From 0fe9e79fba3de363360775dbb3f972c64de874b8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?LGD-Eval=20=E2=80=94=20Independent=20AI=20Evaluation?= Date: Mon, 16 Nov 2026 13:59:02 +0800 Subject: [PATCH 1/4] Add verifier-side interface record (linked contribution to FG-TIDA/themes#46) VR1-VR5 state what a verifier must declare: evidence grade, moment used, performed vs asserted checks, what a pass may claim, and that the verifier's own report is a claim under R7. Assisted-by: WorkBuddy --- FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md | 193 +++++++++++++++++++++++++ 1 file changed, 193 insertions(+) create mode 100644 FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md diff --git a/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md b/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md new file mode 100644 index 0000000..637b40e --- /dev/null +++ b/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md @@ -0,0 +1,193 @@ +# Verifier-side interface to the record requirements + +**A linked contribution record to AAR-REQ (`FG-TIDA/themes#46`)** + +| | | +|---|---| +| Author | Xinghua Zhao (Steven Zhao), independent expert, GitHub `@zhaoxinghua09-cell` | +| Offered in | Individual-expert capacity, as stated on `FG-TIDA/themes#46` on 29 October 2026 | +| Hosted in | `LGD-Eval/evaluation-protocol` — the public home of the evaluation material this record draws on. Hosting is a publication venue, not an authorship claim: the record is offered in individual-expert capacity, and the account relationship is disclosed below. | +| Version | 0.1 — 16 November 2026 | +| Status | Draft for attributed review before the first in-person FG-TIDA meeting (Paris, 30 November – 1 December 2026) | +| Linked work item | AAR-REQ 0.2 — https://tyche-dev.github.io/tida-aar-req/ | +| Licence | CC BY 4.0 | + +--- + +## 1. Why this record exists + +AAR-REQ states requirements on the record. It also states, in Out of Scope, that **"what a conformance verdict asserts and how it is checked"** is not its subject, and hands that question to `FG-TIDA/themes#6`, task 3. + +This record occupies that handed-off position from the verifier side. It asks one question in four parts: given a record that satisfies R1–R8, **what can a party who was not there establish from it, what must that party state in order to be believed, and what is a "pass" allowed to mean when it is stated?** + +It adds nothing to R1–R8 and modifies nothing in them. Every clause below is a statement about the verifier's report, not about the record's content. Where a clause depends on a record-side requirement, it names that requirement and takes it as given. + +## 2. Scope + +Four statements a verifier makes, and one obligation it carries: + +- **VR1** — the grade of the evidence it relied on; +- **VR2** — the moment it evaluated against; +- **VR3** — whether a check was performable, and whether it was performed; +- **VR4** — what a pass may claim, bounded by what was actually checked; +- **VR5** — that its own report is a claim subject to R7. + +## 3. Out of scope + +Stated explicitly so that overlap is decidable rather than implicit: + +- **How a verifier reaches a class.** The class vocabulary is R1–R4; this record does not prescribe decision procedure, thresholds, or evidence weighting. +- **Encodings, schemas, field names, media types, test vectors.** As in AAR-REQ §1.3. +- **What a conformance verdict asserts about an artefact** (`FG-TIDA/themes#6`, task 3). This record is about the verification verdict on a *record of an act*, not on a *product*. +- **Where authority originates and what may be conferred** (`FG-TIDA/themes#5`). +- **Legal effect.** A pass is a statement about checks performed; it is not a legal opinion and confers nothing. + +## 4. The clauses + +Attribution convention follows AAR-REQ Table 2.1: each clause names who proposed it, and in brackets who refined it. + +### VR1 — Evidence grade is a property of the verifier's relation to the evidence, not of the record + +*Proposed by @zhaoxinghua09-cell.* + +A record can be complete and still be unusable by a given verifier, because usability depends on what that verifier can reach. The grade therefore cannot live inside the record: AAR-REQ §1.3 puts "how any value in a record is computed" out of scope, and a grade assigned by a producer would in any case be a claim the verifier has not checked. + +The verifier states one grade, from an ordered set, for the evidence it relied on: + +| Grade | Meaning | +|---|---| +| **A** | Checked by the verifier against a source it can reach without trusting the producer | +| **B** | Checked by a party with no relationship to the producer, named in the report | +| **C** | Corroborated indirectly — consistent with independent material, not checked at source | +| **D** | Asserted only — the verifier relied on the producer's statement | + +Two properties matter more than the number of grades. First, the grade is **about the claim, not the actor**: a strong grade on a claim does not make the actor's authorization valid, and VR1 must not be read as doing so (this keeps the separation argued in `FG-TIDA/themes#42`). Second, the set is **ordered but not arithmetic**: grades are not averaged across a chain, because a chain is only as strong as its weakest link and averaging would hide exactly that. + +### VR2 — The verifier states which moment it evaluated against, chosen from three + +*Proposed by @zhaoxinghua09-cell. The three moments are @OlehVasylenko2000's (12 November 2026) and are adopted here unchanged; the verifier-side obligation is this record's.* + +R6 makes freshness a relation, not a property. A relation needs two terms, and the second term — the moment the verifier used — is currently unstated on the verifier side. So the verifier names which of the three moments it used: + +1. **The moment of the act** — the moment the record describes. +2. **The moment the producer states for a claim** — carried under R7.2. +3. **An independently bounded moment** — which establishes only that the record existed no later than that time, and is not the freshness relation of R6. + +A verdict evaluated at moment 1 and a verdict evaluated at moment 3 are **not the same object**, even when every field is identical, because the state that governs the act (R8.5) may have changed between them. A report that omits this is not reproducible by a later reader, who cannot tell which question was answered. + +The verifier also states whether rechecking was possible, taking R6's "means, if any, of rechecking" as given. Where no means exist, the report says so rather than leaving the field empty — an absent rechecking means is itself a finding. + +### VR3 — "Checkable" is not "checked" + +*Proposed by @zhaoxinghua09-cell.* + +R7.3 requires tamper-evidence that a party with no relationship to the producer **can** check. That is a statement about capability, and it is satisfied by a record no one has ever examined. + +So the verifier states, for each check it reports, which of three happened: + +- **performed by the verifier** — the check was run, and the means are named; +- **performed by a named third party** — the check was run by a party with no relationship to the producer, and the report names it; +- **asserted** — the check was not run by anyone the verifier can name. + +This distinction is load-bearing because the failure it prevents is silent. A record that is checkable but unchecked produces a verdict that looks like a verdict from a check, and nothing in the record distinguishes the two. Where a check was asserted rather than performed, the evidence grade (VR1) cannot exceed **D** for that claim, whatever the record's properties. + +### VR4 — A pass claims no violation found within a stated boundary, and nothing more + +*Proposed by @zhaoxinghua09-cell, carrying forward the question posed in `FG-TIDA/themes` PR #28.* + +A pass carries four things, and is incomplete without any of them: + +1. the **checks performed** (with VR3's status for each); +2. the **checks not performed** — named, not implied; +3. the **boundary** within which the checks ran; +4. the **moment used** (VR2). + +With those four, a pass means exactly: *no violation was found within this stated boundary, using these checks, at this moment.* It does not mean the act was authorized, that the record is true, or that the delegation was sound. R7 says as much for carried claims — that a claim meets R7 does not make it true — and the same restraint has to hold one level up, at the verdict. + +The property this buys is that a pass is **falsifiable by its own declaration**. A later party can withdraw it by showing a single check inside the declared boundary that was not performed, without having to relitigate the facts. A pass that does not declare its boundary cannot be withdrawn this way, and therefore cannot be relied on — it can only be believed. + +### VR5 — The verifier's own report is a claim, subject to R7 + +*Proposed by @zhaoxinghua09-cell. This closes the recursion @OlehVasylenko2000 raised for R7.* + +If the verifier's report matters to a later decision — and VR4 makes it the thing relied on — then the report is itself a carried claim and R7 applies to it unchanged: attribution (R7.1), a stated moment (R7.2), and tamper-evidence a party with no relationship to the producer can check (R7.3). + +The practical consequence is narrow but sharp. **A verifier that does not disclose its own evidence basis is producing an unattributed claim**, and VR4's boundary cannot be read off it. This is not a demand that verifiers publish; it is a statement that an unpublished basis caps the grade the next party can assign. + +## 5. Interface table + +What the verifier reads from each record-side clause, and what it must state in return. The right-hand column is this record's contribution; the left-hand column is AAR-REQ's and is not restated. + +| AAR-REQ | What the verifier reads | What the verifier must state | +|---|---|---| +| R1, R2 | the rejection class and the reported cause | VR1 grade for the claim about the cause; VR3 status per check | +| R3 | the four tier-1 classes are decidable from the record and the verifier's own state | which classes it actually decided, and which it could not | +| R4, R4.1–R4.3 | whether continued standing is checkable, and which moment the relying party used | VR2 moment, stated explicitly rather than inherited | +| R5 | negative and null assertions are claims like any other | VR1 grade per assertion; a null assertion is graded, never skipped | +| R6 | the moment the assertion was true, and the means of rechecking | VR2 moment used; whether rechecking was possible, and if not, that it was not | +| R7, R7.1–R7.3 | attribution, stated moment, tamper-evidence | VR3 performed / by a named third party / asserted; VR5 for its own report | +| Absence and silence | which of the three states holds | which state was established, and the scope and author of any bounded search | +| LR1–LR5 | which lineage case holds | the case reported, with the evidence grade for it | +| R8, R8.1–R8.5 | act, execution, supported relation to a logical agent, effective state | which of the five it established, and at which moment (VR2) | + +## 6. Three boundary cases + +These are the cases where a verifier most easily states more than it checked. + +**Case A — checkable, uncheckable-by-this-verifier.** The record is complete and tamper-evident; the verifier has no relationship to the producer and no route to recheck. Verdict: *no violation found within boundary, grade D, rechecking not possible.* Not a pass in the ordinary sense, and not a rejection either — R3's classes do not apply, because nothing was found to be wrong. + +**Case B — right record, wrong moment.** Every tier-1 check passes at the moment of the act; standing was withdrawn before the verifier looked. Reporting a pass at moment 3 while presenting it as a statement about moment 1 is the error VR2 exists to prevent. The verifier's verdict is about the moment it names, and the report must let a reader see which. + +**Case C — absence is not silence.** No record at all, a bounded search that returned nothing, and a reference that was established are three different findings. Only the second carries a scope and an author; the third is a positive finding. Collapsing them yields a verdict that reads as "nothing wrong" when the correct reading is "nothing was looked at". + +## 7. Open questions + +Carried openly, in the style of AAR-REQ §2.1. + +1. **Is four grades the right cut?** Three would force B and C together, which hides the difference between a check performed elsewhere and a corroboration. Five invites distinctions no one can apply consistently. *(@zhaoxinghua09-cell)* +2. **Should VR2's moment be a required field in the report, or only a required statement?