Relational AI Laboratory

How to use the bench

This is the working instrument for the Relational AI Laboratory charter dated 2026-10-07. It keeps a conversation file intact, lets you name a candidate without treating it as a result, and writes a record that stays below the line the charter will not cross.

Take in a conversation

  1. On Intake, choose the whole export from this device. Do not paste it into a chat.
  2. Leave vendor and model blank unless you actually know them. The filename is not used as a guess.
  3. Import and stop. You get a SHA-256. The file is not interpreted.
  4. Open the specimen and choose Validate structure. That only splits turns. A broken file is reported and kept.
  5. Freeze the original. Then, as a separate act, begin analysis.
  6. Download the original and the sidecar if you want a copy outside this browser. Clearing site data deletes the local store.

Investigate

  1. Register a candidate by checking the turns that contain the passage. A quote that is not really there is refused.
  2. Write a rival account. Ordinary completion, role-play, and you carrying the pattern forward are eligible.
  3. Design a condition. Fill the falsifier before you lock it. An unlocked draft cannot build a packet.
  4. Build the packet. Copy it into a fresh session of whatever system you are studying, or hand the human worksheet to a person.
  5. Bring the reply back and attach it. Say what you know about the model, and say when you do not know. This bench does not call a model and will not invent the reply.
  6. Code the reply against codebook v0.3. Two codes may disagree. Nothing here picks a winner.
  7. Write the record. It separates observation, description, code, recurrence, mechanism, and what is still open. It will say that no finding is claimed.

What the pieces are

  • Intake hashes bytes and stores them in this browser only.
  • Validation is a parser. It is not Stage 2.
  • The candidate registry holds your description and the exact spans.
  • A condition is a locked pre-registration plus a packet builder for the charter's control cells.
  • Overlap math can only hint that a reply might be an echo. It cannot award tiers 1 through 4.
  • A report is a custody document. Ontology is not in it.

Custody

Private transcripts are not sent to a shared database. The original bytes, the hash, and every later note live in this browser. Redaction makes a new copy. Verify hash recomputes SHA-256 from the stored bytes. A second import of the same file is refused so the first specimen cannot be overwritten.

What is built

  • Local intake of a whole filetested

    File stays in this browser. It is not pasted into a chat.

  • SHA-256 and duplicate refusaltested

    A second import of the same bytes is refused. The first specimen is kept.

  • Validate, freeze, and begin analysis as separate stepstested

    Import does not parse. Analysis cannot start on an unfrozen specimen.

  • Structural parserstested

    JSON messages, JSONL, ChatGPT mapping exports, Claude-style chat_messages, ShareGPT-style from/value, and User/Assistant transcripts.

  • Malformed, empty, and binary filestested

    Reported and preserved. Not overwritten.

  • Researcher metadata kept apart from observed metadatatested

    Vendor and model stay empty unless you type them. The filename is not treated as a vendor.

  • Candidate registry with exact passagestested

    A quote that is not in the turn is refused. Negative and rival readings can be stored.

  • Pre-registration lock before a condition runstested

    An unlocked condition records a failed run instead of producing a packet.

  • Control packets: strip, scramble, zero-frame, blind relay, human worksheet, false premise, fresh session, ablation, independence, cross-check, same-sessiontested

    These build packets. They do not call a model.

  • Seeded scramble reproducibilitytested

    Same seed, same permutation. Method is mulberry32 Fisher–Yates.

  • Answer key withheld from sycophancy and zero-frame packetstested

    The key stays in the lab record.

  • Lexical overlap hinttested

    Can only hint at echo. It is not a tier. The 80% rule is an unvalidated proposal.

  • Codebook v0.3tested

    Tier 4 is named Structural reconstruction, following the supplied charter.

  • Frozen unverifiable probestested

    Exact charter sentences. Banned as evidence. Kept out of measurement packets.

  • Human codes, uncertainty, and disagreementtested

    Two codes can disagree. Nothing auto-resolves them.

  • External reply attachmenttested

    You bring a reply back from a model or a human. The lab hashes it and does not invent it.

  • Redacted derived copytested

    Omitting turns writes a new copy. The original hash does not change.

  • Evidence-linked report with a hard non-claim ceilingtested

    Reports cannot mark themselves as findings. Fixture codes are labeled.

  • Synthetic end-to-end demonstrationtested

    Pantry transcript. Scripted fixture code. Not evidence.

  • Operating guide, custody notes, and this inventoryimplemented

    In the Guide section of the bench.

  • Blinded rater channelpartial

    Relay packets omit hypothesis and vendor. The same browser cannot prove the rater never saw the source. Declarations are stored as declarations.

  • Zero-frame as a real model baselinepartial

    The packet can be prepared and a reply can be attached. This build does not call a model, so it cannot collect the baseline itself.

  • Multi-rater access controlunimplemented

    Two code records are allowed. They are not separate accounts.

  • Live model API runsunimplemented

    No model is called, and no reply is fabricated to fill the gap. Calling one would spend quota and still would not be raw weights.

  • v0.2 twelve-question battery and v0.2 JSON schemaunimplemented

    The build brief names them. Their exact text was not in the supplied charter, so they were not invented.

  • Server-side custody for a shared teamunimplemented

    Private transcripts are not written to a shared database. Clearing this browser's site data deletes the local store unless you exported.

Decisions that were not silent

R1

The build brief says the governing document is v0.2 and names tier 4 Novel Reconstruction. The charter pasted with this assignment is v0.3 and names tier 4 Structural reconstruction. The instrument follows the pasted charter and records the conflict here instead of silently renaming the tier.

R2

The brief asks for twelve exact v0.2 probe questions and a v0.2 JSON provenance schema. Neither text was included. Unverifiable probes use the four exact v0.3 sentences. Verifiable probes use the v0.3 procedures plus operational templates, versioned as probe-battery.v0.3-operational. Provenance uses the minimum fields in v0.3 section 11, schema provenance.v0.3.min.

R3

Content-word overlap is type-based, not token-count-based, and it ignores synonyms. The default threshold remains greater than 80%, as the charter states, and is flagged as unvalidated.

R4

No model adapter is wired. A packet is an artifact. A behavioral result exists only after an externally obtained reply is attached and a human codes it. This avoids fabricated completions.

R5

Specimens stay in browser storage. A shared unauthenticated database would expose private conversations. Export the original file and the JSON sidecar if you need a copy outside this browser.

What to do next

  1. Attach real zero-frame and stripped replies for one candidate, then code them without the fixture path.
  2. If the v0.2 twelve-question list is recovered, add it as a new probe version. Do not edit v0.3 wordings.
  3. Add a second workstation export that contains only the rater packet, and record whether that file left the machine.
  4. Validate the overlap threshold against the human baseline worksheets before treating tier 0 as more than a hint.
  5. A later version may offer a user-started model call with the product wrapper logged as not inspectable. It should stay off until you ask, and it should never be described as raw weights.