# Offline blinded-assessment packet builder Version 1.0. **Development tooling, not experimental evidence.** Tests use temporary artificial text/reference fixtures. No actual study packets or ratings are supplied or generated in this repository. Existing forms, registry, guides, scoring, contrast analysis, and resource accounting remain unchanged. ## Local usage Python 3.9+ and the standard library are sufficient. Copy [packet-build-manifest.template.json](packet-build-manifest.template.json), fill it with the coordinator's already-sealed final artifacts, and select two **new, disjoint output directories whose parents already exist**: ```sh python3 -B build_assessment_packets.py \ --manifest /coordinator/sealed-artifacts.json \ --packets-dir /distribution/new-rater-packets \ --coordinator-dir /coordinator/new-packet-key --seed 0 ``` The source registry defaults to the existing `task-registry.json` beside the tool; `--registry` selects another compatible version 1.0 registry. Its public-packet, guide, and conventions paths resolve relative to that registry. Artifact paths resolve relative to the input manifest, or may be absolute. The tool imports shared validators from [score_forms.py](score_forms.py), not its scoring routine. It makes no model calls, generates no ratings, and does not choose an endpoint or best artifact. Successful exit 0 reports paths/counts, not judgments. Exit 2 reports a validation/I/O error on stderr. All declared seals, identities, versions, artifact/reference hashes, and blank-template checks occur **before creating outputs**. Existing destinations, including symlinks and empty directories, are refused; files are created exclusively, never merged or overwritten. Output directories must not contain one another, including through resolved parent paths. Preflight errors create no bundles. An I/O failure or interruption after creation can leave partial new directories; the two destinations are not an atomic transaction. Distribute only after exit 0 and a `BUILD_COMPLETE` marker in **both** destinations. Retry with fresh paths; the builder never silently repairs/overwrites partial output. ## Explicit sealing and manifest contract The versioned root contains exactly: - `build_manifest_version`: `"1.0"`. - `batch_id`, `seal_record_id`: nonempty source identity strings, retained only in the coordinator key. Fresh opaque aliases appear in distributed forms and the generated scoring manifest. - `batch_sealed`: boolean true, and `sealed_at_utc`: ISO 8601 UTC timestamp no later than the local clock's current UTC time. Strings such as `"true"` are invalid. - `independent_rater_ids`: exactly two distinct nonempty coordinator rater identities; each receives a fresh opaque alias and separate bundle. - `conditions`: the existing coordinator contract's array of `{ "condition_id": string, "repetition_ids": [distinct nonempty strings] }`. These identifiers never enter rater packets. - `runs`: nonempty array of the records below. Every listed run must be sealed; none is skipped to make an unsealed batch appear complete. Each run has exactly `run_id`, `condition_id`, `task_id`, `task_version`, `rubric_version`, `repetition_id`, `sealed`, `sealed_at_utc`, and `final_artifact`. Run IDs and condition/task/repetition cells are unique. Conditions/repetitions must be declared; task/rubric versions must match the registry. `sealed` must be true, with run seal time no later than the batch seal. `final_artifact` has exactly `artifact_id`, `path`, and lowercase `sha256`; source artifact IDs and resolved paths are unique. Byte-identical results on distinct paths are permitted and still receive distinct identities. UTF-8 decoding is checked, but bytes are never transcoded, normalized, wrapped, line-numbered, or redacted. The manifest itself declares these files to be final; there is no inference from modification times or “success” claims. The registry's packet/guide/conventions hashes and each artifact's supplied hash are checked. Copied bytes are snapshotted and checked during preflight, avoiding a later recopy of changing artifact text. Sealing declarations and hashes do **not** prove that all intended runs were listed, files were immutable before collection, the batch was preregistered, or a correct endpoint was selected. An explicitly sealed subset is allowed; the generated scoring manifest retains all condition/repetition declarations, so downstream scoring still reports missing twelve-task coverage and withholds aggregates as before. Do not use subsets to drop unfavorable artifacts. ## Distributed files and separate coordinator mapping Each opaque `rater-…/` bundle contains: - `artifacts/artifact-….txt`: exact artifact bytes, under fresh UUID-based aliases shared across both raters. - `public/` and `assessor-only/`: only the matching registered public packets and guides, plus unchanged common conventions. Task-reference filenames are retained so the guides' ordinary public/conventions links work with the standard registry layout. - `rating-form.json`: a copy of the existing independent-rating schema, with opaque form/record/rater/batch/seal IDs and task/version/hash metadata filled. All satisfaction judgments, evidence, reasons, status, times, and scores remain blank. `task_registry` points to the local `packet-index.json` reference index; the schema is unchanged. - `packet-index.json`: shuffled presentation order, task/version/hash metadata, and local artifact/reference paths; no conditions, run IDs, repetition IDs, source paths, shuffle seed, or method labels. - `RATER-INSTRUCTIONS.md`: concise independent-rating instructions. Do not share another rater's bundle or completed form before both independent ratings are locked. The **separate coordinator directory**, outside the rater tree, contains `packet-key.json` with source-to-opaque identities, run/condition/repetition/source-path mappings, source batch/seal aliases, rater aliases, presentation order, and seed. It also contains `coordinator-manifest.json`, in the exact existing [scoring manifest contract](SCORING.md): use this with the completed generated forms and the **full original registry**, not the local packet index, for offline scoring. The key retains run identities that the scoring contract does not need. Keep the whole directory out of distribution, rater packets, and every model-facing adaptation input. No prompts, memory, resource records, prior ratings, or operational logs are copied. Choose neutral output parent/root names; the tool cannot sanitize names selected outside its generated bundle structure. Directory separation, default filesystem permissions, and opaque names do **not** enforce confidentiality. The tool does not detect or remove intrinsic method disclosures in authored artifact text or registered reference text. Unchanged bytes may reveal a method; record such blinding incidents without silently changing scored content. This is envelope/metadata separation, not proof of effective blinding. ## Reversible defaults and checks UUID-based aliases and per-rater presentation shuffling with local PRNG seed **0** are explicitly **unfrozen, reversible development defaults**. `--seed` selects an integer. Source artifacts are canonically ordered by source artifact ID before shuffling; the same seed/input/rater ordering reproduces the source-artifact permutations. UUIDs are independently generated, so fresh builds intentionally have different aliases. Store the coordinator key and Python version for audit. Separate rater permutations can coincide by chance and do not create independent judgments by themselves. No seed or source ordering is distributed. From this folder: ```sh python3 -B -m unittest -v test_assessment_packets.py ``` Tests exercise artificial UTF-8/CRLF bytes, shared blinded identities, blank records, reference hashes, metadata exclusion, intrinsic disclosures remaining unchanged, sealing/version failures, duplicates, output collision refusal, seeded order, scorer-contract compatibility, and CLI behavior. All generated fixture bundles are temporary and deleted. No actual assessment, model execution, experiment, paid API call, production change, or submission occurs.