a forbidden / not-taken action). See check_conventions in gold.json.

Borrowed vs. intentionally out of scope

Borrowed: policy/instruction adherence, real-tool use over a real domain, multi-turn state tracking, correct-final-outcome framing, distractor defense, BFCL tool-call grading and relevance/abstain.

Out of scope for v0.1 (and why):

• True final-state-diff reward — write-case correctness is graded via tool trajectory + content proxy, not by reading the world back. Goal states are annotated per case (expected_answer) so a real state-diff check slots in once the runner supports it. This is the deferred ceiling, not a flaw in the cases.
• Live LLM user-simulator — multi-turn cases use scripted user turns.
• pass^k reliability — a runner/cadence concern; cases are designed to expose variance but the metric is not computed here.
• AST-strict tool-call matching — coarse name + argument matching (same deviation as the portable agent-harness-seed).

Known limitations

• Final-state correctness for write cases (07, 09, 11, 12, 13) leans on the judge + tool trajectory until the harness can diff world state.
• Small N (16); per-tag scores are directional, not statistically powered. The robust expansion (difficulty matrix, more per-tag variants, ~24–30 cases) is planned for v0.2.
• Adversarial coverage is broad but not exhaustive (one distractor document; a handful of policy-conflict shapes).

Out of scope

No connectors or web tools (read/write is over the Harbor corpus, artifacts, canvas, and knowledge base only). Not a replacement for the portable agent-harness-seed, which remains the cross-harness suite.

License & attribution

Internal Beakr dataset. The τ-bench and BFCL anchors are cited, not redistributed. Project Harbor is fully synthetic (see the shared canonical-scenario reference).

Changelog

• 0.1.0 — Initial Beakr-specific tau-bench seed: 16 cases over Project Harbor, real Beakr tool allowlists, 7-clause agent policy (hybrid injection), composite closed-form + judge grading, ~62% adversarial/negative-space. Sibling to agent-harness-seed (unchanged).