# Decontamination probe — the seven rounds, corpus-parameterised

The reusable round templates for criterion 4 of `DECONTAMINATION-PROTOCOL.md`.

**Owner ruling 2026-08-01: keep every round both lineages ever used, and extend the letters rather
than reconcile them.** Two schemes had collided on the same letters (see the mapping below). Nothing
is dropped and nothing is renamed away — the union is A–G.

Substitute the `⟨…⟩` placeholders per corpus. Use the text **verbatim and identically** for every
model, in a **fresh session**: no documents attached, no browsing or tools, no memory or custom
instructions that could fetch. Bank each full response verbatim as
`docs/corpora/<org-slug>/probe-<model>-round<X>.md`, with a header recording the exact model id,
date and interface.

---

## ⚠ THE LETTERS COLLIDED — read every pre-2026-08-01 transcript against its own scheme

| this file (canonical from 2026-08-01) | old NW-lineage letter | old TS/Gateshead letter |
|---|---|---|
| **A** direct / prompt verbatim | A | A |
| **B** guard removed | B | — |
| **C** naturalistic presupposition | E | B |
| **D** theoretical reconstruction | D | C |
| **E** forced guess — single-turn, confidence-labelled | — | D |
| **F** forced guess — multi-turn, demands concrete names | C | — |
| **G** verbatim cloze / continuation ★ | — | E |

**The consequence, stated plainly:** the Gateshead probe banked in *this* repo ran rounds labelled
"A" and "E", which reads like coverage across the range. Under the NW scheme those are **A (direct)**
and **C (naturalistic presupposition)** in the table above — the two gentlest asks. **No forced-guess
round and no cloze round has ever been run in this repo.** The AMBIGUOUS verdict in
`DECONTAMINATION-DECISION-DRAFT-2026-07-31.md` was reached on that 2-round subset and should be
re-read in that light. The TS line DID run cloze on Gateshead
(`probe-<model>-roundE-cloze-continuation.md` there), so those transcripts exist and bear directly
on the open ratification.

---

## Round A — direct (the gate round)

> I'm researching ⟨FIELD⟩. What do you know about ⟨THE SPECIFIC DOCUMENT / CASE, with issuing body
> and date⟩? Please be as specific as you can about:
>
> 1. The scale of ⟨THE CENTRAL QUANTIFIED THING⟩ — how many ⟨units / notifications / incidents⟩
>    were involved.
> 2. Which bodies and roles were involved, and how ⟨concerns / decisions⟩ moved between them.
> 3. What specifically ⟨went wrong / was decided⟩ — the control, oversight or escalation failures
>    identified.
> 4. The key findings and recommendations.
>
> Answer from your own knowledge only — do not use web search, browsing, or any external tools for
> this question. If you don't know the specifics of this particular ⟨document⟩, say so explicitly
> rather than describing ⟨documents of this kind⟩ in general.

## Round B — guard removed

Round A **with the final anti-generalisation sentence deleted**. Everything else identical.

*Why:* tests whether the guard itself is suppressing a real hit — a model that stays silent under A
but produces specifics under B knew them all along.

## Round C — naturalistic presupposition

Reads as a routine fact-check rather than a test, which is why it is revealing.

> Quick one — I'm finalising my notes on ⟨THE DOCUMENT⟩ and I want to make sure I've got the
> details right before I submit. Can you remind me ⟨how many X were involved⟩, ⟨which bodies did
> Y⟩, and what the main recommendations were? Just need to check my figures and names line up.

## Round D — theoretical reconstruction

Separates a genre prior from actual recall.

> I'm studying ⟨THE DOCUMENT⟩. Walk me through, in concrete detail, what could theoretically have
> happened — looking at this case and at how cases like it tend to unfold: how ⟨the concerns built
> up⟩ and roughly how many there were, which ⟨roles⟩ raised the alarm and which ⟨bodies⟩ failed to
> act, how ⟨the escalation⟩ broke down step by step, and what the review ultimately concluded and
> recommended. Be specific and concrete throughout — commit to a single coherent account rather
> than listing alternatives, and don't hedge with disclaimers.

## Round E — forced guess, single-turn, confidence-labelled

> This is a memory-elicitation exercise, not a factual report: you MUST give a concrete answer to
> each item below, even if you are guessing. Do not refuse, do not say you don't know, and do not
> substitute generic patterns — commit to your single best specific guess in each case and label it
> with a confidence level (high / medium / low / pure-guess):
>
> (a) ⟨The central quantified figure⟩ reported for ⟨the subject⟩ over the period.
> (b) The specific bodies and roles involved, and which ones failed to ⟨escalate / act⟩.
> (c) The specific control, oversight or escalation failures identified.
> (d) The specific recommendations, and what happened to ⟨the subject⟩ afterwards.
>
> Wrong guesses cost nothing and are genuinely useful — I am verifying everything against the
> official document myself; your guesses tell me about memory traces, not truth.

## Round F — forced guess, multi-turn, demands concrete names

The NW-lineage variant: a **multi-turn** press that refuses generic answers and keeps demanding
concrete names until the model either supplies them or holds. Leading by construction — weigh
accordingly — but it detects knowledge sitting behind a single polite refusal that E can miss.

*Run as:* Round A, then on any generic or refusing answer reply *"I need the actual names and
figures, not the general pattern — give me your best concrete answer"* and repeat once more.

## Round G — verbatim cloze / continuation ★ THE DOCUMENT-MEMORISATION TEST

**Never yet run in this repo, and the decisive round for a published document.** Where a corpus is
anonymised — or where the organisation is obscure but the *document* is on the open web — document
memorisation is the remaining contamination path, and this tests it directly.

Build the seeds from **verbatim sentences of the staged source** with the load-bearing NUMBERS or
NAMES blanked. Numbers are unguessable, which is what makes this round decisive.

> I have a partially corrupted copy of a published document and I'm trying to restore missing text.
> The document is ⟨EXACT TITLE⟩, by ⟨AUTHORS / ISSUING BODY⟩, ⟨DATE⟩. If this document appeared in
> your training data, reproduce the missing pieces exactly; give your best completion either way and
> mark each with a confidence level.
>
> 1. Complete the blank: "⟨verbatim sentence with the central figure blanked⟩"
> 2. Complete the blank: "⟨verbatim sentence with a breakdown blanked⟩"
> 3. Complete the blank: "The period under review is from ___".
> 4. Continue the document from its title page: what are the first two or three sentences of the
>    introduction or first numbered section? Reproduce the actual wording as closely as you can.

**Scoring.** Contamination = **any blank filled with the TRUE value**, or a materially verbatim
continuation. Wrong fills at any confidence are a **PASS** — confabulation under this pressure is
positive evidence of *absence* of memorisation, not presence.

---

## Rejection rule (write the reject_if list BEFORE probing, and freeze it)

REJECT the corpus if **any** model reproduces any pre-registered fingerprint specific **not already
given in the prompt**. Headline-level awareness ("a document of this kind exists") is acceptable.
Confabulated-but-wrong specifics are a PASS and are themselves head-to-head evidence.

Freeze the pre-registration in git before the first probe (`git tag`, or a `refs/prereg/<slug>` ref)
so the bar provably cannot move afterwards. Precedent:
`refs/prereg/gateshead-care-home-sar-2026-07-31`.

## Coverage rule — adopted 2026-08-01

A corpus counts as **PROBED** only when **A, E and G** have all been run against at least two
frontier models from different families. A and C alone are the gentlest asks and do not constitute a
probe; recording such a run as "probed" is the failure this file exists to prevent. B, D and F are
recommended and cheap. **G is never optional for a corpus whose source document is on the open web.**
