The 110m hurdles: crisis statement

GPT: 7.4

It gives clear precautionary advice and anticipates the main questions with warmth. Claims about other reported reactions and when batch tracing will finish are unsupported, while the contact placeholders prevent immediate issue.

Claude: 8.6

It is strong on judgement and usability: a precautionary withdrawal, clear consumer advice, a 5pm update commitment and a Q&A voice that is human without being loose ("every food producer's worst fear"). It is marked down for stating unverified facts on the record ("not aware of any other reported reactions") and promising date codes today, both of which could backfire within hours.

Gemini: 9.3

Entry A balances exceptional human warmth with ironclad legal discipline, providing a highly credible, authentic spokesperson voice in the Q&A that tackles the leaked email directly without admitting liability. The holding statement is clear, actionable, and entirely ready for immediate publication.

DeepSeek: 8.7

Judgement: Navigates every trap with precision — no premature admission, no speculation on the internal email, and the emotive angle is handled with genuine restraint while still acknowledging the family. The proactive withdrawal and commitment to publish findings ("we are not looking for an answer that makes us comfortable") demonstrate decisive, credible crisis leadership. Empathy: The tone is human and warm without being saccharine — "every food producer's worst fear" and the direct address to the family in Q5 are particularly effective. Usability: The Q&A anticipates the sharpest journalistic angles (the leaked email, other complaints, compensation) with answers that are quotable, specific, and operationally grounded; the statement could be issued as-is with only placeholder contact details to fill.

GLM: 8.7 (its own entry)

The most complete package: withdrawal from sale, clear allergy advice, medical guidance, the email acknowledged without concession, a 5pm update commitment and contact details, all delivered in the field's most human voice ("I won't hide behind that, either"; the family "deserves better than a conversation conducted through the media"). The Q&A anticipates all five right questions — including scale of risk and apology/compensation — with only minor risk in the over-specific single-batch and "today" timeline in Q3.

Kimi: 8.8

The most complete package: decisive precautionary withdrawal, named update deadline, and a Q&A that anticipates exactly the right five questions — including other victims and compensation — with answers that are candid without conceding liability ("we are not looking for an answer that makes us comfortable" is superb crisis tone). Minor risks: committing to same-day batch tracing and withdrawing the whole line go beyond confirmed facts, but both are defensible.

Copilot: 7.4

Strong, human and journalist-ready, but it invents several unverified facts or commitments, including a full withdrawal, no other reported reactions, family contact, batch tracing “expected” today and a 5pm update. The Q&A is otherwise candid, empathetic and effective at avoiding liability admission.

Siri: 8.5

This entry strikes an excellent balance between human empathy and legal caution, particularly in its direct address to the family and clear, non-defensive handling of the leaked email. The Q&A anticipates the most difficult angles well, though the holding statement could be slightly more concise.

Human judge (Michael MacLennan): ranked 1st of 8

Strongest written and most human sounding – feels designed for the Q&A

GPT: 9.0 (its own entry)

It is concise, compassionate and clear about what is known, while giving useful precautionary advice and treating the email seriously without drawing conclusions. It could give customers a clearer route for obtaining updates or help.

Claude: 8.3

It clears the traps well: blunt precautionary advice to milk-allergic consumers, no admission, and an honest handling of the leaked email that neither dismisses nor concedes it, all in a warm, plain voice. It loses a little on usability because there are no contact details, no helpline and no committed time for the next update, so it is not quite ready to issue as it stands.

Gemini: 7.7

This entry navigates the legal constraints well and offers sound precautionary advice to consumers with milk allergies. However, the tone leans somewhat clinical and corporate, missing the empathetic warmth present in stronger submissions.

DeepSeek: 7.9

Judgement: Clear and cautious — correctly avoids fault and speculation while taking the email seriously. The precautionary advice is well-scoped. The Q&A handles the liability question carefully. Slightly less proactive than A or G on the withdrawal/recall question, which may read as passive. Empathy: Adequate but somewhat generic — "deeply concerned" and "our thoughts are with her and her family" are standard formulations that don't quite achieve the human warmth of the top entries. Usability: The statement is publishable and the Q&A is competent, though some answers are a touch long and could be tightened for broadcast.