** A field is machine-checkable; a statement is cheaper and survives formats the group has not settled. *(@zhaoxinghua09-cell)* +3. **Does VR5 bind all verifier output, or only published verification reports?** Binding all output is the stronger position and the harder one to meet, and I would rather the group choose it knowingly than have it arrive by implication. *(@zhaoxinghua09-cell)* +4. **Where does the boundary in VR4 come from?** If the group names a minimum boundary somewhere, this record should cite it rather than let each verifier declare its own. *(@zhaoxinghua09-cell)* + +## 8. Provenance + +- AAR-REQ 0.2, web — https://tyche-dev.github.io/tida-aar-req/ +- AAR-REQ 0.2, PDF — https://tyche-dev.github.io/tida-aar-req/specification.pdf +- Originating thread — https://github.com/FG-TIDA/themes/issues/46 +- Verifier-side charter — https://github.com/FG-TIDA/themes/pull/28 +- Evidence grading (separating evidence quality from authorization validity) — https://github.com/FG-TIDA/themes/issues/42 +- Evidence-graded authorization verdicts — https://github.com/FG-TIDA/themes/issues/43 +- Conformance verdict scope, handed off by AAR-REQ §1.3 — https://github.com/FG-TIDA/themes/issues/6 +- Author identity — https://orcid.org/0009-0001-0512-1237 +- Account relationship and conflict-of-interest disclosure — https://github.com/LGD-Eval/evaluation-protocol/blob/main/CONFLICT_OF_INTEREST_DISCLOSURE.md + +## 9. Limitations + +- **The clause set has had one author.** VR1–VR5 are coherent with each other by construction, which is not the same as being right. Two of the five (VR3, VR4) rest on a distinction that is easy to state and hard to apply; a reviewer should test them against a case they already know the answer to. +- **No implementation accompanies this version.** The reference implementation covers the evidence-gated state machine, not the verifier report format these clauses describe. A worked example set is the obvious next increment and is not claimed here. +- **The material this record draws on has not been through peer review.** `FG-TIDA/themes#42` and `#43` are proposals under discussion, not settled positions, and VR1 is a restatement of their separation argument rather than independent evidence for it. +- **Drafted with AI assistance** under author direction; the author is responsible for every claim above. + +## 10. Version history + +| Version | Date | Change | +|---|---|---| +| 0.1 | 2026-11-16 | First draft, posted for attributed review on `FG-TIDA/themes#46` | + +--- + +## 中文要点对照 + +> 英文为权威文本;以下为逐条中文要点,供中文读者对照,非逐句对译。 + +**本记录的位置。** AAR-REQ 在 Out of Scope 里明写「一致性判定能主张什么、怎么检查」不是它的事,并转给 `themes#6` task 3。本记录就站在这个被让出来的位置上,从验证方一侧回答:一份满足 R1–R8 的记录,一个当时不在场的人能从中确立什么、必须自己声明什么、以及「通过」到底允许主张什么。**不给 R1–R8 增加任何内容。** + +| 条款 | 中文要点 | +|---|---| +| **VR1** | 证据等级是验证方与证据之间关系的属性,不是记录的属性。记录再完整,对某个验证方也可能不可用。四档:A 自己直接核到源头;B 由与产出方无关的具名第三方核过;C 间接佐证;D 仅有断言。等级针对**主张**,不针对**行为人**——等级高不等于授权有效。四档有序但不可做算术平均,链的强度等于最弱一环。 | +| **VR2** | 验证方必须声明自己用的是哪个「时刻」,三选一:①行为发生时 ②产出方为某主张声明的时刻 ③独立可界定的时刻(只能证明记录不晚于某时,不是 R6 的时效关系)。同一份记录、同一个字段,用不同时刻得出的结论**不是同一个东西**,因为 R8.5 的有效状态可能已经变了。还要声明「能否复核」——没有复核手段本身就是一个结论,不能留空。 | +| **VR3** | 「可检查」不等于「已检查」。R7.3 只保证「与产出方无关的人**能**检查」,一份谁都没看过的记录同样满足它。所以每项检查都要标状态:验证方自己做的/具名第三方做的/仅有断言。这条重要,因为失效是静默的——可检查但未检查的记录,产出的结论看起来跟检查过的一样。仅有断言时,该主张的证据等级最高只能到 D。 | +| **VR4** | 通过 = 在声明边界内、用声明的检查、在声明的时刻,未发现违规。**仅此而已**,不等于行为获授权、记录为真、委托成立。必须带四样:做了什么检查、**没做什么检查**(要点名)、边界、所用时刻。这样换来的性质是:**通过可以被它自己的声明推翻**——后人只要指出边界内有一项检查没做,就能撤回它,不必重打事实。不声明边界的通过,无法被这样撤回,因此不能被依赖,只能被相信。 | +| **VR5** | 验证方自己的报告也是主张,R7 照样适用:归属、声明时刻、与产出方无关者可校验的防篡改。推论很窄但很利:**不披露自身证据基础的验证方,产出的是一份无归属主张**,VR4 的边界无从读起。这不是要求验证方必须公开,而是说未公开的基础会压低下一个当事人能给到的等级。 | + +**三个边界情形。** A:记录可检查、但本验证方查不到——结论是「边界内未发现违规、等级 D、无法复核」,既不是通过也不是拒绝(R3 的四类不适用,因为没发现错)。B:记录对、时刻错——行为时全部通过,验证方来看时资格已撤回;把时刻③的结论当时刻①的结论报,正是 VR2 要挡的错。C:无记录 ≠ 静默——「根本没有记录」「有界检索查无」「引用已确立」是三种不同结论,只有第二种带检索范围和作者;混为一谈会把「什么都没查到」读成「什么都没问题」。 From a38672d2608d1fcccf54c7b835ce95ac55dfc02c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?LGD-Eval=20=E2=80=94=20Independent=20AI=20Evaluation?= Date: Mon, 16 Nov 2026 14:12:40 +0800 Subject: [PATCH 2/4] Revise verifier-side interface record to 0.2 Positioning restated as a refinement of R1's unstated accept side, not an unoccupied space. VR2 premise corrected against R4.2: scope extension, not a blank. Verifier defined against relying party. VR1 attribution moved to #42/#43. Grades redefined so each is reachable and each status maps to one. VR4 given a floor; Case A reclassified as unresolved. R8 count corrected. Conformance language added. Limitations extended. Assisted-by: WorkBuddy --- FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md | 202 ++++++++++++++----------- 1 file changed, 113 insertions(+), 89 deletions(-) diff --git a/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md b/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md index 637b40e..5df6bd5 100644 --- a/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md +++ b/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md @@ -7,172 +7,196 @@ | Author | Xinghua Zhao (Steven Zhao), independent expert, GitHub `@zhaoxinghua09-cell` | | Offered in | Individual-expert capacity, as stated on `FG-TIDA/themes#46` on 29 October 2026 | | Hosted in | `LGD-Eval/evaluation-protocol` — the public home of the evaluation material this record draws on. Hosting is a publication venue, not an authorship claim: the record is offered in individual-expert capacity, and the account relationship is disclosed below. | -| Version | 0.1 — 16 November 2026 | +| Version | 0.2 — 16 November 2026 | | Status | Draft for attributed review before the first in-person FG-TIDA meeting (Paris, 30 November – 1 December 2026) | | Linked work item | AAR-REQ 0.2 — https://tyche-dev.github.io/tida-aar-req/ | | Licence | CC BY 4.0 | --- -## 1. Why this record exists +## 0. Terms -AAR-REQ states requirements on the record. It also states, in Out of Scope, that **"what a conformance verdict asserts and how it is checked"** is not its subject, and hands that question to `FG-TIDA/themes#6`, task 3. +Two terms are used throughout and are defined here so that the interface below is readable against AAR-REQ's own text. -This record occupies that handed-off position from the verifier side. It asks one question in four parts: given a record that satisfies R1–R8, **what can a party who was not there establish from it, what must that party state in order to be believed, and what is a "pass" allowed to mean when it is stated?** +**Relying party.** As in AAR-REQ: the party that acts on a claim about an agent after the fact. -It adds nothing to R1–R8 and modifies nothing in them. Every clause below is a statement about the verifier's report, not about the record's content. Where a clause depends on a record-side requirement, it names that requirement and takes it as given. +**Verifier.** The relying party itself, or an agent acting for it, when it evaluates a record and states a result. This record does not introduce a second role: wherever AAR-REQ binds a relying party, the obligation falls on the verifier, and wherever AAR-REQ's text says "relying party" this record's clauses are satisfied by the same party under the other name. The name is used because the clauses are about the act of verifying, not about the act of relying. -## 2. Scope +**Record-side / verifier-side.** Record-side requirements constrain what a record must carry. Verifier-side requirements constrain what a party must state when it reports a result from a record. R1–R8 are record-side. The clauses below are verifier-side. + +## 1. Where this record sits + +AAR-REQ §1.2 puts two things in scope: requirements on the record, and **"what a relying party reports from it"**. The second is not left open in general — R1–R4 enumerate the rejection classes a relying party reports, and R1 states that a bare accept or reject is not enough. + +What is left open is narrower, and it is the gap this record works in: + +- **The rejection side is enumerated; the accept side is not.** R1 forbids a bare accept and R3–R4 give four tier-1 and one tier-2 rejection class with contents. No clause says what a non-bare accept must carry, or what it is allowed to assert. R1 creates the obligation without supplying its content. +- **§1.3 hands away the verdict question.** "What a conformance verdict asserts and how it is checked" is out of scope and assigned to `FG-TIDA/themes#6`, task 3. -Four statements a verifier makes, and one obligation it carries: +So the position is not an empty space: it is an obligation with no content, next to a question that has been moved elsewhere. This record supplies content for the first and stays clear of the second — it says nothing about conformance verdicts on an artefact, only about the report a party makes after verifying a record of an act. -- **VR1** — the grade of the evidence it relied on; -- **VR2** — the moment it evaluated against; -- **VR3** — whether a check was performable, and whether it was performed; +This is a refinement of AAR-REQ's existing verifier-side requirement, not an addition to the record-side set. **It adds nothing to R1–R8 and changes nothing in them.** Where a clause depends on a record-side requirement, it takes that requirement as given and names it. + +## 2. Scope + +- **VR1** — the grade of the evidence the verifier relied on; +- **VR2** — the moment the verifier evaluated against; +- **VR3** — whether each reported check was performed, and by whom; - **VR4** — what a pass may claim, bounded by what was actually checked; -- **VR5** — that its own report is a claim subject to R7. +- **VR5** — the verifier's own report as a claim the next party has to grade. ## 3. Out of scope -Stated explicitly so that overlap is decidable rather than implicit: +Stated so that overlap is decidable rather than implicit: -- **How a verifier reaches a class.** The class vocabulary is R1–R4; this record does not prescribe decision procedure, thresholds, or evidence weighting. -- **Encodings, schemas, field names, media types, test vectors.** As in AAR-REQ §1.3. -- **What a conformance verdict asserts about an artefact** (`FG-TIDA/themes#6`, task 3). This record is about the verification verdict on a *record of an act*, not on a *product*. +- **Decision procedure, thresholds, evidence weighting.** How a verifier reaches a class is R1–R4's subject and is not prescribed here. +- **Encodings, schemas, field names, media types, test vectors.** As in AAR-REQ §1.3. VR1's four grades and VR3's three states are **descriptions of relations, not enumerations to be encoded**: a verifier may state them in its own words, and no field name is proposed. +- **Conformance verdicts on an artefact** (`FG-TIDA/themes#6`, task 3). - **Where authority originates and what may be conferred** (`FG-TIDA/themes#5`). -- **Legal effect.** A pass is a statement about checks performed; it is not a legal opinion and confers nothing. +- **Legal effect.** A pass is a statement about checks performed. It is not a legal opinion and confers nothing. -## 4. The clauses +## 4. Conformance language -Attribution convention follows AAR-REQ Table 2.1: each clause names who proposed it, and in brackets who refined it. +"States" in VR1–VR5 means **SHALL** for any report presented as a pass, and **SHOULD** for any report presented as an unresolved result. The consequence of not stating is stated once here and relied on below: a report that omits a required statement cannot be graded above **D** by the next party, because the next party has nothing to check it against (VR5). No other consequence is claimed. -### VR1 — Evidence grade is a property of the verifier's relation to the evidence, not of the record +## 5. The clauses -*Proposed by @zhaoxinghua09-cell.* +Attribution follows AAR-REQ Table 2.1: each clause names who proposed it and, in brackets, who refined it. Where a clause restates an argument made elsewhere, the origin is named first and the restatement is not presented as an independent proposal. -A record can be complete and still be unusable by a given verifier, because usability depends on what that verifier can reach. The grade therefore cannot live inside the record: AAR-REQ §1.3 puts "how any value in a record is computed" out of scope, and a grade assigned by a producer would in any case be a claim the verifier has not checked. +### VR1 — Evidence grade describes the verifier's relation to the evidence, not the record -The verifier states one grade, from an ordered set, for the evidence it relied on: +*Argument from `FG-TIDA/themes#42` and `#43` (LGD-Eval, with refinements from @dakleyer); restated and applied to the AAR verification case by @zhaoxinghua09-cell.