F.12 — Evaluate Promise Fulfilment from Work Evidence (Service Acceptance)
Preface node
heading:f-12-evaluate-promise-fulfilment-from-work-evidence-service-acceptance:131347
What this page is
This is generated FPF reference text from the specification preface or supporting sections. It helps interpret FPF; it is not FPF Reference product documentation.
Methodology
Use it to understand how the specification wants to be read, then return to a route, pattern, or work packet for active work. Cite generated IDs only when the wording changes the task decision.
Content
“Judge a promise from what happened, in a stated window, with evidence that bears on the promised outcome.”
Type: Architectural pattern.
Status: Stable
Builds on: F.1 Question-Relative Source Selection; F.0.1, F.2, F.3, and F.17 for exact source-local meaning and optional addresses; F.5 for naming; F.9 only for actual relations between local meanings; F.10 for status families and windows; F.11 for the Method, MethodDescription, and Work distinctions; A.2.3 for U.PromiseContent; A.15.1 for evaluation Work; and A.6.1 for the applied evaluation operation and its result binding.
Coordinates with. C.2 and C.16 for observation, characteristic, scale, unit, and measured-value claims; A.3.2 when a selected evaluation MethodDescription edition changes the result; A.10 for evidence use; B.3 only when an actual named assurance claim is current; C.16.P and A.6.RCD when indicator or proxy wording hides an unsupported relation; E.13 only when a proxy is optimized or used as a target, incentive, gate, release argument, reputation signal, repair target, or decision driver; and direct transformation or control patterns when outputs are material.
Scope. F.12 explains the claims needed to justify a promise-fulfilment judgement.
Intent & applicability
Intent. Relate an exact promise-content claim to the delivery Work and outcome being judged, the observations and measured values used as evidence, the relevant window and population, and application of the declared acceptance rule to those inputs. Return a result on its declared scale.
Recover a separate dated evaluation Work when the claim or receiving use consumes that performance history. Map the result to F.10 status when needed, and identify a durable verdict episteme when another use needs that assertion. Keep these objects and relations distinct so a reader can see what would change the result.
Use this when. An SLO, SLA clause, safety margin, response-time target, quality gate, or other promise must be judged from actual occurrences.
Do not use this when. The question is only what the promise says, how the Method is described, or how a measurement is made. Use the direct A.2.3, A.3, A.15, or C.16 pattern. F.12 does not turn a lexical cell or comparison row into a verdict subject.
Problem frame
- Plan ≠ proof. A diagram or playbook is treated as evidence that its promise was met.
- Output ≠ outcome. Commands or setpoints are mistaken for the consumer-relevant result.
- Word ≠ measure. Availability, latency, or incident is used without recovering its exact local meaning and characteristic.
- Proxy by assumption. Monitor output is called a proxy before asking whether the observation and measurement model already concern the promised characteristic directly; when they do not, the needed indicator relation has no defining or testing rule.
- Evaluation disappears. A verdict appears without the selected facts, acceptance rule, calculation or other application that supports its result. A claim about who performed the evaluation additionally needs its actual performance basis.
- Status families collapse.
Satisfied,Violated, andInconclusiveare treated as one universal scale even though the first two are RequirementStatus values and the third is an EvidenceStatus value unless an exact local result scale says otherwise.
Forces
Core idea (didactic)
For an ordinary evaluation, the delivery facts, adequate observations, acceptance rule and supported result may suffice. Recover the evaluation’s performance history when it is claimed or used. Recover the following content as needed to justify the result:
- the exact
U.PromiseContentclaim being evaluated; - the actual Work occurrence or defined Work population whose delivery is in question;
- the promised outcome or characteristic and its scope;
- the relevant observations and measured values, including scale and unit;
- the explicit window and population boundary;
- the evaluation Method and, when its performance history is claimed or consumed, the independently admitted dated evaluation Work that enacts it;
- the selected evaluation inputs, applied rule and result; recover the exact A.6.1 application and result binding when that identity is claimed or required by the governing rule;
- the acceptance rule and the result scale it declares; examples include Boolean, trichotomous, graded,
N/A, andInconclusive-including scales; and - the PromiseContentUse, delivery, fulfilment, measurement, evidence-use, any separately defined indicator or proxy, reliance, and status relations that connect these claims.
The operation returns a value on its declared scale. A separate F.10 rule may warrant a RequirementStatus or EvidenceStatus for that result and use. If the evidence leaves the threshold open, say so. Inconclusive can itself be a local acceptance result when the scale provides for it.
When the receiving decision concerns acceptance of a commissioned result, ask what effect the arrangement gives this evaluation. A passing value settles acceptance if the agreement makes it decisive and the other required conditions hold. A customer empowered to accept a disclosed deviation can accept an item whose measured value still misses the target. The terms determine what the acceptance covers and whether correction is still due. For several linked undertakings, use A.6.C:4.1.
A plain summary may say met, not met, or cannot judge while distinguishing the result on its declared scale, the standing of the evidence and the status of the requirement.
A SchemeSenseCell addresses a local meaning; the evaluation still concerns the exact promise, Work, observations, values, result and status-use relation. A comparison table may display the argument; display alone does not establish those claims.