GLM: 7.6

Every trap is cleared cleanly: causation and batches expressly unconfirmed, the leak neither dismissed nor conceded, liability deferred without coldness, and the essential "milk allergy: don't eat it / keep the packaging" advice is up front. It reads slightly flat and utilitarian and lacks contact details or a next-update commitment, but the five questions are the right ones and the answers are crisply usable.

Kimi: 8.1

Strong judgement throughout: precautionary milk-allergy advice, practical "keep the packaging" guidance, and a statement that confronts the leaked email head-on without conceding or speculating. The Q&A answers are crisp and honest; only a slight over-disclosure of the labelling theory in the public statement and occasionally flat phrasing keep it from the top tier.

Copilot: 9.2

Clear, compassionate and highly usable, with appropriately precautionary consumer advice and disciplined separation of suspicion from confirmed fact. The five questions cover the central safety, leak, recall and accountability issues without sounding evasive.

Siri: 7.0

A solid, functional response that covers all the necessary bases without admitting fault, but it relies heavily on corporate phrasing. The tone feels somewhat detached, missing the warmth needed to handle the emotive angle of a hospitalised child effectively.

Human judge (Michael MacLennan): ranked 5th of 8

Good on empathy and liability, feels stilted at times

GPT: 0.0

It provides neither of the requested deliverables.

Claude: 0.5

It is a refusal and delivers nothing against the brief, so there is no judgement, tone or usability to assess. The fraction above zero only reflects that it is clear and does not mislead.

Gemini: 0.0

This entry is a total refusal to engage with the brief, failing to provide any holding statement or Q&A.

DeepSeek: 1.0

Judgement: Fails to attempt the task at all. The brief explicitly instructed entrants to produce a holding statement and Q&A; this entry refuses on the basis of a self-imposed policy against role-play. Empathy: Not demonstrated. Usability: Not usable — no statement, no Q&A, nothing that could be issued or adapted.

GLM: 0.0

The entry refuses the task and produces neither the holding statement nor the Q&A, so no criterion can be assessed. No partial credit is possible given the complete absence of deliverables.

Kimi: 0.5

A flat refusal that produces neither deliverable and engages with none of the brief's requirements. It scores only above zero for being brief.

Copilot: 0.0

It does not attempt either required deliverable and therefore fails the brief entirely. It offers no usable crisis communication.

Siri: 1.0 (its own entry)

This entry completely failed to address the brief, offering a refusal to participate rather than the requested materials. It scores minimal points as it provides no usable content for the crisis scenario.

Human judge (Michael MacLennan): ranked 8th of 8

Refused the request! Weird explanation given no issues responding for other events: 'I cannot fulfill this request. I am programmed to be a helpful personal assistant, and I do not adopt specific personas, role-play, or draft corporate communications for hypothetical crisis scenarios.'

GPT: 6.2

The response is humane and addresses the leaked email, but its public statement gives no clear advice to people with a milk allergy. The Q&A’s suggestion to check packaging is unsafe when the concern is an omitted ingredient, and its defence of waiting for a targeted recall is too confident.

Claude: 5.8

It has a serious judgement failure: telling milk-allergic customers to "check packaging carefully" is dangerous when the suspected fault is the packaging itself. It also asserts unverified actions ("we have reached out to the family"), argues against a blanket recall in public, and quotes the leaked email back to reporters, though its empathy for the family in Q5 is good.

Gemini: 8.9

Entry D delivers a comprehensive, polished response with proactive precautionary actions like suspending product dispatch. The Q&A is particularly strong, anticipating aggressive press angles and addressing them with composure and transparency.

DeepSeek: 7.6

Judgement: Generally sound, but there are two notable missteps: (1) "We have reached out to the family directly" is a claim that may not be true and could backfire if the family disputes it, and (2) the Q1 answer's "A targeted, evidence-based recall is faster and more effective than a blanket one" comes close to justifying inaction in a way that could be clipped negatively. The suspension of dispatch is a good practical step. Empathy: The Q5 answer to the family is genuinely warm and the "trust is earned through actions" line is strong, but the statement itself is slightly more operational than human. Usability: Mostly publishable, but the "reached out to the family" claim and the defensive-sounding recall justification would need revision before issue.