* -| Grade | Meaning | -|---|---| -| **A** | Checked by the verifier against a source it can reach without trusting the producer | -| **B** | Checked by a party with no relationship to the producer, named in the report | -| **C** | Corroborated indirectly — consistent with independent material, not checked at source | -| **D** | Asserted only — the verifier relied on the producer's statement | +A record can be complete and still be unusable by a given verifier, because usability depends on what that verifier can reach and what it has actually done. A grade assigned by the producer would therefore be a claim the verifier has not checked, and AAR-REQ §1.3 already puts "how any value in a record is computed" out of scope. -Two properties matter more than the number of grades. First, the grade is **about the claim, not the actor**: a strong grade on a claim does not make the actor's authorization valid, and VR1 must not be read as doing so (this keeps the separation argued in `FG-TIDA/themes#42`). Second, the set is **ordered but not arithmetic**: grades are not averaged across a chain, because a chain is only as strong as its weakest link and averaging would hide exactly that. +The verifier states one grade for each claim it relies on. The four grades are defined by two questions — *was the check performed, and against what?* — so that each grade has a route to being reached and each reported status has a grade: -### VR2 — The verifier states which moment it evaluated against, chosen from three +| Grade | Was it performed? | Against what? | +|---|---|---| +| **A** | Yes, by the verifier | A source the verifier can reach without trusting the producer | +| **B** | Yes, by a named party with no relationship to the producer | A source that party can reach without trusting the producer | +| **C** | Yes, by the verifier | Material independent of the producer, which the verifier can reach, but not at source | +| **D** | No | Nothing — the verifier relied on the producer's statement | + +Two things follow from the table and are worth stating because they are where it is usually got wrong. First, the grade is **about the claim, not the actor**: a grade A on a claim says nothing about whether the actor's authorization was valid, and VR1 must not be read as saying so — this is the separation argued in `FG-TIDA/themes#42`. Second, the grades are **ordered but not arithmetic**: they are not averaged across a chain, because averaging would hide the weakest hop, which is the only hop that matters. + +Where corroboration is claimed at grade C, this record does **not** set how much agreement counts as corroboration. That is open question 2. + +### VR2 — The verifier states which moment it evaluated against + +*Three moments: @OlehVasylenko2000 (12 November 2026), recorded in AAR-REQ §2.1 against open question 12, where they are proposed as a possible R7.2a and are therefore not yet text. The verifier-side obligation to name one, and to extend it beyond the standing question, is @zhaoxinghua09-cell's.* -*Proposed by @zhaoxinghua09-cell. The three moments are @OlehVasylenko2000's (12 November 2026) and are adopted here unchanged; the verifier-side obligation is this record's.* +R6 makes freshness a relation rather than a property. R4.2 already requires a relying party to record **which moment it used** — but only for the tier-2 standing question (R4.1–R4.3). A verdict on the other tier-1 classes, and a pass as a whole, carries no such requirement, and a pass whose moment is unstated is not reproducible by a later reader: they cannot tell which question was answered. -R6 makes freshness a relation, not a property. A relation needs two terms, and the second term — the moment the verifier used — is currently unstated on the verifier side. So the verifier names which of the three moments it used: +So the verifier names one of three moments for every verdict it reports: -1. **The moment of the act** — the moment the record describes. -2. **The moment the producer states for a claim** — carried under R7.2. -3. **An independently bounded moment** — which establishes only that the record existed no later than that time, and is not the freshness relation of R6. +1. **the moment of the act** — the moment the record describes; +2. **the moment the producer states for a claim** — carried under R7.2; +3. **an independently bounded moment** — which establishes only that the record existed no later than that time, and is not the freshness relation of R6. -A verdict evaluated at moment 1 and a verdict evaluated at moment 3 are **not the same object**, even when every field is identical, because the state that governs the act (R8.5) may have changed between them. A report that omits this is not reproducible by a later reader, who cannot tell which question was answered. +A verdict at moment 1 and a verdict at moment 3 are not the same object, even when every field is identical, because the effective state that governed the act (R8.5) may have changed between them. The increment over R4.2 is scope, not substance: R4.2 covers standing; VR2 covers every reported verdict, and requires the moment to be named rather than inferable. -The verifier also states whether rechecking was possible, taking R6's "means, if any, of rechecking" as given. Where no means exist, the report says so rather than leaving the field empty — an absent rechecking means is itself a finding. +The verifier also states whether rechecking was possible, taking R6's "means, if any, of rechecking" as given. Where no means exist, the report says so: an absent rechecking means is a finding, not a blank. ### VR3 — "Checkable" is not "checked" *Proposed by @zhaoxinghua09-cell.* -R7.3 requires tamper-evidence that a party with no relationship to the producer **can** check. That is a statement about capability, and it is satisfied by a record no one has ever examined. +R7.3 requires tamper-evidence that a party with no relationship to the producer **can** check. That is a capability, and it is satisfied by a record no one has ever examined. -So the verifier states, for each check it reports, which of three happened: +So each reported check carries its status: **performed by the verifier**, **performed by a named third party**, or **asserted**. The mapping to VR1 is deliberate and total: A and C are the verifier's own checks, B is a named third party's, D is asserted with nothing performed. -- **performed by the verifier** — the check was run, and the means are named; -- **performed by a named third party** — the check was run by a party with no relationship to the producer, and the report names it; -- **asserted** — the check was not run by anyone the verifier can name. +This matters because the failure it prevents is silent. A record that is checkable but unchecked produces a report that reads like a report from a check, and nothing in the record distinguishes the two. The status is what distinguishes them, and it is why a claim whose checks were asserted cannot carry a grade above D. -This distinction is load-bearing because the failure it prevents is silent. A record that is checkable but unchecked produces a verdict that looks like a verdict from a check, and nothing in the record distinguishes the two. Where a check was asserted rather than performed, the evidence grade (VR1) cannot exceed **D** for that claim, whatever the record's properties. - -### VR4 — A pass claims no violation found within a stated boundary, and nothing more +### VR4 — A pass claims no violation found within a stated boundary *Proposed by @zhaoxinghua09-cell, carrying forward the question posed in `FG-TIDA/themes` PR #28.* -A pass carries four things, and is incomplete without any of them: +A pass carries four statements, and is incomplete without any of them: -1. the **checks performed** (with VR3's status for each); +1. the **checks performed**, with VR3's status for each; 2. the **checks not performed** — named, not implied; 3. the **boundary** within which the checks ran; 4. the **moment used** (VR2). -With those four, a pass means exactly: *no violation was found within this stated boundary, using these checks, at this moment.* It does not mean the act was authorized, that the record is true, or that the delegation was sound. R7 says as much for carried claims — that a claim meets R7 does not make it true — and the same restraint has to hold one level up, at the verdict. +With those four, a pass means exactly: *no violation was found within this stated boundary, using these checks, at this moment.* It does not mean the act was authorized, that the record is true, or that the delegation was sound. AAR-REQ says as much one level down — "R7 does not make a claim true" — and the same restraint has to hold at the verdict. + +**A pass also has a floor.** Where every reported check is C or D, the report is *unresolved within the boundary*, not a pass, whatever else it states. Without the floor, a report in which nothing was checked against source could satisfy all four statements and present itself as a pass — which is the failure VR3 exists to name. -The property this buys is that a pass is **falsifiable by its own declaration**. A later party can withdraw it by showing a single check inside the declared boundary that was not performed, without having to relitigate the facts. A pass that does not declare its boundary cannot be withdrawn this way, and therefore cannot be relied on — it can only be believed. +The floor is what makes a pass **falsifiable by its own declaration**: a later party can withdraw it by showing one check inside the declared boundary that was not performed as reported, without relitigating the facts. A pass that declares no boundary cannot be withdrawn that way, and therefore cannot be relied on — only believed. -### VR5 — The verifier's own report is a claim, subject to R7 +AAR-REQ enumerates rejection classes and does not enumerate accept classes; R1 requires that a bare accept is not enough without saying what is. Whether "pass" and "unresolved within boundary" should become named classes in AAR-REQ is the group's decision, not this record's — it is open question 4. -*Proposed by @zhaoxinghua09-cell. This closes the recursion @OlehVasylenko2000 raised for R7.* +### VR5 — The verifier's report is a claim the next party has to grade -If the verifier's report matters to a later decision — and VR4 makes it the thing relied on — then the report is itself a carried claim and R7 applies to it unchanged: attribution (R7.1), a stated moment (R7.2), and tamper-evidence a party with no relationship to the producer can check (R7.3). +*Proposed by @zhaoxinghua09-cell. The recursion in R7 is @OlehVasylenko2000's; this applies it one level up.* -The practical consequence is narrow but sharp. **A verifier that does not disclose its own evidence basis is producing an unattributed claim**, and VR4's boundary cannot be read off it. This is not a demand that verifiers publish; it is a statement that an unpublished basis caps the grade the next party can assign. +VR4 makes the verifier's report the thing a later party relies on. When it is relied on, it is a carried claim in R7's sense, and R7's three properties apply by its own terms: attribution (R7.1), a stated moment (R7.2), and tamper-evidence a party with no relationship to the producer can check (R7.3). This record does not extend R7; it notes that R7 already reaches the report whenever the report is what a later party carries. -## 5. Interface table +The practical consequence is narrow. A verifier that does not disclose the basis of its own report is producing an unattributed claim, and the next party cannot read VR4's boundary off it. That does not cap the grade at D by itself — but it caps it at whatever the next party can establish without the basis, which in practice is D. This is not a demand that verifiers publish; it is a statement about what an unpublished basis costs the next party. -What the verifier reads from each record-side clause, and what it must state in return. The right-hand column is this record's contribution; the left-hand column is AAR-REQ's and is not restated. +## 6. Interface table + +What the verifier reads from each record-side clause, and what it must state in return. The left column is AAR-REQ's and is not restated; the right column is this record's. | AAR-REQ | What the verifier reads | What the verifier must state | |---|---|---| -| R1, R2 | the rejection class and the reported cause | VR1 grade for the claim about the cause; VR3 status per check | -| R3 | the four tier-1 classes are decidable from the record and the verifier's own state | which classes it actually decided, and which it could not | -| R4, R4.1–R4.3 | whether continued standing is checkable, and which moment the relying party used | VR2 moment, stated explicitly rather than inherited | -| R5 | negative and null assertions are claims like any other | VR1 grade per assertion; a null assertion is graded, never skipped | +| R1, R2 | the rejection class and the reported cause | VR1 