Minimal vocabulary
- Promise-content claim — the exact
U.PromiseContentunder A.2.3, including its subject, scope, target, and conditions. - Work — the actual dated
U.Workoccurrence, or an explicitly defined population of Work occurrences, being judged. - Observation — the actual observation occurrence and its result under C.2 and C.16.
- Measured value — a value for a named characteristic on an explicit scale and unit.
- Window — the time, batch, episode, phase, or other bounded evaluation interval under F.10.
- Evaluation Work — the dated Work in which a System enacts the evaluation Method over the selected facts and states; identify it when the evaluation account or receiving use consumes that performance history.
- Evaluation application and result — the application of the rule to its selected inputs and the result on the acceptance specification's declared scale. An exact application identity and argument/result binding use A.6.1 when that formal claim is current.
- Evaluation rule and result scale — the stated comparison or aggregation and its declared admissible results. Boolean, trichotomous, graded,
N/A, andInconclusive-including scales are examples, not defaults. - Status use — a separate F.10 application of an exact EvidenceStatus or RequirementStatus value to its exact target, scope, window, and use after the direct result is recovered.
- Verdict episteme — an optional C.2.1 episteme that states the evaluation result or status when another use needs a durable assertion; it is not the operation result or the fulfilment relation.
- Indicator or proxy relation — only a separately defined or tested relation in which one observed characteristic or result stands in for another subject or outcome for this use, with exact participants, coverage, and loss. The word proxy does not create it.
- Evidence use and reliance — the independently established direct relations supporting the acceptance claim, represented in A.10's descriptive evidence-provenance path, with a local reliance disposition for the bounded use. B.3 enters only for an actual named assurance claim. Neither the path nor the disposition creates an indicator relation or the acceptance result.
The binding, as eight practical rules
R1 — Match the promise to delivery.
Use A.2.3 to keep the exact promise content, PromiseContentUse, delivered outcome, and fulfilment relations distinct for the Work occurrence or population being judged. An abstract service label or lexical cell is not enough, and the later evaluation does not make those delivery-side relations obtain.
R2 — Test for direct measurement first. Ask whether the observation and measurement model directly concern the promised characteristic of the relevant Work outcome inside the window. If they do, use C.16 for measurement and A.10 for evidence use; add no proxy relation. Commands, approvals, and MethodDescriptions are not outcome evidence by themselves.
R3 — Recover an indicator relation when one is needed.
If a distinct observed indicator stands in for the promised characteristic or outcome, name both participants and cite the pattern that defines or tests that exact relation. Use C.16.P to recover the construction and distortion risk. If no current rule supplies the relation, stop with A.6.RCD missing-governor; do not treat the word proxy, A.10 evidence use, or B.3 reliance as its substitute.
Use E.13 only when the indicator is optimized or used as a target, incentive, gate, release argument, reputation signal, repair target, or decision driver.
R4 — Perform the evaluation. Select the delivery facts, affected-state references and measured values required by the acceptance rule. Apply its comparison or aggregation with the specified population, exclusions, uncertainty treatment and result scale, then state the supported result. For example, an incident rule can bind observed start and restoration times and compare their difference with the allowed duration. If the required observations or rule are unavailable, return the precise missing basis rather than a fulfilment verdict.
When the claim or receiving use consumes a dated evaluation occurrence, recover each evaluator's A.13 core and let A.15.1 independently admit that Work before asserting that it enacts the evaluation Method, as A.2.3:4.3 requires. Add F.6 only for precise assignment-bound attribution.
Recover an exact A.6.1 application and its argument/result binding when claimed or required by the rule. Cite a particular MethodDescription edition when its content changes the evaluation or replay.
R5 — Use the declared result scale.
Name the characteristic, scale, unit, aggregation, comparison, and admissible result values through the acceptance specification's verdictScaleDescriptionRef. Typical calculation shapes include:
- a value at or above, or at or below, a threshold;
- a stated percentile at or below a target;
- a share such as good time divided by total time;
- an event count within a limit;
- all relevant values remaining inside a band.
The calculation shape does not select a verdict scale. Use the exact scale declared by the acceptance specification.
R6 — State every needed relation directly. Use A.2.3 for promise use, delivery, and fulfilment; C.16 for observation and measurement; the defining or testing pattern for any indicator relation; A.10 for the descriptive evidence-provenance path and bounded reliance on independently established direct relations; B.3 only for an actual named assurance claim; and F.9 only when distinct local meanings themselves require a semantic relation.
One generic Bridge cannot establish clause–Work fit, measurement, indicator validity, evidence, evaluation, status, or fulfilment. A passing reliance disposition requires the evidence demanded by the acceptance claim and its direct rule; a descriptive path alone is insufficient.
R7 — Keep the window and population explicit. A monthly verdict, a batch verdict, and an incident verdict are different claims. A new promise, monitor, or window does not rewrite an earlier verdict.
R8 — Preserve result, status, and assertion boundaries.