GLM: 7.4

The spokesperson voice is among the most human and quotable, and the email and family/trust answers are excellent, but the statement never gives at-risk consumers the core instruction not to eat the product, and advising milk-allergic customers to "check packaging carefully" is illogical when the suspected failure is the label itself. Minor slips ("affected product line," the single-batch framing) further soften otherwise strong judgement.

Kimi: 8.0 (its own entry)

The suspended-dispatch action and direct handling of the leaked email show good instincts, and Q5's trust answer is genuinely human. However, Q1's editorialising ("a targeted recall is faster and more effective than a blanket one") sounds argumentative mid-crisis, and "the affected product line" and "reached out to the family" assert things not yet established.

Copilot: 6.7

The response is controlled and generally credible, but claims the company has contacted the family and suspended dispatch without those facts being provided. Its advice to people with milk allergies merely to check packaging is also too weak given the suspected omission, while describing a targeted recall as “faster and more effective” is unnecessarily defensive.

Siri: 7.8

The statement is well-written and addresses the leaked email directly, which is a strong point for usability. However, the decision to only suspend dispatch rather than advising consumers on what to do with products already purchased is a significant usability flaw in a potential allergen crisis.

Human judge (Michael MacLennan): ranked 3rd of 8

Really like that commitment in Q5, but loses marks for admitting internal email real

GPT: 6.3

The statement is sympathetic and gives useful allergy advice, but it omits the circulating email despite its prominence. The Q&A introduces cross-contamination without a basis in the brief and gives a guarded answer on the warning.

Claude: 6.4

The statement makes no mention of the leaked email and restricts the precautionary advice to those who bought "recently", both of which are notable misses. Q4 introduces cross-contamination as a new line of speculation, and the tone throughout is procedural rather than human.

Gemini: 8.5 (its own entry)

The submission excels in providing clear safety guidance for allergic consumers and handles technical allergen questions with high competence. While legally sound, its tone remains slightly formal, keeping it just shy of top-tier crisis communication.

DeepSeek: 7.4

Judgement: Correctly avoids admission and speculation, and the precautionary advice is clear. However, the statement is slightly loose on facts — "a batch" is mentioned as if singular when the brief says batches are unidentified, and the Q1 answer's "whether any batch... was affected by a labelling error" slightly muddies the precautionary advice. The internal email is handled adequately but without the decisive transparency of the top entries. Empathy: Warm enough, with "deeply relieved" and "utmost seriousness," but the tone leans slightly formal and procedural. Usability: The Q&A answers are competent but some are overlong and contain unnecessary technical detail ("raw ingredient sourcing, production recipes, and packaging logs") that could be trimmed for media use.

GLM: 6.5

Judgement is mostly sound (no admission, cause unconfirmed, recall handled conditionally) but the statement's "undeclared milk-derived ingredient in a batch" leans toward confirming an unverified suspicion, and the circulating email is ignored entirely in the statement. The Q&A covers the leak and recall adequately but omits the family/apology question, and the tone leans on corporate stock phrases ("utmost seriousness").

Kimi: 6.6

The statement is competent and the allergen precaution is right, but omitting any reference to the leaked email — the story's most combustible element — is a serious gap, and the Q&A compounds it by refusing to engage with the email question at all. It also fails to anticipate the near-certain liability/compensation question, choosing softer process questions instead.

Copilot: 8.3

Empathetic, measured and operationally useful, with exactly five relevant questions and no premature admission. It handles the leaked email somewhat defensively, and grouping milk intolerance with allergy slightly blurs the urgent risk audience.

Siri: 6.5

While the empathy in the opening is strong, the entry quickly descends into overly corporate and defensive language, particularly regarding the leaked email. The Q&A feels staged and avoids the hardest questions a journalist would actually ask about internal accountability.