grade for the claim about the cause; VR3 status per check. Where nothing was rejected, VR4 governs instead, since R1–R4 give no content to an accept | +| R3 | the four tier-1 classes are decidable from the record and the relying party's own state | which classes it actually decided, and which it could not | +| R4, R4.1–R4.3 | whether continued standing is checkable, and which moment the relying party used | VR2 moment, named rather than inherited, extended from standing to every verdict | +| R5 | negative and null assertions are claims like any other (proposed by @cgorlla; null case @OlehVasylenko2000, refined by @zhaoxinghua09-cell) | VR1 grade per assertion; a null assertion is graded, never skipped | | R6 | the moment the assertion was true, and the means of rechecking | VR2 moment used; whether rechecking was possible, and if not, that it was not | -| R7, R7.1–R7.3 | attribution, stated moment, tamper-evidence | VR3 performed / by a named third party / asserted; VR5 for its own report | +| R7, R7.1–R7.3 | attribution, stated moment, tamper-evidence | VR3 status per check; VR5 for its own report | | Absence and silence | which of the three states holds | which state was established, and the scope and author of any bounded search | -| LR1–LR5 | which lineage case holds | the case reported, with the evidence grade for it | -| R8, R8.1–R8.5 | act, execution, supported relation to a logical agent, effective state | which of the five it established, and at which moment (VR2) | - -## 6. Three boundary cases +| LR1–LR5 | which lineage case holds | the case reported, with its evidence grade | +| R8, R8.1–R8.5 | four things across five sub-clauses: the act, the execution that performed it, that execution's supported relation to a logical agent where one exists, and the effective state that governed the act at that time | which of the four it established, at which moment (VR2), with the grade for each | -These are the cases where a verifier most easily states more than it checked. +## 7. Three boundary cases -**Case A — checkable, uncheckable-by-this-verifier.** The record is complete and tamper-evident; the verifier has no relationship to the producer and no route to recheck. Verdict: *no violation found within boundary, grade D, rechecking not possible.* Not a pass in the ordinary sense, and not a rejection either — R3's classes do not apply, because nothing was found to be wrong. +These are the cases where a report most easily states more than was checked. -**Case B — right record, wrong moment.** Every tier-1 check passes at the moment of the act; standing was withdrawn before the verifier looked. Reporting a pass at moment 3 while presenting it as a statement about moment 1 is the error VR2 exists to prevent. The verifier's verdict is about the moment it names, and the report must let a reader see which. +**Case A — checkable, but not checked by this verifier.** The record is complete and tamper-evident; the verifier has no relationship to the producer and no route to a source. Every check it can make is C or D, so by VR4's floor the output is **unresolved within the boundary**, not a pass. It is also not a rejection: R1–R4's classes do not apply, because nothing was found to be wrong. The report names both facts and leaves the result unclassified, which is a gap in AAR-REQ's class set rather than in this record (open question 4). -**Case C — absence is not silence.** No record at all, a bounded search that returned nothing, and a reference that was established are three different findings. Only the second carries a scope and an author; the third is a positive finding. Collapsing them yields a verdict that reads as "nothing wrong" when the correct reading is "nothing was looked at". +**Case B — right record, wrong moment.** Every tier-1 check passes at the moment of the act; standing was withdrawn before the verifier looked. Reporting a pass at moment 3 while presenting it as a statement about moment 1 is the error VR2 exists to prevent. The verdict is about the moment named, and the report must let a reader see which — otherwise the reader supplies the wrong one. -## 7. Open questions +**Case C — absence is not silence.** No record at all, a bounded search that returned nothing, and a reference that was established are three different findings. Only the second carries a scope and an author; the third is a positive finding. A report that collapses them reads as "nothing wrong" when the correct reading differs in each case: nothing was recorded, nothing was found within a stated search, or the thing was established. -Carried openly, in the style of AAR-REQ §2.1. +## 8. Open questions -1. **Is four grades the right cut?** Three would force B and C together, which hides the difference between a check performed elsewhere and a corroboration. Five invites distinctions no one can apply consistently. *(@zhaoxinghua09-cell)* -2. **Should VR2's moment be a required field in the report, or only a required statement?** A field is machine-checkable; a statement is cheaper and survives formats the group has not settled. *(@zhaoxinghua09-cell)* -3. **Does VR5 bind all verifier output, or only published verification reports?** Binding all output is the stronger position and the harder one to meet, and I would rather the group choose it knowingly than have it arrive by implication. *(@zhaoxinghua09-cell)* -4. **Where does the boundary in VR4 come from?** If the group names a minimum boundary somewhere, this record should cite it rather than let each verifier declare its own. *(@zhaoxinghua09-cell)* +1. **Should the grade set and the status set be one table or two?** They are defined from the same two questions here, which makes them total but couples them: a fifth status would force a fifth grade. *(@zhaoxinghua09-cell)* +2. **How much agreement counts as corroboration at grade C?** This record does not set a threshold, and §3 excludes thresholds by scope. If the group wants C to be operable across implementers, the threshold has to come from somewhere; naming that somewhere is the open part. *(@zhaoxinghua09-cell)* +3. **Does VR5 bind all verifier output, or only published verification reports?** Binding all output is stronger and harder to meet, and I would rather the group choose it knowingly than have it arrive by implication. *(@zhaoxinghua09-cell)* +4. **Should "pass" and "unresolved within boundary" become named classes in AAR-REQ?** VR4's floor produces a result that R1–R4's class set does not name. Either the class set grows, or the report leaves it unclassified as Case A does. *(@zhaoxinghua09-cell)* +5. **Where does the boundary in VR4 come from?** If a minimum boundary is named somewhere in FG-TIDA's output, this record should cite it rather than let each verifier declare its own. *(@zhaoxinghua09-cell)* -## 8. Provenance +## 9. Provenance - AAR-REQ 0.2, web — https://tyche-dev.github.io/tida-aar-req/ - AAR-REQ 0.2, PDF — https://tyche-dev.github.io/tida-aar-req/specification.pdf - Originating thread — https://github.com/FG-TIDA/themes/issues/46 - Verifier-side charter — https://github.com/FG-TIDA/themes/pull/28 -- Evidence grading (separating evidence quality from authorization validity) — https://github.com/FG-TIDA/themes/issues/42 +- Evidence grading, separating evidence quality from authorization validity — https://github.com/FG-TIDA/themes/issues/42 - Evidence-graded authorization verdicts — https://github.com/FG-TIDA/themes/issues/43 - Conformance verdict scope, handed off by AAR-REQ §1.3 — https://github.com/FG-TIDA/themes/issues/6 - Author identity — https://orcid.org/0009-0001-0512-1237 - Account relationship and conflict-of-interest disclosure — https://github.com/LGD-Eval/evaluation-protocol/blob/main/CONFLICT_OF_INTEREST_DISCLOSURE.md -## 9. Limitations +## 10. Limitations -- **The clause set has had one author.** VR1–VR5 are coherent with each other by construction, which is not the same as being right. Two of the five (VR3, VR4) rest on a distinction that is easy to state and hard to apply; a reviewer should test them against a case they already know the answer to. -- **No implementation accompanies this version.** The reference implementation covers the evidence-gated state machine, not the verifier report format these clauses describe. A worked example set is the obvious next increment and is not claimed here. +- **Only the clause table and the scope chapter were read.** AAR-REQ states that its clause text "stands in 2 comments posted on FG-TIDA/themes#1" and that where the table differs, those comments govern. This record was written against the table in §2 and the scope in §1, and has **not** been reconciled against those two comments. Where they differ, they win, and VR2 in particular should be re-read against them. +- **The three moments are a proposal, not text.** AAR-REQ §2.1 records them against open question 12 as a possible R7.2a — "if the proposed R7.2a becomes text". VR2 adopts them as the set to choose from; if the group settles differently, VR2's enumeration changes with it. +- **The grade set is single-author and untested.** VR1's four grades are coherent by construction, which is not the same as being right. C is the one to test: it is the only grade whose boundary is a matter of degree, and open question 2 is unresolved. +- **No consequence semantics beyond §4.** The clauses say what a report must state and what the next party can do without it. They do not say what happens to a verifier that states nothing — that is AAR-REQ's conformance question, which §1.3 moves elsewhere. +- **No implementation accompanies this version.** The reference implementation covers the evidence-gated state machine, not the report format these clauses describe. A worked example set is the obvious next increment and is not claimed here. - **The material this record draws on has not been through peer review.** `FG-TIDA/themes#42` and `#43` are proposals under discussion, not settled positions, and VR1 is a restatement of their separation argument rather than independent evidence for it. -- **Drafted with AI assistance** under author direction; the author is responsible for every claim above. +- **Drafted with AI assistance** under author direction, and red-teamed before posting; the author is responsible for every claim above. -## 10. Version history +## 11. Version history | Version | Date | Change | |---|---|---| -| 0.1 | 2026-11-16 | First draft, posted for attributed review on `FG-TIDA/themes#46` | +| 0.1 | 2026-11-16 | First draft | +| 0.2 | 2026-11-16 | Position restated as a refinement of R1's unstated accept side rather than an unoccupied space; VR2 premise corrected against R4.2 (scope extension, not a blank); verifier defined against relying party; VR1 attribution corrected to #42/#43; grades redefined so each has a route to being reached and each status maps to a grade; VR4 given a floor, and Case A reclassified as unresolved; R8 count corrected; conformance language added; limitations extended to the posted-comments risk and the proposal status of the three moments | --- @@ -180,14 +204,14 @@ Carried openly, in the style of AAR-REQ §2.1. > 英文为权威文本;以下为逐条中文要点,供中文读者对照,非逐句对译。 -**本记录的位置。** AAR-REQ 在 Out of Scope 里明写「一致性判定能主张什么、怎么检查」不是它的事,并转给 `themes#6` task 3。本记录就站在这个被让出来的位置上,从验证方一侧回答:一份满足 R1–R8 的记录,一个当时不在场的人能从中确立什么、必须自己声明什么、以及「通过」到底允许主张什么。**不给 R1–R8 增加任何内容。** +**这份记录站的位置。** AAR-REQ §1.2 把「依赖方要报什么」写进了自己的范围,R1–R4 也逐个列了**拒绝**的类别。真正空着的不是整块,而是一条:R1 说「光说接受或拒绝不够」,但**从没说一个不 bare 的「接受」要带什么**;§1.3 又把「一致性判定能主张什么」让给了 themes#6 task 3。本文补前一条,不碰后一条。**对 R1–R8 一字不加、一字不改。** | 条款 | 中文要点 | |---|---| -| **VR1** | 证据等级是验证方与证据之间关系的属性,不是记录的属性。记录再完整,对某个验证方也可能不可用。四档:A 自己直接核到源头;B 由与产出方无关的具名第三方核过;C 间接佐证;D 仅有断言。等级针对**主张**,不针对**行为人**——等级高不等于授权有效。四档有序但不可做算术平均,链的强度等于最弱一环。 | -| **VR2** | 验证方必须声明自己用的是哪个「时刻」,三选一:①行为发生时 ②产出方为某主张声明的时刻 ③独立可界定的时刻(只能证明记录不晚于某时,不是 R6 的时效关系)。同一份记录、同一个字段,用不同时刻得出的结论**不是同一个东西**,因为 R8.5 的有效状态可能已经变了。还要声明「能否复核」——没有复核手段本身就是一个结论,不能留空。 | -| **VR3** | 「可检查」不等于「已检查」。R7.3 只保证「与产出方无关的人**能**检查」,一份谁都没看过的记录同样满足它。所以每项检查都要标状态:验证方自己做的/具名第三方做的/仅有断言。这条重要,因为失效是静默的——可检查但未检查的记录,产出的结论看起来跟检查过的一样。仅有断言时,该主张的证据等级最高只能到 D。 | -| **VR4** | 通过 = 在声明边界内、用声明的检查、在声明的时刻,未发现违规。**仅此而已**,不等于行为获授权、记录为真、委托成立。必须带四样:做了什么检查、**没做什么检查**(要点名)、边界、所用时刻。这样换来的性质是:**通过可以被它自己的声明推翻**——后人只要指出边界内有一项检查没做,就能撤回它,不必重打事实。不声明边界的通过,无法被这样撤回,因此不能被依赖,只能被相信。 | -| **VR5** | 验证方自己的报告也是主张,R7 照样适用:归属、声明时刻、与产出方无关者可校验的防篡改。推论很窄但很利:**不披露自身证据基础的验证方,产出的是一份无归属主张**,VR4 的边界无从读起。这不是要求验证方必须公开,而是说未公开的基础会压低下一个当事人能给到的等级。 | +| **VR1** | 证据等级描述的是「验证方与证据的关系」,不是记录自带的字段。四档由两个问题定义——**查没查、对什么查**——所以每一档都走得通、每个状态都有对应等级:A 自己查、对能直达且无需信任产出方的源;B 与产出方无关的具名方查、对同类源;C 自己查、对独立于产出方的材料但不在源头;D 没查,只有产出方的说法。等级针对**主张**不针对**行为人**;四档有序但**不可平均**,平均会把最弱那一跳藏起来。 | +| **VR2** | R6 说时效是关系不是属性;R4.2 **已经**要求依赖方记录用了哪个时刻——但只管 tier-2 的资格问题。本条的增量是**范围**不是内容:把「点名一个时刻」从资格问题推广到每一份判定,并把三时刻(行为时/产出方声明时/独立可界定时)作为可选集合。三时刻是 @OlehVasylenko2000 的**提案**(可能成为 R7.2a),不是既成文本,这一点正文已写明。 | +| **VR3** | 「可检查」不等于「已检查」。R7.3 只保证「与产出方无关的人**能**查」,一份没人看过的记录同样满足它。所以每项检查标状态:验证方自做/具名第三方做/仅有断言。与 VR1 是全覆盖映射:A、C 是自做,B 是第三方,D 是断言。 | +| **VR4** | 通过 = 在声明边界内、用声明的检查、在声明的时刻,未发现违规。必带四样:做了什么、**没做什么**(要点名)、边界、时刻。**另有一条地板**:若全部检查都是 C 或 D,结论是「边界内未决」,不是通过——否则「什么都没对源查过」的报告也能凑齐四条伪装成通过,而这正是 VR3 要挡的失效。有了地板,通过才能**被它自己的声明推翻**。 | +| **VR5** | 验证方的报告是下一方要评的东西;一旦被依赖,它就是 R7 意义上的 carried claim,R7 的三项(归属/时刻/防篡改)按其自身条款即已适用。本条不扩张 R7,只指出它本来就够得着报告。推论很窄:不披露自身依据的验证方,产出的是一份归属不明的主张,下一方读不出 VR4 的边界。 | -**三个边界情形。** A:记录可检查、但本验证方查不到——结论是「边界内未发现违规、等级 D、无法复核」,既不是通过也不是拒绝(R3 的四类不适用,因为没发现错)。B:记录对、时刻错——行为时全部通过,验证方来看时资格已撤回;把时刻③的结论当时刻①的结论报,正是 VR2 要挡的错。C:无记录 ≠ 静默——「根本没有记录」「有界检索查无」「引用已确立」是三种不同结论,只有第二种带检索范围和作者;混为一谈会把「什么都没查到」读成「什么都没问题」。 +**三个边界情形。** A:记录可检查但本验证方查不到——按地板判「边界内未决」,既不是通过也不是拒绝(R1–R4 没有对应类,这是 AAR-REQ 类集的缺口,列为开放问题 4)。B:记录对、时刻错——行为时全通过、验证方来看时资格已撤回;把时刻③的结论当时刻①报,正是 VR2 要挡的。C:无记录 ≠ 静默——「没有记录」「有界检索查无」「引用已确立」是三种结论,只有第二种带检索范围和作者。 From f6a4a01d8c1c4d08324325c53faa49f865fe3e3f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?LGD-Eval=20=E2=80=94=20Independent=20AI=20Evaluation?= Date: Mon, 16 Nov 2026 14:16:56 +0800 Subject: [PATCH 3/4] Revise verifier-side interface record to 0.3 Positioning restated as a refinement of R1's unstated accept side, not an unoccupied space. VR2 premise corrected against R4.2: scope extension, not a blank. Verifier defined against relying party. VR1 attribution moved to #42/#43. Grades redefined so each is reachable and each status maps to one. VR4 given a floor; Case A reclassified as unresolved. R8 count corrected. Conformance language added. Limitations extended. Assisted-by: WorkBuddy --- FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md | 170 ++++++++++++------------- 1 file changed, 80 insertions(+), 90 deletions(-) diff --git a/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md b/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md index 5df6bd5..6489254 100644 --- a/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md +++ b/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md @@ -6,8 +6,8 @@ |---|---| | Author | Xinghua Zhao (Steven Zhao), independent expert, GitHub `@zhaoxinghua09-cell` | | Offered in | Individual-expert capacity, as stated on `FG-TIDA/themes#46` on 29 October 2026 | -| Hosted in | `LGD-Eval/evaluation-protocol` — the public home of the evaluation material this record draws on. Hosting is a publication venue, not an authorship claim: the record is offered in individual-expert capacity, and the account relationship is disclosed below. | -| Version | 0.2 — 16 November 2026 | +| Hosted in | `LGD-Eval/evaluation-protocol` — the public home of the evaluation material this record draws on. Hosting is a publication venue, not an authorship claim; the account relationship is linked in §9. | +| Version | 0.3 — 16 November 2026 | | Status | Draft for attributed review before the first in-person FG-TIDA meeting (Paris, 30 November – 1 December 2026) | | Linked work item | AAR-REQ 0.2 — https://tyche-dev.github.io/tida-aar-req/ | | Licence | CC BY 4.0 | @@ -16,158 +16,146 @@ ## 0. Terms -Two terms are used throughout and are defined here so that the interface below is readable against AAR-REQ's own text. - **Relying party.** As in AAR-REQ: the party that acts on a claim about an agent after the fact. -**Verifier.** The relying party itself, or an agent acting for it, when it evaluates a record and states a result. This record does not introduce a second role: wherever AAR-REQ binds a relying party, the obligation falls on the verifier, and wherever AAR-REQ's text says "relying party" this record's clauses are satisfied by the same party under the other name. The name is used because the clauses are about the act of verifying, not about the act of relying. +**Verifier.** The relying party itself, or an agent acting for it, when it evaluates a record and states a result. This record introduces no second role: wherever AAR-REQ binds a relying party, the obligation falls on the verifier, and the name is used because these clauses are about the act of verifying rather than the act of relying. + +**Record-side and verifier-side.** Record-side requirements constrain what a record must carry. Verifier-side requirements constrain what a party must state when it reports a result from a record. R1–R8 are record-side; the clauses below are verifier-side. -**Record-side / verifier-side.** Record-side requirements constrain what a record must carry. Verifier-side requirements constrain what a party must state when it reports a result from a record. R1–R8 are record-side. The clauses below are verifier-side. +**Source.** A party, register or artefact against which a check can be run **without taking the producer's word for the result**. The record is the object checked, not a source: reading it establishes what it says, not whether what it says is true. This definition matters because it is the only thing standing between a verification and a recital. ## 1. Where this record sits -AAR-REQ §1.2 puts two things in scope: requirements on the record, and **"what a relying party reports from it"**. The second is not left open in general — R1–R4 enumerate the rejection classes a relying party reports, and R1 states that a bare accept or reject is not enough. +AAR-REQ §1.2 puts two things in scope: requirements on the record, and what a relying party reports from it. R1 requires that a bare accept or reject is not enough. R1–R4 then give the rejection classes content; **the accept side is required to be non-bare but is nowhere given content** — no clause states which checks a report of acceptance covers, or what bounds them. -What is left open is narrower, and it is the gap this record works in: +Two neighbouring clauses come close and stop short, and it is worth being exact about how: -- **The rejection side is enumerated; the accept side is not.** R1 forbids a bare accept and R3–R4 give four tier-1 and one tier-2 rejection class with contents. No clause says what a non-bare accept must carry, or what it is allowed to assert. R1 creates the obligation without supplying its content. -- **§1.3 hands away the verdict question.** "What a conformance verdict asserts and how it is checked" is out of scope and assigned to `FG-TIDA/themes#6`, task 3. +- **R4.4** makes a carried standing answer a claim, subject to R7. So a carried answer already owes attribution, a stated moment and tamper-evidence. +- **R7** governs the properties of any carried claim. -So the position is not an empty space: it is an obligation with no content, next to a question that has been moved elsewhere. This record supplies content for the first and stays clear of the second — it says nothing about conformance verdicts on an artefact, only about the report a party makes after verifying a record of an act. +Neither says **what was checked** or **within what boundary** — R7 regulates how a claim is carried, not what an acceptance asserts. That distinction is the whole subject of this record, and it is narrower than "a gap in AAR-REQ": the properties are supplied, the content is not. -This is a refinement of AAR-REQ's existing verifier-side requirement, not an addition to the record-side set. **It adds nothing to R1–R8 and changes nothing in them.** Where a clause depends on a record-side requirement, it takes that requirement as given and names it. +§1.3 separately puts "what a conformance verdict asserts and how it is checked" out of scope and assigns it to `FG-TIDA/themes#6`, task 3. This record does not enter that question: it is about a verdict on a **record of an act**, not on an artefact. + +**Nothing here adds to or alters R1–R8.** Every clause constrains the verifier's report. ## 2. Scope -- **VR1** — the grade of the evidence the verifier relied on; +- **VR1** — what the verifier reports about each check: who performed it, and against what; - **VR2** — the moment the verifier evaluated against; -- **VR3** — whether each reported check was performed, and by whom; -- **VR4** — what a pass may claim, bounded by what was actually checked; -- **VR5** — the verifier's own report as a claim the next party has to grade. +- **VR3** — what a pass may claim, bounded by the checks reported; +- **VR4** — the verifier's report as a claim the next party has to assess. ## 3. Out of scope -Stated so that overlap is decidable rather than implicit: - -- **Decision procedure, thresholds, evidence weighting.** How a verifier reaches a class is R1–R4's subject and is not prescribed here. -- **Encodings, schemas, field names, media types, test vectors.** As in AAR-REQ §1.3. VR1's four grades and VR3's three states are **descriptions of relations, not enumerations to be encoded**: a verifier may state them in its own words, and no field name is proposed. +- **Decision procedure, thresholds, evidence weighting.** AAR-REQ §1.3 puts how a relying party reaches a class out of scope, and this record does not bring it back in. +- **Encodings, schemas, field names, media types, test vectors.** As in §1.3. The two components in VR1 are **descriptions of what happened, not fields to be implemented**, and no encoding is proposed. - **Conformance verdicts on an artefact** (`FG-TIDA/themes#6`, task 3). - **Where authority originates and what may be conferred** (`FG-TIDA/themes#5`). - **Legal effect.** A pass is a statement about checks performed. It is not a legal opinion and confers nothing. ## 4. Conformance language -"States" in VR1–VR5 means **SHALL** for any report presented as a pass, and **SHOULD** for any report presented as an unresolved result. The consequence of not stating is stated once here and relied on below: a report that omits a required statement cannot be graded above **D** by the next party, because the next party has nothing to check it against (VR5). No other consequence is claimed. +Where a clause says the verifier **states**, **reports**, **names** or **carries**, that is **SHALL** for any report presented as a pass, and **SHOULD** for any report presented as unresolved. The consequence of omission is stated once, here: a report that omits a required statement cannot be assessed by the next party above the level that party can establish without it, which under VR1 is "not performed, no source" — the lowest pair. Two further consequences follow from the clauses themselves rather than from this section: an asserted check cannot support a pass (VR3), and a pass whose boundary is undeclared cannot be withdrawn and therefore cannot be relied on (VR3). ## 5. The clauses Attribution follows AAR-REQ Table 2.1: each clause names who proposed it and, in brackets, who refined it. Where a clause restates an argument made elsewhere, the origin is named first and the restatement is not presented as an independent proposal. -### VR1 — Evidence grade describes the verifier's relation to the evidence, not the record +### VR1 — Each check is reported by who performed it and against what -*Argument from `FG-TIDA/themes#42` and `#43` (LGD-Eval, with refinements from @dakleyer); restated and applied to the AAR verification case by @zhaoxinghua09-cell.* +*Proposed by @zhaoxinghua09-cell. The separation of evidence quality from authorization validity is `FG-TIDA/themes#42`'s argument, restated in `#43` as LGD-Eval with @dakleyer's refinements; this clause applies it to the AAR case and does not claim the argument.* -A record can be complete and still be unusable by a given verifier, because usability depends on what that verifier can reach and what it has actually done. A grade assigned by the producer would therefore be a claim the verifier has not checked, and AAR-REQ §1.3 already puts "how any value in a record is computed" out of scope. +A record can be complete and still support nothing, because what a report supports depends on what the verifier did and what it did it against. Both parts are reported, as a pair — not as a single rank: -The verifier states one grade for each claim it relies on. The four grades are defined by two questions — *was the check performed, and against what?