darkone01/contextual-evidence-grounded-text-reconstruction

Dataset Description

Overview

Synthetic Attested Operational Text Repair v4 is an original procedural fictional dataset for generating auditable document repairs under evidence tampering and revocation. The intended use is to evaluate whether a system distinguishes authenticated evidence from plausible forged capsules, follows scoped delegation chains, repairs only permitted document bytes, and produces verifiable support for both original and counterfactual repairs.

There are 4,800 cases in 400 families of 12 variants. Each case supplies an operational draft with an immutable archived quotation, 21 candidate source capsules, a signed evidence snapshot, a scoped authority graph and two independent source-removal queries. Eighteen capsules authenticate; three plausible forged candidates fail membership. No validity label is supplied as a feature. Each target contains a base and two counterfactual certificates.

This is simulated operational data and simulated signing identities. It does not represent a real organization, real delegation documents or a production cryptographic system. No downloaded corpus or model calls are used by the generator. Code, template text and documentation were AI-assisted; platform eligibility is a separate requirement.

File Structure

raw_dataset_v4.zip contains these seven creator/reviewer files at its root:

• records.csv: 4,800 labeled cases in UTF-8 CSV, with four string columns.
• generation_metadata.json: reproducibility, counts, scheme names and raw-file checksum; not a sample table.
• generate_raw.py: v4 corpus generator, snapshot signing and forged-candidate construction.
• synthetic_text.py: fictional equipment, person/location sampling and draft/fact rendering.
• policy_logic.py: creator authority, byte-span, minimal-replay and revocation semantics.
• attestation.py: signature, source-membership and certificate-binding implementation.
• case_logic.py: authentication-aware creator reference construction and validation.

attested_challenge_v4.zip is a separate complete creator package containing that raw ZIP, self-contained preparation/grading scripts, full descriptions, research text, mandatory.md, rubrics, a labeled public training example, validation results and agent-status instructions. Upload only the raw ZIP under Data Files.

Features

Raw column	Type	Format and meaning
id	string	Unique opaque item- identifier with 128 random bits; generated separately from semantic values
family	string	family-0000 through family-0399; 12 variants share equipment, a root source and a simulated signing anchor
input_text	string	Serialized JSON case: draft, byte intervals, policy metadata, public trust anchor, signed manifest, source capsules and two revocation queries
target_text	string	Serialized JSON program: base and two counterfactuals, each with record, edits, provenance and attestation

All required fields are populated. Raw IDs are unique. There is no template_group column.

Input JSON

Key	Type	Meaning / units
version	string	attested-repair-v4
ticket	string	Opaque identity of the document
draft	string	Original Unicode document with an archived quotation containing obsolete values
field_spans	object[string,array[int]]	Mutable status/location/owner intervals; zero-based half-open UTF-8 byte offsets
protected_spans	array[array[int]]	Immutable quotation intervals in UTF-8 byte units
trusted_roots	array[string]	Explicit trusted authority source IDs
cutoff	integer	Latest permitted fact time in relative episode minutes
sources	array[object]	21 candidate root/grant/fact/comment capsules in shuffled order
revocation_queries	array[string]	Two independently applied removals: one sampled fact and one sampled grant
trust_anchor	string	Simulated curator's 32-byte Ed25519 public key in lowercase hexadecimal
attestation	object	scheme string, signed manifest object and 64-byte signature encoded in lowercase hexadecimal

Each capsule contains id (string), kind (root/grant/fact/comment), text (string) and commitment (object). The commitment contains leaf_index (integer) and siblings (array of 32-byte hexadecimal SHA-256 hashes in bottom-up order). The complete capsule excluding commitment is the authenticated payload.

Roots add issuer (string), scope (array of field names), start/end (inclusive integer relative minutes). Grants additionally add parent and grantee (strings). Facts add issuer,authority,ticket (strings), time (integer relative minutes) and field_spans (object of UTF-8 byte intervals into their own text). Comments provide no authority privileges.

The manifest contains context_sha256 and source_root (32-byte lowercase hexadecimal strings) and tree_size (integer, 18). The context digest binds version, ticket, draft, mutable/protected spans, trusted roots, cutoff, queries and trust anchor. It contains no target or target hash. The signature authenticates that manifest. The source tree commits to authentic input payloads; it is not a commitment to reference outputs.

