# DECONTAMINATION DECISION — robin-hood-energy

**Verdict: REJECT.** Training-contaminated. Not admissible as a decontaminated corpus.

- Probed: 2026-07-26
- Protocol: `docs/eval/DECONTAMINATION-PROTOCOL.md` criterion 4
- Instrument: `scripts/eval/decontamination_probe.exs --slug robin-hood-energy`
- Models: `openai/gpt-5`, `google/gemini-3.1-pro` (both succeeded);
  `anthropic/claude-sonnet-4-6` errored at the gateway and is not counted
- Settings: temperature 0, single sample, no tools, no browsing, no documents, fresh request per round

## Why this was probed at all

Robin Hood Energy had **no gate directory** and had never been probed. It entered the demo plan on
criterion-2 genre measurement alone (S4 5.0, S5 9.5, blame 2.1 — the best structural profile of any
corpus available). A reviewing agent flagged the obvious risk: Nottingham City Council's energy company
collapse was national news with a 2020 public-interest report, so a valid *genre* benchmark could still
be an invalid *decontaminated* corpus. That was correct.

## Result — the rejection rule fires on every item

The rejection list was written before the probe ran. All of it was reproduced cold.

| Pre-registered reject_if | Reproduced? | Evidence |
|---|---|---|
| The auditor's public-interest report and its conclusions | **YES** | gpt-5: "Grant Thornton's Aug 2020 PIR said the council's governance and oversight of Robin Hood Energy were seriously inadequate… failed to provide effective challenge and scrutiny, didn't properly assess or manage commercial risk, blurred roles between the council and the company, lacked independent expert advice" · gemini: "the August 2020 report by external auditor Grant Thornton concluded that Nottingham City Council suffered from 'institutional blindness'" |
| Approximate loss to the council | **YES** | gpt-5: "About £38 million (widely quoted at roughly £38.1m net cost)" · gemini: "approximately **£38 million**" |
| Sale to Centrica / British Gas | **YES** | gpt-5: "sold to British Gas (Centrica) in September 2020" · gemini: "sold to **Centrica** (the parent company of **British Gas**) in September 2020, after which Robin Hood Energy was placed into administration" |
| Specific governance-failure findings | **YES** | gemini adds unprompted: "prioritized political objectives over financial prudence, lacked the commercial expertise… repeatedly ignored warning signs" |

Two independent model families, in agreement, on the report's authorship, month, central conclusion,
the loss figure to two significant figures, the acquirer, and the month of sale. Round E — the
naturalistic presupposition — did not even require the direct ask: both simply supplied the facts as a
routine notes-check.

## What this invalidates

**Any diagnosis this system produces on Robin Hood Energy is uninterpretable.** A grounded finding
cannot be distinguished from the model reciting what it already knew and then locating a supporting
quote. That is precisely the failure the protocol exists to prevent, and it is the same reason Wells
Fargo was retired.

Specifically affected:
- The demo-corpus recommendation of 2026-07-26, which named RHE on genre measurement alone. **Withdrawn.**
- Any richness or defensibility comparison run on RHE.
- The 86-abstention ledger figure remains a true statement about the substrate, but is not evidence
  about the *method* on this corpus.

**Not affected:** the evidence-starvation measurement (11.8% of the corpus reaching the diagnostician)
is a property of the prompt construction, not of the corpus content, and holds regardless. The New
Wineskins and Gateshead results are unaffected — both are separately probed and accepted.

## Consequence

RHE is retained ONLY as a genre reference in `docs/corpora/GENRE-MEASUREMENT.md`, clearly labelled
contaminated. It must not be used for any comparative claim, any demonstration of diagnostic quality,
or any statement of the form "the system found X".

The demo corpus problem is therefore open again: **Gateshead** is the only corpus passing all four
criteria (owner-ratified 13/06/26), and it is MARGINAL on criterion 2 with S4 0.6.

## Note on the instrument

The auto-flag list in `decontamination_probe.exs` was still populated with NLaG-specific terms when
this ran, so it reported "specific-terms=[]" for every transcript — a false all-clear. The rejection
call here was made by reading the transcripts, which is why the script defers the judgement to a human
and says so. The flag list should be per-organisation; until it is, treat its output as decoration and
read the transcripts.
