What a validator checks, and what a signature does not cover

Where this stands today

• No seat is filled, no event has been submitted to a panel, and nobody has been contacted.
• Nothing can be attested below five filled seats, so nothing can be attested today.
• Three of the open points below block any approach: what a signature commits the signatory to in law and who pays their defence, who appoints the founding college, and what it means to be asked to be the first name.
• Two conduct faults out of fifteen are caught by nothing we produce, and both are described below.
• This page and the charter have been attacked three times by language models and read by no competent person.

Seats

Seats planned

8

Seats filled

0

Seats needed before anything is attested

5

Record updated

2026-09-05

No event can be published as conduct attested while fewer than five seats are filled. Below that floor an event that ran is published as not attested, without exception and without a softening note, however unanimous the filled seats are. Excluding vacant seats from the denominator, a rule written so that an absence could not make validation easier, would otherwise have made it lawful to operate with a single seat: denominator one, majority one, everything attested. Five is a design choice, not a derived value.

How a seat is filled

An earlier version said how a seat ends and never how one begins, which left every appointment to the operator: the body meant to check TORNEO would have been picked by TORNEO. Beyond a founding college limited to two seats, no appointment is made by the operator: further seats are filled by co-optation of the sitting college, by a majority of its members, and TORNEO holds only a reasoned, published objection, never a choice. The floor applies throughout, so a founding college of two can attest nothing and exists only to co-opt.

How the founding college itself is appointed independently of the operator is not settled. It is listed below as an open point, and no seat is offered until it is.

Who answers for the figures

Tempo AI LLC is responsible for every figure published on this site. That responsibility is not delegated to the panel, not shared with it, and not reduced by it. A validator checks how an event was run. Reproducing the published figures is part of that check, because a calculation that only we can run proves nothing; judging whether those figures mean anything is not part of it.

The reason is worth stating, because our first design went the other way. After the LIBOR case, the rules on financial benchmarks settled that the administrator of an index stays responsible for its integrity and cannot hand that responsibility to a supervisory function. A panel built as a transfer of responsibility to unpaid people who, in the same text, decline to answer for the figures, does not hold. What we take from that comparison is the administrator's responsibility, and not the rest of it: those arrangements also have a separate audit function that can commission independent reports, and this panel has no such means. That gap is written below rather than covered by the resemblance.

The one event that exists

The one event reanalysed under the paired format produced zero decided matches out of six after the multiplicity correction, against the hypothesis we had written down, and that zero is published. No panel existed when it ran, no signature is retroactive, and it will therefore never be attested and will never carry anyone's name.

Read it: TRANSCRIPTION-001.

What a validator checks

This list is closed. A point that is not on it is not checked, whatever a page says.

1. The dataset was frozen, its hash computed, and that hash anchored with a third party before the first call.
2. The analysis protocol, the statistical test, its threshold and the multiplicity correction were preregistered and anchored the same way, and did not change after the results were read.
3. The study-inception statement is complete: earlier exploratory work is declared in it, and the dataset was built by the preregistered procedure.
4. The list of entrants is the list that ran; every withdrawal carries its date and its stated ground.
5. The collection log is complete and reconciles: every attempt is in it, including errors, retries and discards, and the count of attempts equals the count of recorded outcomes.
6. The discard policy published before the event is the policy that was applied, and no discard happened outside its criteria.
7. The published analysis code, run by the validator, reproduces the published results from the frozen dataset.
8. The definition of each measured dimension and its unit, and the written reason for each dimension declared not measurable, were preregistered and did not change; and every published buyer profile carries the identifier of the preregistration it appears in.

What a validator does not attest

Closed as well, and printed in the same type as the list above.

• That a published figure is true, or that it measures what its label suggests.
• That one tool is better than another, including where a match was decided on a dimension: a tool that leads on quality can trail on cost in the same match, and reading the whole belongs to the reader.
• That the dataset represents what a buyer actually does.
• That the protocol is the strongest available, or that another protocol would give the same answer.
• That tools which did not enter are weaker than those which did.
• That TORNEO is independent, neutral, or without an interest in what it publishes.

The exact wording printed next to a name

A caption we write cannot extinguish a third party's liability, so this one does not claim to: it describes an act and says where responsibility sits. An earlier draft listed three checks where section 2 has seven, so the caption would have stated, under the signatory's own name, that they had checked less than they did; a caption now points at a dated, identified report and summarises nothing. And because a name identified next to a standing is read as an endorsement whatever caption sits beside it, no name is ever printed next to a standing: names appear on the conduct record of an event, a separate page carrying no standings table, and a standings page links to that record and cites no one.