* — so that each grade has a route to being reached and each reported status has a grade: +| Component | Values | +|---|---| +| **Performed by** | the verifier / a named party / not performed (asserted) | +| **Against** | source (per §0) / corroborated material, not at source / none | -| Grade | Was it performed? | Against what? | -|---|---|---| -| **A** | Yes, by the verifier | A source the verifier can reach without trusting the producer | -| **B** | Yes, by a named party with no relationship to the producer | A source that party can reach without trusting the producer | -| **C** | Yes, by the verifier | Material independent of the producer, which the verifier can reach, but not at source | -| **D** | No | Nothing — the verifier relied on the producer's statement | +Nine combinations, each nameable, and no ordering is claimed between them. That is deliberate: an ordering would have to decide, for instance, whether a named party's check against source outranks the verifier's own check against corroborating material, and any answer to that is a policy choice the group has not made (open question 1). + +Two points about what the pair does **not** say: -Two things follow from the table and are worth stating because they are where it is usually got wrong. First, the grade is **about the claim, not the actor**: a grade A on a claim says nothing about whether the actor's authorization was valid, and VR1 must not be read as saying so — this is the separation argued in `FG-TIDA/themes#42`. Second, the grades are **ordered but not arithmetic**: they are not averaged across a chain, because averaging would hide the weakest hop, which is the only hop that matters. +- **It does not grade the actor.** A report about a claim says nothing about whether the actor's authorization was valid. Keeping the two apart is `#42`'s argument, and VR1 reports only the first. +- **It does not grade the record.** The values describe the verifier's relation to the evidence, not a property the record carries. A grade assigned by a producer would describe how the producer evaluated the record, which is a different object and is not what a later party needs. -Where corroboration is claimed at grade C, this record does **not** set how much agreement counts as corroboration. That is open question 2. +Corroboration is deliberately weaker than a check at source: corroborating material cannot distinguish an independent source from a shared upstream one, which is the case `FG-TIDA/themes#60` raises. Where a report rests on corroboration, it says so, and the next party can see that the independence question is open. ### VR2 — The verifier states which moment it evaluated against -*Three moments: @OlehVasylenko2000 (12 November 2026), recorded in AAR-REQ §2.1 against open question 12, where they are proposed as a possible R7.2a and are therefore not yet text. The verifier-side obligation to name one, and to extend it beyond the standing question, is @zhaoxinghua09-cell's.* +*Three moments: @OlehVasylenko2000 (12 November 2026), recorded in AAR-REQ §2.1 against open question 12 as a **possible** R7.2a and therefore not yet text. The obligation to name one for every verdict, rather than for the standing question alone, is @zhaoxinghua09-cell's.* -R6 makes freshness a relation rather than a property. R4.2 already requires a relying party to record **which moment it used** — but only for the tier-2 standing question (R4.1–R4.3). A verdict on the other tier-1 classes, and a pass as a whole, carries no such requirement, and a pass whose moment is unstated is not reproducible by a later reader: they cannot tell which question was answered. +R6 makes freshness a relation, not a property. R4.2 already requires a relying party to record which moment it used — for the tier-2 standing question (R4.1–R4.3). Outside that question the requirement does not run, and a verdict whose moment is unstated cannot be reproduced by a later reader: they cannot tell which question was answered. -So the verifier names one of three moments for every verdict it reports: +So the verifier names one of three: 1. **the moment of the act** — the moment the record describes; 2. **the moment the producer states for a claim** — carried under R7.2; 3. **an independently bounded moment** — which establishes only that the record existed no later than that time, and is not the freshness relation of R6. -A verdict at moment 1 and a verdict at moment 3 are not the same object, even when every field is identical, because the effective state that governed the act (R8.5) may have changed between them. The increment over R4.2 is scope, not substance: R4.2 covers standing; VR2 covers every reported verdict, and requires the moment to be named rather than inferable. +The increment over R4.2 is scope, not substance: R4.2 covers standing, VR2 covers every reported verdict. A verdict at moment 1 and one at moment 3 are not the same object even when every field is identical, because the effective state that governed the act (R8.5) may have changed between them. -The verifier also states whether rechecking was possible, taking R6's "means, if any, of rechecking" as given. Where no means exist, the report says so: an absent rechecking means is a finding, not a blank. +The verifier also reports whether rechecking was possible, taking R6's "means, if any, of rechecking" as given. Where no means exist, the report says so: an absent means of rechecking is a finding, not a blank. -### VR3 — "Checkable" is not "checked" - -*Proposed by @zhaoxinghua09-cell.* - -R7.3 requires tamper-evidence that a party with no relationship to the producer **can** check. That is a capability, and it is satisfied by a record no one has ever examined. - -So each reported check carries its status: **performed by the verifier**, **performed by a named third party**, or **asserted**. The mapping to VR1 is deliberate and total: A and C are the verifier's own checks, B is a named third party's, D is asserted with nothing performed. - -This matters because the failure it prevents is silent. A record that is checkable but unchecked produces a report that reads like a report from a check, and nothing in the record distinguishes the two. The status is what distinguishes them, and it is why a claim whose checks were asserted cannot carry a grade above D. - -### VR4 — A pass claims no violation found within a stated boundary +### VR3 — A pass claims no violation found within a stated boundary *Proposed by @zhaoxinghua09-cell, carrying forward the question posed in `FG-TIDA/themes` PR #28.* A pass carries four statements, and is incomplete without any of them: -1. the **checks performed**, with VR3's status for each; +1. the **checks performed**, with VR1's pair for each; 2. the **checks not performed** — named, not implied; 3. the **boundary** within which the checks ran; 4. the **moment used** (VR2). -With those four, a pass means exactly: *no violation was found within this stated boundary, using these checks, at this moment.* It does not mean the act was authorized, that the record is true, or that the delegation was sound. AAR-REQ says as much one level down — "R7 does not make a claim true" — and the same restraint has to hold at the verdict. +A pass then means no more than this: *no violation was found within this stated boundary, using these checks, at this moment.* It does not mean the act was authorized, that the record is true, or that the delegation was sound. AAR-REQ says as much one level down — "R7 does not make a claim true" — and the same restraint holds at the verdict. -**A pass also has a floor.** Where every reported check is C or D, the report is *unresolved within the boundary*, not a pass, whatever else it states. Without the floor, a report in which nothing was checked against source could satisfy all four statements and present itself as a pass — which is the failure VR3 exists to name. +**A pass also has a floor.** At least one reported check must be performed, against source. Without the floor, a report in which every check was corroborated or asserted could state all four things and present itself as a pass — and nothing in the report would distinguish it from a report in which something was actually checked. The floor is why corroboration is weaker than source in VR1: corroboration alone cannot carry a pass, because it cannot establish that the evidence is independent of the producer. -The floor is what makes a pass **falsifiable by its own declaration**: a later party can withdraw it by showing one check inside the declared boundary that was not performed as reported, without relitigating the facts. A pass that declares no boundary cannot be withdrawn that way, and therefore cannot be relied on — only believed. +The floor is what makes a pass **withdrawable by its own declaration**: a later party can withdraw it by showing one check inside the declared boundary that was not performed as reported, without relitigating the facts. A pass that declares no boundary cannot be withdrawn that way, and therefore cannot be relied on — only believed. -AAR-REQ enumerates rejection classes and does not enumerate accept classes; R1 requires that a bare accept is not enough without saying what is. Whether "pass" and "unresolved within boundary" should become named classes in AAR-REQ is the group's decision, not this record's — it is open question 4. +AAR-REQ names rejection classes and not accept classes. Whether "pass" and "unresolved within boundary" should become named classes is the group's decision (open question 3). -### VR5 — The verifier's report is a claim the next party has to grade +### VR4 — The verifier's report is a claim the next party has to assess *Proposed by @zhaoxinghua09-cell. The recursion in R7 is @OlehVasylenko2000's; this applies it one level up.* -VR4 makes the verifier's report the thing a later party relies on. When it is relied on, it is a carried claim in R7's sense, and R7's three properties apply by its own terms: attribution (R7.1), a stated moment (R7.2), and tamper-evidence a party with no relationship to the producer can check (R7.3). This record does not extend R7; it notes that R7 already reaches the report whenever the report is what a later party carries. +VR3 makes the verifier's report the thing a later party relies on. Once it is relied on, it is carried, and R7 reaches it by its own terms — attribution (R7.1), a stated moment (R7.2), tamper-evidence a party with no relationship to the producer can check (R7.3). This record does not extend R7. It notes that R7 already covers the case, and that a verifier which has not disclosed the basis of its report has produced a claim the next party cannot assess above the lowest VR1 pair. -The practical consequence is narrow. A verifier that does not disclose the basis of its own report is producing an unattributed claim, and the next party cannot read VR4's boundary off it. That does not cap the grade at D by itself — but it caps it at whatever the next party can establish without the basis, which in practice is D. This is not a demand that verifiers publish; it is a statement about what an unpublished basis costs the next party. +This is not a demand that verifiers publish. It is a statement about what an undisclosed basis costs the next party. ## 6. Interface table -What the verifier reads from each record-side clause, and what it must state in return. The left column is AAR-REQ's and is not restated; the right column is this record's. +What the verifier reads from each record-side clause, and what it reports in return. The left column is AAR-REQ's; the right column is this record's. -| AAR-REQ | What the verifier reads | What the verifier must state | +| AAR-REQ | What the verifier reads | What the verifier reports | |---|---|---| -| R1, R2 | the rejection class and the reported cause | VR1 grade for the claim about the cause; VR3 status per check. Where nothing was rejected, VR4 governs instead, since R1–R4 give no content to an accept | -| R3 | the four tier-1 classes are decidable from the record and the relying party's own state | which classes it actually decided, and which it could not | +| R1, R2 | the rejection class and the reported cause | VR1 pair for the claim about the cause. Where nothing was rejected, VR3 governs, since R1–R4 give no content to an accept | +| R3 | the four tier-1 classes are decidable from the record and the relying party's own state | which classes it decided, and which it could not | | R4, R4.1–R4.3 | whether continued standing is checkable, and which moment the relying party used | VR2 moment, named rather than inherited, extended from standing to every verdict | -| R5 | negative and null assertions are claims like any other (proposed by @cgorlla; null case @OlehVasylenko2000, refined by @zhaoxinghua09-cell) | VR1 grade per assertion; a null assertion is graded, never skipped | -| R6 | the moment the assertion was true, and the means of rechecking | VR2 moment used; whether rechecking was possible, and if not, that it was not | -| R7, R7.1–R7.3 | attribution, stated moment, tamper-evidence | VR3 status per check; VR5 for its own report | +| R4.4, R7 | the properties a carried answer owes | satisfied by VR4's application of R7 to the report itself | +| R5 | negative and null assertions are claims like any other (proposed by @cgorlla; null case @OlehVasylenko2000, refined by @zhaoxinghua09-cell) | VR1 pair per assertion; a null assertion is reported, never skipped | +| R6 | the moment the assertion was true, and the means of rechecking | VR2 moment used, and whether rechecking was possible | | Absence and silence | which of the three states holds | which state was established, and the scope and author of any bounded search | -| LR1–LR5 | which lineage case holds | the case reported, with its evidence grade | -| R8, R8.1–R8.5 | four things across five sub-clauses: the act, the execution that performed it, that execution's supported relation to a logical agent where one exists, and the effective state that governed the act at that time | which of the four it established, at which moment (VR2), with the grade for each | +| LR1–LR5 | which lineage case holds | the case reported, with its VR1 pair | +| R8, R8.1–R8.5 | four things across five sub-clauses: the act, the execution that performed it, that execution's supported relation to a logical agent where one exists, and the effective state that governed the act at that time | which of the four it established, at which moment (VR2), with a VR1 pair for each | ## 7. Three boundary cases -These are the cases where a report most easily states more than was checked. - -**Case A — checkable, but not checked by this verifier.