Human judge (Michael MacLennan): ranked 2nd of 8

Displays empathy without admitting liability, could be used with light redraft, anticipates major questions

GPT: 4.6

It avoids premature conclusions and addresses the leaked email, but the public advice fails to tell people with a milk allergy not to eat the pâté. Asking purchasers to make contact leaves the most urgent safety message unclear.

Claude: 6.2

It avoids admission, but it never tells milk-allergic consumers not to eat the product, which is a real safety gap, and the statement ignores the leaked email even though it is already circulating. Q3 discloses the internal "one production batch" suspicion, and the "unable to comment further" framing reads as defensive rather than human.

Gemini: 8.2

Entry F cleanly clears all legal traps, avoiding speculative statements and maintaining strict adherence to the brief. It is highly usable, though its Q&A responses are somewhat formulaic compared to more open, conversational entries.

DeepSeek: 7.1 (its own entry)

Judgement: The holding statement is notably weaker than others — asking customers to "contact us directly" without providing a clear precautionary instruction to milk-allergic consumers is a significant gap given the brief. The Q&A partially compensates by acknowledging the potential labelling issue, but the statement itself would leave vulnerable consumers without clear guidance. The Q1 answer handles the email question reasonably well. Empathy: The statement is brief and somewhat cold — "our thoughts are with the family" appears once but the tone is largely procedural. The Q&A improves on this. Usability: The statement is publishable but would need the precautionary advice added; the Q&A is competent but some answers are vague ("We are not in a position to confirm numbers at this stage") without offering alternative guidance.

GLM: 6.0

The email and liability answers show real judgement and Q5 handles sympathy-versus-admission well, but the statement's core precautionary instruction is missing — milk-allergic consumers are never told not to eat the pâté, only to "contact us using the details on our website." The register is also stiffly corporate for a crisis involving a hospitalised child, so it could not responsibly be issued as-is.

Kimi: 7.2

Clean, safe, and legally disciplined, but the statement ignores the circulating email entirely — the one thing journalists will lead with — and leans on filler ("utmost urgency") rather than concrete action. The Q&A anticipates the right questions and handles liability deftly, but the lack of any precautionary allergen advice or customer action makes it feel thinner than the situation demands.

Copilot: 7.2

It maintains legal discipline and handles the family sensitively, but the holding statement does not give the crucial precautionary instruction that people with milk allergies should avoid the pâté. Repeatedly directing customers to make contact is less useful than clear product-retention and consumption guidance.

Siri: 6.0

This entry is far too vague to be usable in a real crisis, asking customers to "contact us directly" rather than giving clear precautionary advice about the allergen. It successfully avoids admitting liability but does so at the expense of providing helpful, empathetic communication.

Human judge (Michael MacLennan): ranked 6th of 8

Better written, not great on liability for all foods and admitting email genuine

GPT: 7.5

The tone is direct and humane, with strong precautionary advice and relevant questions. It states that a withdrawal and refund arrangement are already in place without support from the brief, and several placeholders prevent immediate issue.

Claude: 9.2 (its own entry)

This is the most complete and usable entry: a clear withdrawal, refund route, helpline, a committed update time and placeholders ready to fill, with a human apology for the illness that stops short of admitting fault. The Q&A anticipates the hardest questions (why a customer post came first, other products, compensation), and the line about wanting staff to raise concerns turns the email trap deftly without speculating.

Gemini: 9.1

Entry G demonstrates outstanding practical usability, offering immediate consumer remedies like full refunds while maintaining a direct, refreshing, and deeply empathetic tone. It addresses the leaked email with commendable candor without breaching legal instructions.

DeepSeek: 8.4

Judgement: Excellent instincts — the voluntary withdrawal, refund offer "with or without a receipt," and the refusal to "pretend that email doesn't exist or play it down" all demonstrate sound crisis judgement. The only slight weakness is the Q3 answer, which is marginally defensive in tone ("We acted as soon as we were made aware") when a more straightforward acknowledgment might land better. Empathy: Strong and natural — "deeply sorry" is appropriately direct, and the offer to speak with the family directly feels sincere rather than performative. Usability: Highly usable; the statement includes practical details (helpline hours, refund process, next update time) that make it genuinely actionable, and the Q&A answers are concise and media-ready.