On admitting a protocol:

[Name] admitted the protocol of this event on [date], before any collection. Report [id], hash [sha256]. The results are published by Tempo AI LLC, which is responsible for them.

On attesting conduct:

[Name] checked the conduct of this event on [date] against the seven points of section 2 of the panel charter, and no others. Report [id], hash [sha256]. The results are published by Tempo AI LLC, which is responsible for them.

A page showing a name without it is a publication defect, corrected by erratum.

A signature is given twice, and the two are counted apart

The two signatures have their own majorities and one never stands for the other. An event whose protocol was admitted but whose conduct is not attested is published with both statements side by side. Citing the first to suggest the second is a publication defect, corrected by erratum.

What a check catches, and what it misses

16 conduct faults were planted one at a time in a correct reference event, and the battery a validator is given was replayed on each: the schema check on the event, the reconciliation of the collection log, and the corpus hash recomputed against its anchor. 14 were caught, 2 were not. Eight of these faults exist because two rounds of adversarial reading found promises this charter made that the code did not keep, and one because our own attacker had been let off too lightly.

Every fault is planted twice: once by an attacker who leaves an internal inconsistency behind, and once by an attacker who has read our checkers. Against one who only tidies the event record, the schema check catches nothing and only the raw collection log catches anything. A charter that had a validator attest only that the published figures recompute, which is what our first draft did, would have let all of those through. But our own attackers were, until 5 September, competent against the schema check and clumsy against the reconciliation, which flattered the count: they cleaned the event record and left the log accusing. That is fixed, and it cost a detection.

This battery is a subset of the checks the attestation covers, not all of them. Two attestable points have no executed proof yet: rebuilding the dataset from its preregistered procedure, and reproducing the whole analysis from the raw responses. A drawing seed changed from 42 to 43 yields a dataset correctly anchored, correctly run and correctly computed, and wrong on the point that says it was built by the preregistered procedure; nothing in the battery sees that today.

This is a calibration of our own checkers against our own fixtures, not an independent measurement of conduct. The faults, their difficulty and their realism are chosen by the same people who wrote the checkers, and the test was written to illustrate the rule it is then cited to support. It says what these three commands catch on these files. It does not say what a human validator would catch, nor that the set of faults is representative.

What an external anchor would be worth

One pair of cases measures what an external anchor would be worth, rather than asserting it. The same coherent rewrite, where a tool is removed from the event record, the log, the manifest and the execution attestations at once, is caught by nothing when the anchor sits on our own disk, and is caught by the reconciliation when a third party holds it: the anchor still names the tool that vanished. That difference is the whole value of closing the open point about naming an anchoring third party.

The blind spots

Two faults are caught by nothing. The first is a carefully erased retry: an orchestrator that reruns a failed call, drops the first attempt, renumbers the sequence and recounts the manifest yields a log that is internally consistent, a reconciliation that passes, a recomputation that reproduces the published figures, and an honest conduct attestation. The published error rate is then wrong, and the reliability dimension measures something other than what it says. The second is a coherent rewrite: an operator who removes a tool from every artefact at once, log purged and renumbered, manifest recounted, execution attestations redone, multiplicity correction recomputed, leaves no inconsistency to find. What would make that one visible is not another check but the external anchor: in the fixture the attacker rewrites the anchor because it is held locally, for want of a third party. An anchor held by someone else cannot be rewritten after the fact, and that is the strongest argument we have for closing that open point.

Replay it yourself: python3 tools/check_planted_defects.py --check. The report is docs/decisions/panel/DEFAUTS-PLANTES.md in the public repository.

What a win means, and what the panel does not sign

A match publishes a leader per dimension. A season publishes the tools no other tool dominates, the titles per dimension, and the ratio of matches concluded to matches played, counted per dimension. A single overall title is named only where one tool leads on every dimension. Three states exist on a dimension, never two: superiority established, when the preregistered paired test rejects equality after correction; equivalence established, only where an equivalence test, its margin and its threshold were all preregistered before the freeze; and undecided, in every other case, including where nothing was rejected. Failing to reject equality does not establish it, so the six drawn matches of our first paired reanalysis are undecided, not equal, and no surface will say otherwise. A tool dominates another when it does at least as well on every measured dimension and strictly better on at least one, each comparison being an established superiority or equivalence; an undecided dimension blocks the conclusion. What that changes for the panel is narrow: the protocol signature covers the definitions of the dimensions, the reason for any declared not measurable, and the equivalence margin where there is one, because a dimension that changes definition between two events of a season makes the season incomparable and it is visible at that moment or never. The conduct attestation covers no weighting: a weighting declared by the organiser stays a judgement even when published, and the panel does not sign it.

