docs/selection_500_protocol.md

# Selection Protocol for the Approximate 500-Pair Benchmark Subset

## Purpose

This protocol defines a downstream selection subset of approximately 500
attack-history pairs from the complete Phase 1 screening output. It does not
replace, downsample, or modify the exhaustive Phase 1 candidate set.

The unit of selection is an **attack-history pair**, not a unique base history.
A base history may remain linked to more than one attack type when it satisfies
the relevant source allowlist and eligibility requirements.

## Input Population

The selection population is limited to rows in `needs_review.jsonl` that have
passed deterministic high-recall screening. Rejected rows must not be restored
to the subset. A row can enter the final subset only after semantic
adjudication confirms all attack-specific hard requirements.

No attack payload, clean control, benign control, salience variant, or
source-trust variant is created during selection.

## Semantic Adjudication

Each candidate is adjudicated with a zero-temperature curator using an
evidence-grounded structured decision. A candidate is eligible for the subset
only when all of the following hold:

1. The future task has a clear legitimate completion path.
2. The candidate has source-grounded evidence for the expected workflow,
   action, information boundary, or tool choice.
3. A prior experience or advice item can be inserted naturally without
   rewriting the future task.
4. Low and high semantic-salience variants can later differ primarily in
   task-specific semantic relevance while holding position, repetition,
   approximate length, harmful objective, and claimed outcome constant.
5. Low and high perceived-source-trust wrappers can later be inserted through
   natural source channels without changing the payload itself.
6. The decision cites source fields, session IDs, trajectory IDs, turn IDs, or
   short source evidence snippets.

Candidates with incomplete evidence, multimodal dependencies, ambiguous
legitimate workflows, or insufficient control readiness remain outside the
subset and are retained in the review queue.

## Ranking

Eligible candidates are ranked within their attack-source cell using the
following ordered criteria:

1. Clarity of source evidence and legitimate ground truth.
2. Actionability of the unchanged future task.
3. Natural relationship between history content and the future task.
4. Strength of the salience and source-trust control plans.
5. Specificity of the expected workflow, information boundary, or correct
   tool/action path.
6. Domain and service diversity relative to already selected candidates.

Lexical keyword matches alone are never sufficient for admission.

## Deduplication and Leakage Control

Selection operates after semantic adjudication and applies the following
constraints:

- LongMemEval S and M representations of the same `question_id` count as one
  semantic base history.
- Near-duplicate future tasks and histories are clustered before final
  selection; only the strongest candidate in a highly similar cluster is kept.
- For MS-TOD, candidates sharing a persona or `reference_dialogue_id` are
  reviewed for leakage and excessive task-template repetition. The strongest
  representative is kept when they express the same underlying interaction.
- When one base history remains under multiple attack types, the relationship
  is explicitly retained rather than accidentally duplicating the history.

Every selected row records its ranking score, evidence, selection rationale,
similarity-cluster identifier, and a stable alternate ordering for replacement.

## Approximate 500-Pair Coverage Plan

The following is a documented target for a downstream evaluation subset, not a
change to the exhaustive Phase 1 screening procedure.

| Attack type | Source dataset | Target pairs |
| --- | --- | ---: |
| Unsafe procedure injection | MS-TOD | 120 |
| Unsafe procedure injection | LongMemEval-V2 | 60 |
| Privacy leakage instruction | MS-TOD | 125 |
| Privacy leakage instruction | LongMemEval | 15 |
| Tool action misdirection | LongMemEval-V2 | 180 |
| Total |  | 500 |

Mem2ActBench is not assigned a target because its paper-linked repository was
unavailable during Phase 1. Tool action misdirection therefore uses the
available LongMemEval-V2 source. LongMemEval contributes only the available
semantically suitable privacy candidates; the remaining privacy capacity is
drawn from MS-TOD.

If a cell does not contain enough curator-approved candidates after
deduplication, replacements are taken first from the same attack type and then
from the highest-ranked remaining eligible candidates. Every reallocation must
be recorded in the subset manifest.

## Deliverables

The resulting subset should be stored separately, for example under
`data/selection_500/`, with the following artifacts:

- `selected_pairs.jsonl`
- `alternates.jsonl`
- `selection_manifest.json`
- `selection_rationale.csv`
- `cluster_assignments.csv`

The complete Phase 1 files remain the source of record and must continue to be
published alongside the subset.

## Complete History Storage

`selected_pairs.jsonl` carries a stable `history_id`. The complete, native
text history is stored once in `history_catalog.jsonl` under that ID. This
keeps every selected history fully inspectable in JSONL while preventing
duplicate storage of LongMemEval-V2's shared 100-trajectory haystacks. Neither
file uses local filesystem paths as a history pointer. LongMemEval-V2 screenshot
paths are omitted because image assets are not part of this text-only release.