Verified Transcript Cache Protocol

Reference implementation: donotreadagain (dnr)

Read once, never again. A small protocol for faithful, signed transcripts of expensive-to-parse source files, so AI agents stop re-OCR/re-parsing the same PDF, image, scan, spreadsheet, or audio every time.

· status: v0.2 early release

Tell your agent:

Use dnr for this folder.

That one line is the adoption path. The agent fetches SKILL.md, checks cached transcripts before parsing files, and records any expensive read it had to do anyway.

This repo contains the Verified Transcript Cache Protocol, its record schema and conformance vectors under spec/, and dnr, the reference CLI implementation. Harnesses can call the CLI today or implement compatible records natively later.

The problem

AI agents re-parse the same file every time they touch it — re-OCR a scan, re-run vision on a screenshot, re-transcribe an audio clip, re-extract a PDF. It's slow, it burns tokens and model calls, and it's non-deterministic. In repeat-access corpora (legal, research, compliance) the same documents get read dozens of times; with multi-agent setups every agent re-parses independently. The waste compounds exactly where it hurts.

The idea

The protocol defines a cache/trust/index layer for expensive source-file reads. A local extractor, local ASR model, or the calling AI agent reads a file once; dnr stores the resulting transcript + structured metadata as a signed JSON record in a folder .dnr.db by default, so original files stay byte-identical. Any agent that opens it later reads the cached transcript instead of re-parsing. A per-folder SQLite + FTS5 index makes a whole folder searchable without opening anything. Portable in-file records remain available with explicit --embed, but file modification is never the default.

The second view is the win:

…and the cache is trustworthy: a record is used only if it's signed by a trusted key and its content_hash still matches the file, so "fast" never means "stale or forged."

Protocol contract

For agents and harnesses, the protocol is a small pre-read loop:

1. Known file: run dnr read <file> before parsing it. If stdout has text, use it and do not re-read.
2. Miss: if the task still needs the file, parse/look/listen once, then cache that result with dnr ingest or dnr record.
3. Folder question: run dnr index <folder>, then dnr query <folder> ... before opening files.
4. Folder preparation: use dnr status <folder> --pending; run dnr backfill <folder> only when the user wants a folder pass.
5. Boundary: never bulk-transcribe just because files are pending; only cache work the task actually needs.

Harness maintainers can copy the integration contract and reference adapters from HARNESS.md, or implement the protocol directly from PROTOCOL.md.

Demo

$ dnr ingest contract.pdf            # transcribe once → sign → cache in .dnr.db
ingested contract.pdf  [db-only (index)]
  method=text-extract transcriber=pymupdf
  signed key_id=ce6d170a497238f7

$ dnr read contract.pdf              # later (or from any agent): verified cache hit — no re-parsing
LOAN AGREEMENT
Lender: Acme Capital LLC
Borrower: Jordan Smith
Principal: USD 1,200,000
...

$ dnr index ./contracts
$ dnr query ./contracts --match damages --context 40    # search a whole folder, no files opened
contract.pdf
    … Principal: USD 1,200,000  Maturity: 2026-12-31  Damages clause: section 7.

The transcript lives in the folder's .dnr.db by default, so contract.pdf itself stays untouched. If you explicitly need the cache to travel inside the file, add --embed.

Quickstart

Recommended install:

pipx install donotreadagain
dnr --version

# one-off/fallback when installing is not available:
uvx --from donotreadagain dnr <cmd>

# audio ASR:
pipx inject donotreadagain faster-whisper   # ffmpeg may also be needed for decoding

dnr ingest report.pdf              # extract once (local) → sign → store in .dnr.db
dnr backfill ./case-folder         # folder pass: local-provider files now, agent/vision worklist after
dnr read   report.pdf              # print the cached transcript (verified), or fall back
dnr index  ./case-folder           # build .dnr.db
dnr status ./case-folder --pending # honest usable/pending/repair coverage
dnr query  ./case-folder --match "손해배상" --tag 가압류 --since 2025-01-01

For a scan / image / anything you must look at, the agent transcribes it and records the result:

dnr record scan.png --transcript-file t.md --method vision --transcriber <your-model>

How it fits together

File = canonical truth                    Index .dnr.db = default cache
┌────────────────────────────┐  harvest  ┌────────────────────────────┐
│  original file bytes         │ ───────▶  │  signed record + FTS5 search │
│  content_hash · transcript  │           │  path · tags · transcript … │
│  provenance · fields · sig  │           └────────────────────────────┘
└────────────────────────────┘                 ▲ query via sqlite3 — dnr CLI optional for reads
   ▲ transcribe once · sign · store db-only (expensive)

Where the record lives (no sidecar files):

• db-only by default in the folder's .dnr.db, so original files stay byte-identical. If the source file changes, the stale record is removed and the file must be re-ingested/re-recorded.
• Optional in-file with explicit --embed for formats with a metadata slot — PDF→XMP, MP3→ID3, M4A/MP4/MOV→MP4 freeform atom, FLAC/OGG/OPUS→Vorbis/Opus comments, PNG→iTXt, JPEG→APP segment. This makes the transcript travel with the file, but it rewrites file bytes and is never the default.
• Nothing for already-readable text (.txt/.md/.csv) — an agent just reads it.