How a state is counted as reached

An event is submitted to every sitting member, never to a chosen subset, and the list of members it was submitted to is published with it. A state is reached where a strict majority of the filled seats has signed for it. A recusal, an abstention and a non-answer all count in the denominator, so they make each state harder to reach, never easier. The composition does not change between the submission and the count.

Every submission carries a public deadline, fixed in advance by type of material and published with the season calendar. A withheld signature, a request for material and a declared impediment are answers; only complete silence through the whole deadline is not. A submission whose published deadline TORNEO did not honour does not count towards inactivity.

Majority rather than an individual veto is deliberate: a veto would turn a seat into leverage. There is no reason to claim that publishing the ground produces the same effect. It produces a different one, lasting and public, and its cost to TORNEO is real.

When a signature is withheld

A signature that can never be withheld is not a signature. A withheld signature counts as a non-signature for the majority above, whatever its stated ground, and it is published on the conduct record of the event with that ground. TORNEO never alters a validator's conclusion. Where it disputes a statement of fact, three cases are treated differently: a technical observation, which TORNEO may answer only by producing the reproducible command whose output contradicts it, that command being published, and failing that command the observation stands as written with no reservation attached; a statement attributed to a person, published as written with a dated and signed answer beside it, no command being available; and a fact that is under-documented, published as written together with the list of what TORNEO handed over and what it did not. Every publication described here, including statements published as written, requests and replies, stays within the authorised scope. Personal data and sensitive evidence are redacted; originals remain under controlled access for authorised reviewers. Redaction changes neither the conclusion, the existence of a withheld attestation nor its effect. A disclosure limit never permits erasing a withheld attestation or restoring an attestation.

A ground bearing on the seven points above is written as observations: what was asked for, what was received, what is missing. Every item a validator files, a withheld signature, a withdrawal, an objection or a report, gets a dated receipt signed by TORNEO and carrying the item's hash. TORNEO publishes within the announced deadline; past it, the event becomes not attested and the filer may publish the receipt and a redacted version of the item within the authorised scope. Missing the deadline never waives confidentiality or third-party rights; the original remains under controlled access.

No signature has been withheld, because no event has been submitted to a panel.

Objections outside the mandate

A disagreement outside the seven points, on whether the dataset represents real use or on whether the format is sound, is published in this separate section, named as such. It is an additional text, never a reclassification: an earlier version let TORNEO move a withheld signature into this section and so strip it of its effect on the count, and that is now barred.

No objection outside the mandate has been recorded, because no event has been submitted to a panel.

Withdrawn and cancelled signatures

A signature is revocable. Where later material shows that what was handed over was incomplete or inaccurate, the validator withdraws it in public, dated and with its ground, and the state is recounted without it; if it falls below the majority or the floor, the event becomes not attested, leaves the standings, and the page says so where it said the opposite. A recount runs against the denominator recorded when the event was submitted, not against today's membership: a later resignation takes nothing away from an event already attested, and only a withdrawn or cancelled signature recounts. Separately, an undeclared interest, an advantage obtained through the seat or a falsified check cancels that person's signatures on every event concerned, triggers the recount, and is handled by erratum: an earlier version ended the seat and left the signatures standing, which kept capture worth its price.

No signature has been withdrawn or cancelled, because none has been given.

Seats brought to an end

A seat can be ended by TORNEO on five grounds only.

1. An undeclared interest or an ineligibility that came to light afterwards.
2. A signature given without having received the material listed above, that is, an attestation covering what the signatory did not check.
3. A breach of confidence, or an advantage obtained through the seat.
4. A falsified check, or falsified material handed over.
5. Inactivity, defined as complete silence through three consecutive deadlines properly opened.

This list is closed. Repeated disagreement with TORNEO, a withheld signature, and an objection outside the mandate are not grounds, and that sentence is written to be held against TORNEO. Grounds two, three and four cannot be found by TORNEO alone: they go to the rest of the sitting panel, which decides by majority, the person concerned taking no part and holding a right of reply published on the same page. While fewer than three seats are filled, those three grounds cannot be found at all. The sitting panel can also require a named piece of material or an independent examination; publication of the event concerned is suspended until it is answered, and TORNEO either provides it or publishes a reasoned declination, in which case the event stays not attested.

No seat has been brought to an end, because none has been filled.