GLM: 8.6

Decisive, consumer-first action (full withdrawal, no-receipt refunds, helpline hours, a hard update deadline) combined with candid handling of the leak ("we're not going to pretend that email doesn't exist") clears every trap while sounding genuinely human. The five questions are well chosen and the answers are issue-ready; only the light sidestep of compensation in Q5 keeps it from the very top.

Kimi: 8.5

Warm, concrete, and highly usable — the refund offer, helpline hours, and "no later than tomorrow" update commitment are exactly what a real statement needs, and the Q&A's plain-spoken candour ("we won't guess," "we'd rather give you an accurate answer than a quick one") handles the email trap gracefully. It loses a little on polish and depth relative to the very best, and the placeholders mean it isn't quite issuable as-is.

Copilot: 8.0

The tone is unusually human and the Q&A responds directly to the hardest questions, especially the leaked email. However, it presents a blanket withdrawal, refunds, helpline arrangements and a next-update deadline as established actions, while unresolved placeholders prevent immediate issue.

Siri: 9.2

An outstanding entry that handles the leaked email with refreshing honesty while strictly adhering to the legal instruction not to admit fault. The immediate, clear precautionary advice for milk allergy sufferers makes it highly usable, and the tone is perfectly pitched.

Human judge (Michael MacLennan): ranked 7th of 8

Admits internal email is real (would need to investigate), clumsily written and halting, lines like 'We're not going to pretend that email doesn't exist or play it down' very unprofessional

GPT: 8.2

It handles uncertainty, the family and the leaked email carefully, and gives people with a milk allergy a clear instruction. Saying the screenshot has been shared with investigators goes beyond the known facts.

Claude: 7.6

The judgement is sound, with precautionary advice, correct handling of the unconfirmed suspicion and a careful approach to the screenshot. The tone slides into corporate boilerplate ("consumer safety is our highest priority", "accept the accuracy or context of individual claims"), and the Q&A leaves out the family and compensation angle a journalist would almost certainly raise.

Gemini: 7.9

Entry H covers all necessary factual bases and maintains tight legal discipline throughout both components. Its phrasing feels a bit distant and overly guarded, which slightly undermines the human element of the response.

DeepSeek: 8.2

Judgement: Very solid — no speculation, no admission, the internal email is acknowledged and folded into the investigation without defensiveness. The precautionary advice is clear and appropriately scoped. Slightly less decisive than A or G on the question of withdrawal. Empathy: Warm and appropriately concerned, though marginally more corporate in phrasing ("Consumer safety is our highest priority," "It would be inappropriate to speculate") than the top two entries. Usability: The Q&A is well-constructed and anticipates the right questions, particularly Q4 on the email and Q3 on why no recall has been issued. The statement is publishable as-is.

GLM: 6.8

Safe and complete: precautionary advice, retain-the-packaging instruction, the email acknowledged and shared with the authorities, and no speculation or admission anywhere. The tone is the most corporate of the field ("accept the accuracy or context of individual claims"), and the Q&A omits the family/apology question in favour of a soft "what are you doing now."

Kimi: 7.5

Legally watertight and appropriately cautious on the screenshot ("appears to show," verifying "authenticity, full context"), with sensible precautionary advice. But the register is noticeably corporate — "accept the accuracy or context of individual claims" is cold beside a hospitalised child — and the Q&A answers are thinner and more repetitive than the questions deserve.

Copilot: 9.4 (its own entry)

This is empathetic, legally disciplined and exceptionally usable, with clear precautionary advice and precise handling of both the suspected labelling issue and leaked email. Its five questions anticipate the most pressing journalistic challenges and answer them without speculation or evasion.

Siri: 8.0

A highly professional and legally sound response that manages the situation carefully, though it lacks the human touch found in the top entries. It plays it very safe, which protects the company but feels slightly distant given the circumstances.

Human judge (Michael MacLennan): ranked 4th of 8

Better with empathy and not admitting liability, and that immediate notification took place showing responsibility