** The record is complete and tamper-evident; the verifier has no relationship to the producer and no route to a source. Every check it can make is C or D, so by VR4's floor the output is **unresolved within the boundary**, not a pass. It is also not a rejection: R1–R4's classes do not apply, because nothing was found to be wrong. The report names both facts and leaves the result unclassified, which is a gap in AAR-REQ's class set rather than in this record (open question 4). +**Case A — the record is readable but nothing can be checked against source.** The verifier can read the record and can reach no source. Every check it reports is corroborated or asserted, so by VR3's floor the result is **unresolved within the boundary**, not a pass. It is also not a rejection: R1–R4's classes do not apply, because nothing was found to be wrong. The report states both facts and leaves the result unclassified — a gap in the class set, not in this record (open question 3). **Case B — right record, wrong moment.** Every tier-1 check passes at the moment of the act; standing was withdrawn before the verifier looked. Reporting a pass at moment 3 while presenting it as a statement about moment 1 is the error VR2 exists to prevent. The verdict is about the moment named, and the report must let a reader see which — otherwise the reader supplies the wrong one. -**Case C — absence is not silence.** No record at all, a bounded search that returned nothing, and a reference that was established are three different findings. Only the second carries a scope and an author; the third is a positive finding. A report that collapses them reads as "nothing wrong" when the correct reading differs in each case: nothing was recorded, nothing was found within a stated search, or the thing was established. +**Case C — absence is not silence.** No record at all, a bounded search that returned nothing, and a reference that was established are three different findings. Only the second carries a scope and an author. A report that collapses them reads as "nothing wrong", and the correct reading differs in each case: nothing was recorded; nothing was found within a stated search; the thing was established. ## 8. Open questions -1. **Should the grade set and the status set be one table or two?** They are defined from the same two questions here, which makes them total but couples them: a fifth status would force a fifth grade. *(@zhaoxinghua09-cell)* -2. **How much agreement counts as corroboration at grade C?** This record does not set a threshold, and §3 excludes thresholds by scope. If the group wants C to be operable across implementers, the threshold has to come from somewhere; naming that somewhere is the open part. *(@zhaoxinghua09-cell)* -3. **Does VR5 bind all verifier output, or only published verification reports?** Binding all output is stronger and harder to meet, and I would rather the group choose it knowingly than have it arrive by implication. *(@zhaoxinghua09-cell)* -4. **Should "pass" and "unresolved within boundary" become named classes in AAR-REQ?** VR4's floor produces a result that R1–R4's class set does not name. Either the class set grows, or the report leaves it unclassified as Case A does. *(@zhaoxinghua09-cell)* -5. **Where does the boundary in VR4 come from?** If a minimum boundary is named somewhere in FG-TIDA's output, this record should cite it rather than let each verifier declare its own. *(@zhaoxinghua09-cell)* +1. **Is there an order between the nine pairs?** This record deliberately claims none, and VR3's floor uses only "performed, against source". If the group needs a single rank for comparison, the ordering is a policy decision and should be taken as one. *(@zhaoxinghua09-cell)* +2. **Does the floor belong at "performed, against source", or lower?** A named party's check against source is admitted by the floor; a verifier's own corroboration is not. That asymmetry follows from independence, not from the verifier's effort, and it is worth testing against a case where the verifier is the only party with access. *(@zhaoxinghua09-cell)* +3. **Should "pass" and "unresolved within boundary" become named classes in AAR-REQ?** VR3's floor produces a result R1–R4 do not name. Either the class set grows, or the report leaves it unclassified as Case A does. *(@zhaoxinghua09-cell)* +4. **Does VR4 bind all verifier output, or only published reports?** Binding all output is stronger and harder to meet, and I would rather the group choose it knowingly. *(@zhaoxinghua09-cell)* +5. **Where does the boundary in VR3 come from?** If a minimum boundary is named in FG-TIDA's output, this record should cite it rather than let each verifier declare its own. *(@zhaoxinghua09-cell)* ## 9. Provenance @@ -177,26 +165,29 @@ These are the cases where a report most easily states more than was checked. - Verifier-side charter — https://github.com/FG-TIDA/themes/pull/28 - Evidence grading, separating evidence quality from authorization validity — https://github.com/FG-TIDA/themes/issues/42 - Evidence-graded authorization verdicts — https://github.com/FG-TIDA/themes/issues/43 +- Independence between references — https://github.com/FG-TIDA/themes/issues/60 - Conformance verdict scope, handed off by AAR-REQ §1.3 — https://github.com/FG-TIDA/themes/issues/6 - Author identity — https://orcid.org/0009-0001-0512-1237 - Account relationship and conflict-of-interest disclosure — https://github.com/LGD-Eval/evaluation-protocol/blob/main/CONFLICT_OF_INTEREST_DISCLOSURE.md ## 10. Limitations -- **Only the clause table and the scope chapter were read.** AAR-REQ states that its clause text "stands in 2 comments posted on FG-TIDA/themes#1" and that where the table differs, those comments govern. This record was written against the table in §2 and the scope in §1, and has **not** been reconciled against those two comments. Where they differ, they win, and VR2 in particular should be re-read against them. -- **The three moments are a proposal, not text.** AAR-REQ §2.1 records them against open question 12 as a possible R7.2a — "if the proposed R7.2a becomes text". VR2 adopts them as the set to choose from; if the group settles differently, VR2's enumeration changes with it. -- **The grade set is single-author and untested.** VR1's four grades are coherent by construction, which is not the same as being right. C is the one to test: it is the only grade whose boundary is a matter of degree, and open question 2 is unresolved. -- **No consequence semantics beyond §4.** The clauses say what a report must state and what the next party can do without it. They do not say what happens to a verifier that states nothing — that is AAR-REQ's conformance question, which §1.3 moves elsewhere. -- **No implementation accompanies this version.** The reference implementation covers the evidence-gated state machine, not the report format these clauses describe. A worked example set is the obvious next increment and is not claimed here. -- **The material this record draws on has not been through peer review.** `FG-TIDA/themes#42` and `#43` are proposals under discussion, not settled positions, and VR1 is a restatement of their separation argument rather than independent evidence for it. -- **Drafted with AI assistance** under author direction, and red-teamed before posting; the author is responsible for every claim above. +- **Only the clause table and the scope chapter were read.** AAR-REQ states that its clause text stands in two comments on `FG-TIDA/themes#1` and that where the table differs, those comments govern. This record was written against the table in §2 and the scope in §1, and has **not** been reconciled against those comments. Where they differ, they win; VR2 should be re-read against them first. +- **The three moments are a proposal, not text.** AAR-REQ §2.1 records them against open question 12 as a possible R7.2a. If the group settles differently, VR2's enumeration changes with it. +- **The two components are a design that has not been tested.** Nine pairs can be named, but whether two readers assign the same pair to the same facts is untested. The likely disagreements are whether a given artefact counts as source under §0, and whether a check was performed by the verifier or by a party acting for it — a distinction §0 does not draw and probably needs to. +- **The floor is a judgement, not a derivation.** "At least one check performed against source" is where I would put it; a group that wants lower can lower it, and the case for doing so is open question 2. +- **No consequence semantics beyond §4.** The clauses say what a report must state and what the next party can do without it. They do not say what happens to a verifier that states nothing; that is AAR-REQ's conformance question, which §1.3 moves elsewhere. +- **No implementation accompanies this version.** The reference implementation covers the evidence-gated state machine, not the report format these clauses describe. +- **The material this record draws on has not been through peer review.** `FG-TIDA/themes#42` and `#43` are proposals under discussion, and VR1 restates their separation argument rather than providing independent evidence for it. +- **Drafted with AI assistance** under author direction, and red-teamed twice before posting; the author is responsible for every claim above. ## 11. Version history | Version | Date | Change | |---|---|---| | 0.1 | 2026-11-16 | First draft | -| 0.2 | 2026-11-16 | Position restated as a refinement of R1's unstated accept side rather than an unoccupied space; VR2 premise corrected against R4.2 (scope extension, not a blank); verifier defined against relying party; VR1 attribution corrected to #42/#43; grades redefined so each has a route to being reached and each status maps to a grade; VR4 given a floor, and Case A reclassified as unresolved; R8 count corrected; conformance language added; limitations extended to the posted-comments risk and the proposal status of the three moments | +| 0.2 | 2026-11-16 | Position restated as a refinement of R1's unstated accept side; VR2 premise corrected against R4.2; verifier defined; VR1 attribution moved to #42/#43; VR4 given a floor; conformance language added | +| 0.3 | 2026-11-16 | Grade lattice replaced by a two-component report, removing an unclaimed ordering and two nameless combinations; VR1 and the former VR3 merged, so each reported value has one home; "no clause says" corrected — R4.4 and R7 supply properties, not content; **source** defined in §0; VR3's floor reason aligned with the rule; conformance language extended to every obligation verb | --- @@ -204,14 +195,13 @@ These are the cases where a report most easily states more than was checked. > 英文为权威文本;以下为逐条中文要点,供中文读者对照,非逐句对译。 -**这份记录站的位置。** AAR-REQ §1.2 把「依赖方要报什么」写进了自己的范围,R1–R4 也逐个列了**拒绝**的类别。真正空着的不是整块,而是一条:R1 说「光说接受或拒绝不够」,但**从没说一个不 bare 的「接受」要带什么**;§1.3 又把「一致性判定能主张什么」让给了 themes#6 task 3。本文补前一条,不碰后一条。**对 R1–R8 一字不加、一字不改。** +**这份记录的位置。** R1 要求「只说接受或拒绝不够」,R1–R4 又把**拒绝**的类别逐条给了内容;**接受这一侧被要求「不 bare」,却始终没被给内容**——没有条款说一份接受报告覆盖了哪些检查、边界在哪。R4.4 和 R7 在附近,但只到一半:R4.4 让「被携带的资格答复」受 R7 约束,R7 管的是**怎么携带**(归属、时刻、防篡改),不是**查了什么、边界在哪**。这就是本文的全部题目,比「AAR-REQ 有个洞」要窄:属性已给,内容未给。§1.3 另把「一致性判定能主张什么」让给 themes#6 task 3,本文不进去。 | 条款 | 中文要点 | |---|---| -| **VR1** | 证据等级描述的是「验证方与证据的关系」,不是记录自带的字段。四档由两个问题定义——**查没查、对什么查**——所以每一档都走得通、每个状态都有对应等级:A 自己查、对能直达且无需信任产出方的源;B 与产出方无关的具名方查、对同类源;C 自己查、对独立于产出方的材料但不在源头;D 没查,只有产出方的说法。等级针对**主张**不针对**行为人**;四档有序但**不可平均**,平均会把最弱那一跳藏起来。 | -| **VR2** | R6 说时效是关系不是属性;R4.2 **已经**要求依赖方记录用了哪个时刻——但只管 tier-2 的资格问题。本条的增量是**范围**不是内容:把「点名一个时刻」从资格问题推广到每一份判定,并把三时刻(行为时/产出方声明时/独立可界定时)作为可选集合。三时刻是 @OlehVasylenko2000 的**提案**(可能成为 R7.2a),不是既成文本,这一点正文已写明。 | -| **VR3** | 「可检查」不等于「已检查」。R7.3 只保证「与产出方无关的人**能**查」,一份没人看过的记录同样满足它。所以每项检查标状态:验证方自做/具名第三方做/仅有断言。与 VR1 是全覆盖映射:A、C 是自做,B 是第三方,D 是断言。 | -| **VR4** | 通过 = 在声明边界内、用声明的检查、在声明的时刻,未发现违规。必带四样:做了什么、**没做什么**(要点名)、边界、时刻。**另有一条地板**:若全部检查都是 C 或 D,结论是「边界内未决」,不是通过——否则「什么都没对源查过」的报告也能凑齐四条伪装成通过,而这正是 VR3 要挡的失效。有了地板,通过才能**被它自己的声明推翻**。 | -| **VR5** | 验证方的报告是下一方要评的东西;一旦被依赖,它就是 R7 意义上的 carried claim,R7 的三项(归属/时刻/防篡改)按其自身条款即已适用。本条不扩张 R7,只指出它本来就够得着报告。推论很窄:不披露自身依据的验证方,产出的是一份归属不明的主张,下一方读不出 VR4 的边界。 | +| **VR1** | 每项检查报**两个分量**,不是打一个分:**谁查的**(验证方/具名第三方/没查即断言)× **对什么查的**(源头/佐证材料(不在源头)/无)。九个组合,个个有名,**不主张彼此排序**——排序等于替组里做政策决定(开放问题 1)。两条边界:「source」§0 定义为「无需采信产出方的说法即可运行的对照物」,**记录本身不是 source**(读记录只能知道它说了什么,不能知道它说的是不是真的);佐证**刻意弱于源头**,因为佐证分不开「独立来源」和「共上游」——正是 #60 那个问题。 | +| **VR2** | R6 说时效是关系;R4.2 **已经**要求记录用了哪个时刻,但**只管 tier-2 资格问题**。本文的增量是**范围**:推广到每一份判定。三时刻(行为时/产出方声明时/独立可界定时)是 @OlehVasylenko2000 的**提案**(可能的 R7.2a),不是既成文本,正文写明。还要报「能否复核」——没有复核手段本身是结论,不是空白。 | +| **VR3** | 通过必带四样:做了什么检查(含 VR1 两分量)、**没做什么**(点名)、边界、时刻。通过的含义**不多于**:在该边界内、用这些检查、在该时刻,未发现违规。**另设地板**:至少一项检查是「做了、且对源头」——否则一份全部靠佐证或断言、却凑齐四样的报告,也能自称通过,而且看不出区别。有了地板,通过才能**被它自己的声明推翻**。 | +| **VR4** | 验证方的报告一旦被下一方依赖,就是被携带的主张,R7 按其自身条款即已适用(归属/时刻/防篡改)。本文**不扩张** R7,只指出它本来就够得着;不披露自身依据的验证方,产出的是一份下一方无法评到最低一档以上的主张。这不是要求公开,是说未公开的代价由下一方承担。 | -**三个边界情形。** A:记录可检查但本验证方查不到——按地板判「边界内未决」,既不是通过也不是拒绝(R1–R4 没有对应类,这是 AAR-REQ 类集的缺口,列为开放问题 4)。B:记录对、时刻错——行为时全通过、验证方来看时资格已撤回;把时刻③的结论当时刻①报,正是 VR2 要挡的。C:无记录 ≠ 静默——「没有记录」「有界检索查无」「引用已确立」是三种结论,只有第二种带检索范围和作者。 +**三个边界情形。** A:记录可读但无可对源——按地板判「边界内未决」,既不是通过也不是拒绝(R1–R4 无对应类,属类集缺口,开放问题 3)。B:记录对、时刻错——把时刻③的结论当时刻①报,正是 VR2 要挡的。C:无记录 ≠ 静默——「没有记录」「有界检索查无」「引用已确立」三种结论,只有第二种带检索范围和作者。 From f66c4658fd1c99b1e6fef99b99eed3865dc8e45f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?LGD-Eval=20=E2=80=94=20Independent=20AI=20Evaluation?= Date: Mon, 16 Nov 2026 14:20:48 +0800 Subject: [PATCH 4/4] Revise verifier-side interface record to 0.4 Positioning restated as a refinement of R1's unstated accept side, not an unoccupied space. VR2 premise corrected against R4.2: scope extension, not a blank. Verifier defined against relying party. VR1 attribution moved to #42/#43. Grades redefined so each is reachable and each status maps to one. VR4 given a floor; Case A reclassified as unresolved. R8 count corrected. Conformance language added. Limitations extended. Assisted-by: WorkBuddy --- FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md | 25 +++++++++++++++++-------- 1 file changed, 17 insertions(+), 8 deletions(-) diff --git a/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md b/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md index 6489254..543d274 100644 --- a/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md +++ b/FG-TIDA-AAR-VERIFIER-SIDE-INTERFACE.md @@ -7,7 +7,7 @@ | Author | Xinghua Zhao (Steven Zhao), independent expert, GitHub `@zhaoxinghua09-cell` | | Offered in | Individual-expert capacity, as stated on `FG-TIDA/themes#46` on 29 October 2026 | | Hosted in | `LGD-Eval/evaluation-protocol` — the public home of the evaluation material this record draws on. Hosting is a publication venue, not an authorship claim; the account relationship is linked in §9. | -| Version | 0.3 — 16 November 2026 | +| Version | 0.4 — 16 November 2026 | | Status | Draft for attributed review before the first in-person FG-TIDA meeting (Paris, 30 November – 1 December 2026) | | Linked work item | AAR-REQ 0.2 — https://tyche-dev.github.io/tida-aar-req/ | | Licence | CC BY 4.0 | @@ -20,7 +20,7 @@ **Verifier.** The relying party itself, or an agent acting for it, when it evaluates a record and states a result. This record introduces no second role: wherever AAR-REQ binds a relying party, the obligation falls on the verifier, and the name is used because these clauses are about the act of verifying rather than the act of relying. -**Record-side and verifier-side.** Record-side requirements constrain what a record must carry. Verifier-side requirements constrain what a party must state when it reports a result from a record. R1–R8 are record-side; the clauses below are verifier-side. +**Record-side, rejection-side, accept-side.** R3–R8 constrain what a record must carry. R1–R4 constrain what a relying party reports **when it rejects**. Neither constrains what it reports when it accepts — that is the subject of the clauses below. **Source.** A party, register or artefact against which a check can be run **without taking the producer's word for the result**. The record is the object checked, not a source: reading it establishes what it says, not whether what it says is true. This definition matters because it is the only thing standing between a verification and a recital. @@ -56,7 +56,7 @@ Neither says **what was checked** or **within what boundary** — R7 regulates h ## 4. Conformance language -Where a clause says the verifier **states**, **reports**, **names** or **carries**, that is **SHALL** for any report presented as a pass, and **SHOULD** for any report presented as unresolved. The consequence of omission is stated once, here: a report that omits a required statement cannot be assessed by the next party above the level that party can establish without it, which under VR1 is "not performed, no source" — the lowest pair. Two further consequences follow from the clauses themselves rather than from this section: an asserted check cannot support a pass (VR3), and a pass whose boundary is undeclared cannot be withdrawn and therefore cannot be relied on (VR3). +Every clause statement about what the verifier does — **states**, **reports**, **names**, **carries**, **says** — is **SHALL** for any report presented as a pass, and **SHOULD** for one presented as unresolved. Where a required statement is omitted, the next party cannot assess the report above the pair "not performed, against none", which is the lowest pair under the within-component order in VR1. Two further consequences follow from the clauses themselves rather than from this section: an asserted check cannot support a pass (VR3), and a pass whose boundary is undeclared cannot be withdrawn and therefore cannot be relied on (VR3). ## 5. The clauses @@ -64,7 +64,7 @@ Attribution follows AAR-REQ Table 2.1: each clause names who proposed it and, in ### VR1 — Each check is reported by who performed it and against what -*Proposed by @zhaoxinghua09-cell. The separation of evidence quality from authorization validity is `FG-TIDA/themes#42`'s argument, restated in `#43` as LGD-Eval with @dakleyer's refinements; this clause applies it to the AAR case and does not claim the argument.* +*Proposed by @zhaoxinghua09-cell. The separation of evidence quality from authorization validity is `FG-TIDA/themes#42`'s argument, restated in `#43`, published under the LGD-Eval account, with @dakleyer's refinements; this clause applies it to the AAR case and does not claim the argument.* A record can be complete and still support nothing, because what a report supports depends on what the verifier did and what it did it against. Both parts are reported, as a pair — not as a single rank: @@ -73,7 +73,15 @@ A record can be complete and still support nothing, because what a report suppor | **Performed by** | the verifier / a named party / not performed (asserted) | | **Against** | source (per §0) / corroborated material, not at source / none | -Nine combinations, each nameable, and no ordering is claimed between them. That is deliberate: an ordering would have to decide, for instance, whether a named party's check against source outranks the verifier's own check against corroborating material, and any answer to that is a policy choice the group has not made (open question 1). +The values do not combine freely: where nothing was performed there is nothing to check against, so "not performed" pairs only with "none". The pairs that can occur are five: + +- performed by the verifier, against source; +- performed by the verifier, against corroborated material; +- performed by a named party, against source; +- performed by a named party, against corroborated material; +- not performed, against none. + +**No order is claimed between the pairs.** The values within each component are ordered — performed before not performed, source before corroborated before none — and that within-component order is the only one used below. Ordering the pairs against each other would have to decide, for instance, whether a named party's check against source outranks the verifier's own check against corroborating material, and any answer to that is a policy choice the group has not made (open question 1). Two points about what the pair does **not** say: @@ -111,7 +119,7 @@ A pass carries four statements, and is incomplete without any of them: A pass then means no more than this: *no violation was found within this stated boundary, using these checks, at this moment.* It does not mean the act was authorized, that the record is true, or that the delegation was sound. AAR-REQ says as much one level down — "R7 does not make a claim true" — and the same restraint holds at the verdict. -**A pass also has a floor.** At least one reported check must be performed, against source. Without the floor, a report in which every check was corroborated or asserted could state all four things and present itself as a pass — and nothing in the report would distinguish it from a report in which something was actually checked. The floor is why corroboration is weaker than source in VR1: corroboration alone cannot carry a pass, because it cannot establish that the evidence is independent of the producer. +**A pass also has a floor.** At least one reported check must be performed, against source. Without the floor, a report in which every check was corroborated or asserted could state all four things and present itself as a pass — and nothing in the report would distinguish it from a report in which something was actually checked. The same property that makes corroboration weaker than source in VR1 is what keeps it from carrying a pass: corroboration cannot establish that the evidence is independent of the producer. The floor is what makes a pass **withdrawable by its own declaration**: a later party can withdraw it by showing one check inside the declared boundary that was not performed as reported, without relitigating the facts. A pass that declares no boundary cannot be withdrawn that way, and therefore cannot be relied on — only believed. @@ -162,7 +170,7 @@ What the verifier reads from each record-side clause, and what it reports in ret - AAR-REQ 0.2, web — https://tyche-dev.github.io/tida-aar-req/ - AAR-REQ 0.2, PDF — https://tyche-dev.github.io/tida-aar-req/specification.pdf - Originating thread — https://github.com/FG-TIDA/themes/issues/46 -- Verifier-side charter — https://github.com/FG-TIDA/themes/pull/28 +- Verifier-side charter, for the theme from #27 — https://github.com/FG-TIDA/themes/pull/28 - Evidence grading, separating evidence quality from authorization validity — https://github.com/FG-TIDA/themes/issues/42 - Evidence-graded authorization verdicts — https://github.com/FG-TIDA/themes/issues/43 - Independence between references — https://github.com/FG-TIDA/themes/issues/60 @@ -179,7 +187,7 @@ What the verifier reads from each record-side clause, and what it reports in ret - **No consequence semantics beyond §4.** The clauses say what a report must state and what the next party can do without it. They do not say what happens to a verifier that states nothing; that is AAR-REQ's conformance question, which §1.3 moves elsewhere. - **No implementation accompanies this version.** The reference implementation covers the evidence-gated state machine, not the report format these clauses describe. - **The material this record draws on has not been through peer review.** `FG-TIDA/themes#42` and `#43` are proposals under discussion, and VR1 restates their separation argument rather than providing independent evidence for it. -- **Drafted with AI assistance** under author direction, and red-teamed twice before posting; the author is responsible for every claim above. +- **Drafted with AI assistance** under author direction, and red-teamed in three independent passes before posting; the author is responsible for every claim above. ## 11. Version history @@ -205,3 +213,4 @@ What the verifier reads from each record-side clause, and what it reports in ret | **VR4** | 验证方的报告一旦被下一方依赖,就是被携带的主张,R7 按其自身条款即已适用(归属/时刻/防篡改)。本文**不扩张** R7,只指出它本来就够得着;不披露自身依据的验证方,产出的是一份下一方无法评到最低一档以上的主张。这不是要求公开,是说未公开的代价由下一方承担。 | **三个边界情形。** A:记录可读但无可对源——按地板判「边界内未决」,既不是通过也不是拒绝(R1–R4 无对应类,属类集缺口,开放问题 3)。B:记录对、时刻错——把时刻③的结论当时刻①报,正是 VR2 要挡的。C:无记录 ≠ 静默——「没有记录」「有界检索查无」「引用已确立」三种结论,只有第二种带检索范围和作者。 +| 0.4 | 2026-11-16 | R1–R8 split into rejection-side (R1–R4) and record-side (R3–R8) so the accept side is defined negatively and accurately; the five occurrable VR1 pairs enumerated, with a within-component order stated and no order between pairs; §4 conformance language extended to every obligation verb and its "stated once" claim removed |