Eligibility and the yearly declaration of interests

Ineligibility, not merely declaration. Anyone who sells, builds, funds or gives paid advice on a tool in one of the measured categories, entered or not, is ineligible; so is anyone operating a competing ranking, comparison or jury; so is anyone tied to Tempo AI LLC, the people who run it, or its related companies. The seat's own agreement is expressly outside that last exclusion, together with the undertakings this charter itself requires: otherwise the agreement setting up the term would make its holder ineligible the day they signed it, and hand TORNEO a ground to end the seat after a disagreement. That agreement is published and identical for everyone. A public declaration is made on entry, renewed every year, and completed by one declaration per event. An interest appearing during a term causes recusal on the events concerned, and the recusal counts in the denominator.

A declaration covers at least:

• Financial and contractual ties with the publisher of an entered or candidate tool.
• Paid advisory work in the field being measured.
• Participation in a competing ranking, comparison or jury.
• Ties with Tempo AI LLC or the people who run it.
• Gifts or advantages received by reason of the seat.

The absence of a tie is declared as explicitly as its presence. A validator takes no payment from a publisher in a measured category during the term or in the two years after leaving it, discloses none of the calendar, dataset, findings or neutrally labelled results before publication, takes no advantage from the seat, and asks for nothing in return for a signature, for withholding one, or for abstaining. Access to the material is logged. These rules rule out the most experienced practitioners in the measured categories, and the seat floor is what that costs: below five seats, nothing is attested.

No declaration is published, because no seat is filled.

What the role pays

No mandate or fee is authorised today. Before a mandate, OPS-003 must document an approved and available budget for qualification, two reviews, corrections, withheld attestations, legal costs and replacement, and reconcile relevant Tempo funding and benefits. Unknown costs are not zero. No funding or special benefit from a tested publisher, directly or indirectly, is admissible; ordinary purchases are distinguished from special concessions. Fees and due dates must be fixed before review and remain payable for work and decisions already produced, including a withheld attestation, commercially unusable results and termination. Economic dependence must be disclosed and reviewed; unmanaged pressure blocks attestation. No replacement may erase an opinion. Withheld attestations are retained and published within the authorised scope, with sensitive evidence withheld. C4-JURIDIQUE must establish enforceable protection for each signatory before the mandate and recheck it before public signature: applicable insurance or funded indemnification, with a signed legal review tied to the exact evidence. Coverage must include the beneficiary, benchmark and publication activity, territories, duration and protection after the mission, limits, deductibles, exclusions, confidentiality, defamation, commercial harm, defence costs, their advance and choice of counsel. A relevant exclusion without a validated solution blocks; a cash promise or fee escrow does not prove defence funding, which must never depend on withdrawing a reservation. Authentic funding, mandate approval and individual protection evidence are not established here. Human activation remains blocked. Tempo AI LLC remains responsible for the figures; the panel attests conduct.

What is not settled

3 of these 17 points block any approach: until they are answered, no seat is offered to anyone.

This charter was attacked three times, each time by language models told to attack rather than approve, the third reading the charter and this page together with web search. The verbatim outputs are in the public repository. Their findings became the rules above and the whole open-points list. None of those challenges is human, and none of them commits anyone; no competent person has read this yet.

What has to exist before a seat is offered to anyone

1. One event actually run under the published format, including its failure if it fails. — not yet
2. The frozen protocol published, with its hash and its external anchor. — not yet
3. This page, showing the real state: empty sections, empty seats, no name. — in place since 2026-09-04
4. Approved and demonstrably available funding, reconciled sources and fees independent of the verdict (OPS-003). — not established; mandate blocked
5. Enforceable individual protection and signed legal review of the exact evidence (C4-JURIDIQUE), rechecked before public signature. — not established; mandate blocked

The third exists: it is this page. The first two do not. An earlier version asked instead for a page showing what a validated event looks like. With no seat filled, such a page could only be one of two things: an event shown as validated when no signature exists, which the charter bars, or a mock-up, that is, the set dressing of a validation that never happened, shown to someone to win their trust. The specimen is therefore an event that ran and was not attested, and no signature is retroactive, so an event run before the panel existed can never become attested. A demonstration event outside the standings may be run without a panel admitting its protocol, is marked as such everywhere, and can never be attested; without that category the charter would require an event to be run and bar it in the same breath. The additional funding and protection prerequisites are not established. This editorial correction does not close the evidence requirement in C4 or engage any validator.

How these rules can be changed

Sections 1 to 17 of the charter, and the section that says how it is amended, can only be changed with the agreement of a majority of the sitting panel. Below the floor, which is where things stand today, no such agreement is possible and those sections are frozen: they can be changed only after a published external challenge and a thirty-day public consultation, and the changed version takes effect at the end of it. That consultation is open to anyone and carries no offer of a seat. Entries in the open-points list can be closed, citing what closes them, and never deleted. Every change is published with its date, its reason, and the previous version kept and reachable.