Target JSON

base is a view object; counterfactuals maps the exact two query IDs to view objects. Each view contains record (string), edits (array), provenance (object) and attestation (object).

An edit has field (string), start/end (integer draft byte offsets) and value (replacement string). A field proof has source (fact ID), start/end (integer source-text byte offsets) and path (ordered root-to-fact source IDs). Proofs are required for unchanged fields too.

Output attestation contains scheme,trust_anchor,manifest,signature,revoked,certificate_sha256,witnesses. revoked is null for base or the query ID for a counterfactual. Witnesses contain the complete payload, integer leaf_index and array of sibling hashes for exactly the sources used by that view's paths. The certificate digest binds the normalized repair to the signed manifest and revocation view; it is generated in the target, never disclosed in public inputs. The solver does not need a signing key.

Generation Process

The main PRNG seed is 20261012. Identifier sampling uses an independent stream seeded by XOR with 0x5A17B3. Forged-candidate and query sampling use another independent stream seeded by XOR with 0xA77E57. These are reproducible creator settings, not public solver features.

Each family samples a fictional equipment context and one of four document formats. Each variant samples old and candidate statuses from six statuses, owners from 576 fictional first/last-name combinations and locations from combinations of sites, buildings, rooms and bays. Curly quotations, em dashes and accented names exercise UTF-8 byte offsets. Fact text also uses four formats.

The authentic graph contains one root, seven grants, nine facts and one comment. Its valid narrow and complementary delegations require assembling the base result from multiple fact sources. Distractors include scope widening, expired authority, an unrelated ticket, cyclic authority, a future fact and instruction-like comments. Fact times are distinct integers sampled from 10 through 169; the cutoff is the eighth sorted time. A trusted-root baseline remains after either query.

Authentic sources are shuffled into an 18-leaf SHA-256 Merkle tree using RFC 6962 leaf/node domain separation and split shape. A family-specific simulated Ed25519 key is reproducibly derived in creator code and signs the context-bound manifest. Three candidates are then added: a tampered newest fact and a forged delegation/fact chain with copied witnesses for different payloads. They contain no attack labels. Source order and leaf order do not encode selected answers.

Queries are sampled from all candidate facts/grants before selecting the target; they include relevant, irrelevant and unauthenticated candidates. Among 9,600 query views, 3,131 change the repaired record in this release. All 4,800 base views combine multiple authentic fact sources. A controlled resolver with authenticity filtering omitted produced a different base record in all 4,800 cases; this control is not an agent score.

The generator validates every output's signature, source witnesses, authority paths, byte spans and exact minimal edit replay. No random label flips are introduced.

Notes

• Use matched v4 raw data, scripts and descriptions; earlier raw versions and predictions are incompatible.
• Preparation keeps complete families together and merges any shared roots, signing anchors or canonical semantic cases. The split is 3,840 training / 960 test rows and 320 / 80 source groups. Cryptographic encodings are excluded from semantic duplicate grouping so random-looking bytes cannot conceal duplicate source content.
• Public train columns are id,family,source_group,input_text,prediction. Public test columns are id,family,source_group,input_text. Private answers contain exactly id,prediction, preserving the prepared-integrity correction.
• Whole cases receive joint exact attested certificate credit. JSON formatting, edit order and output witness order are normalized; document bytes, evidence IDs/spans and path order are exact.
• Public inputs contain no private signing keys, output digest, held-out targets or target-validity flag. Full raw labels, seed-bearing metadata, generator code and reference construction remain creator/reviewer artifacts before closing.
• Signatures authenticate simulated snapshot bytes. They do not certify real-world truth, real institutional authority or an agent-generated output signature. Reproducible synthetic keys must not be reused for production.
• Short schema/prose spans can repeat. Complete-case/source separation does not prove absence of leakage or clear a platform rare-span warning.
• The creator-selected license is CC0-1.0; no external dataset was imported. Verify authorship, redistribution rights and platform eligibility before submission. No genuine external v4 hosting URL has been verified.

Reproducibility metadata

generation_metadata.json records version/seed, family and row counts, source/authentication counts, crypto scheme/hash names, changing-query count, multi-source count, authenticity-omission control, units, source type, external/model-call flags, synthetic-key flag, creator-selected license and records_sha256. Counts and seed are integers; flags are booleans; remaining values are strings. Metadata is creator provenance and is excluded from solver inputs.