Current format support:

Using it

• Read (consumer): dnr read <file> returns the cached transcript only if it's present, trusted, and still matches (self-validating — a changed file silently misses). No dnr tool? An agent can read .dnr.db directly with ambient sqlite3 (the db's _dnr_readme table self-describes).
• Transcribe (producer): dnr ingest (local: PyMuPDF→pypdf / python-docx / openpyxl / optional faster-whisper) or dnr record (agent supplies a vision/OCR transcript). dnr is an opportunistic cache: do this when the current task already requires reading/parsing that file, not just because a folder has pending files. If a needed file's cached transcript is empty/garbled/unusable, repair that file immediately; ask only before expanding into whole-folder OCR/searchability work.
• Backfill a folder: dnr backfill <folder> (also dnr ingest <folder>) ingests locally-processable files in one pass, skips already-readable text, and prints a worklist for images/scans/videos or low-quality results that need agent/vision repair.
• Query a folder: dnr query <folder> combines --match (FTS, Korean/CJK ok) ∩ --tag a,b ∩ --since/--until ∩ restricted --where over fixed columns; plus --any (OR sweep), --dedup, --context (KWIC), --format json. Save composed queries with --save/--use; accumulate labels with dnr tag.
• Agents onboard locally: point an agent at a dnr folder and it fetches SKILL.md once for that task/folder. Recommended install is pipx install donotreadagain; one-off fallback is uvx --from donotreadagain dnr .... dnr init just ensures a signing key by default; to persist the bootstrap in a project instruction file, run dnr init --agent-file AGENTS.md (or --agent-file CLAUDE.md). Harness authors can integrate the same read-through cache hook with HARNESS.md.

Design principles

• Protocol first, CLI as proof. The Verified Transcript Cache Protocol is the portable contract; dnr is the reference implementation that proves it works and gives harnesses an optional hook today.
• Original files stay untouched by default. The default write path is .dnr.db; in-file records require explicit --embed.
• Not a general knowledge format. dnr stores faithful transcripts tied to original files. Curated concepts, runbooks, summaries, and knowledge-base pages belong in higher-level docs; dnr can feed those systems, but it is not trying to replace them.
• dnr is the deterministic substrate; the agent is the intelligence. dnr does verifiable primitives (hash, sign, full-text/structured query); it never infers metadata (dates, parties, topics) or does fuzzy semantic search — that's the agent's job. Set metadata explicitly with dnr tag / dnr date.
• File = truth, index = regenerable cache. Delete .dnr.db and rebuild it from the files anytime.
• Transcriber-agnostic. dnr ships a contract (the verbatim guide) + a trust layer, not a model. Fidelity is the transcriber's; provenance is recorded so a consumer can apply its own quality policy (trusted ≠ faithful).

Status & honest limits

v0.2 early release. Published on PyPI as donotreadagain; the recommended path is pipx install donotreadagain. uvx remains a one-off/fallback path, but it still requires uv and can be slower than a normal install for repeated use. Standalone binaries remain a future packaging option for Python-less environments. Works today for repeat-access corpora; validated by real-corpus dogfooding. Known limits we're explicit about:

• Adoption is the real lever. The value compounds when agents know dnr (a skill, eventually native support) — not from the tool alone.
• trusted ≠ faithful. A signature proves who made it + that it matches the file, not that the transcription is accurate. Low-quality/garbled transcripts are flagged (dnr status), not silently trusted.
• Coverage is still growing. PDF/PNG/JPEG/HEIC/DOCX/XLSX and common audio containers are useful today; OOXML carriers, more video containers, pre-query auto-scan, and larger-corpus concurrency are still roadmap work.
• Benchmarks are early. The README numbers are illustrative dogfood timings; see BENCHMARKS.md and experiments/content-hash-invariance for the current proof/measurement status. A broader latency/token benchmark remains a release-readiness item.
• Python packaging is the product path. Use pipx for the cleanest install; uvx remains the one-off/fallback route.

See PROTOCOL.md (protocol contract) · spec/ (schema and vectors) · HARNESS.md (harness integration) · vision.md (design) · SECURITY.md (threat model) · qna.md (settled design decisions) · MILESTONES.md (roadmap).

Development

git clone https://github.com/melodysdreamj/donotreadagain
cd donotreadagain
python -m venv .venv && . .venv/bin/activate
pip install -e ".[dev]"
pytest                       # the suite is green and fast

Contributions welcome — see CONTRIBUTING.md. Release automation is documented in RELEASING.md.

License

MIT © 2026 june lee