Keep the result on its declared acceptance scale. When the evidence leaves more than one fulfilment outcome open, state the unresolved decision and its grounds. Assert a RequirementStatus such as Satisfied, Violated or Pending, or an EvidenceStatus such as Inconclusive, only under its applicable F.10 rule. Create a verdict episteme when another use needs that durable assertion.
Evaluation shapes
The calculation shapes and examples below preserve the delivery Work or population and illustrate an explicit evaluation-Work account. Use that account when its history is current. For an ordinary result, apply the same population, measurement and acceptance conditions through R4 without separately identifying the evaluation occurrence.
Availability share
Promise: availability is at least 99.9% for a calendar month. Delivery Work: the defined service-delivery occurrences or population during that month. Evidence: observations and measurement results for the promised availability characteristic. A System performs evaluation Work; its A.6.1 application binds the in-scope values and returns the declared result.
If the observation model already concerns the promised characteristic, there is no proxy relation. If synthetic probes measure a different characteristic, identify the rule relating it to user experience. State which regions or degradations the probes leave uncovered and how that limits use of their evidence. Without a defining or testing rule return missing-governor. If the rule exists, apply it and limit the verdict to what the probe evidence supports.
Latency percentile
Promise: p95 response latency is at most 120 ms for a stated request population and window. Evidence: response-time observations for that population. Evaluation Work applies the declared sampling, exclusion, and percentile rule and binds its result on the declared acceptance scale. Sampling bias or missing paths can leave the percentile threshold open. Apply any F.10 status rule or declared local result scale separately.
Safety or quality band
Promise: temperature remains within [L,U] during the batch phase. Evidence: calibrated temperature observations for the relevant EntityOfConcern and interval. Evaluation Work applies the stated sampling, uncertainty, and band rule; the exact application binds the values and returns a result on the declared scale. A RequirementStatus follows only through its own rule.
Incident duration
Promise: restoration occurs within 60 minutes for each in-scope incident. Delivery Work: each handling occurrence. Evidence: observations of the defined start and restoration events. Evaluation Work applies the elapsed-time rule to those bindings and returns its declared result. A BPMN design may be a MethodDescription under A.3.2, but it is not either Work occurrence or evidence of the result.
Invariants
- Exact promise. The evaluation identifies the
U.PromiseContentclaim, not merely an SLO label or cell. - Delivery-side relations. The judged delivery Work or population and the applicable A.2.3 promise-use, delivered-outcome, and fulfilment relations remain distinct from evaluation.
- Outcome evidence. Observations and values concern the promised characteristic of that Work; outputs and approvals alone are insufficient.
- Direct-before-proxy. When observation and measurement directly concern the promised characteristic, use C.16 and A.10 and add no proxy. For a distinct indicator, identify its participants and the rule that defines or tests their relation. Return
missing-governorif the rule is missing. If it is known, check its conditions for this case. - Window and population. Both are explicit and match the promise.
- Evaluation basis. The selected inputs, applied acceptance rule and result remain recoverable. A claimed or rule-required dated evaluation Work and exact A.6.1 application satisfy their separate admission and binding conditions.
- Declared result scale. Characteristic, scale, unit, aggregation, threshold, exclusions, and admissible result values are stated as applicable. Boolean, trichotomous, graded,
N/A, andInconclusive-including scales are examples, not defaults. - Status separation. Distinguish the acceptance result, any EvidenceStatus and any RequirementStatus. Each status requires its applicable F.10 rule. When the evidence leaves the fulfilment decision open, report that limit even if the status rules supply no value.
- Optional verdict episteme. A durable C.2.1 assertion is created only for a named later use and never replaces the application result, status-use occurrence, evidence relation, or fulfilment relation.
- Bounded reliance. A.10 supplies the descriptive evidence-provenance path and local bounded-reliance disposition; B.3 is used only for an actual named assurance claim; E.13 is used only for an optimized or decision-driving proxy.
- Non-retroactivity. Later promise, monitor, MethodDescription edition, or interpretation changes do not silently alter past evaluations or assertions.
- Cells are addresses only. An F.17 cell may identify local meaning but establishes none of the substantive claims above.
Micro-examples
SaaS uptime
- Promise content: availability ≥ 99.9% for the named service scope in June.
- Delivery Work population: the in-scope service-delivery occurrences during June, connected to the promise through the applicable A.2.3 relations.
- Evidence: synthetic-probe observations, with regions and outage-detection coverage stated. If their measurement model directly concerns the promised availability characteristic, no proxy is added; otherwise the distinct probe-to-user indicator relation needs a defining or testing pattern.
- Evaluation: the evaluating System applies the rule to the June window. Verified good time gives a lower bound. Include any unclassified time that could still be good when finding the upper bound. Check whether either bound already settles the threshold, using the indicator relation and regional coverage above.
- Status or summary: report whether the threshold was met, missed or remains open. Where a formal status is needed, apply the rule for that status and retain its basis.
Furnace temperature band
The promise content states [720,740] °C during the soak phase. The delivery Work is the actual batch soak occurrence. Calibrated thermocouple observations either measure the product characteristic directly or use a separately defined sensor-location indicator relation.
For this batch, evaluation Work applies the band rule and binds its result. An out-of-band result can support RequirementStatus=Violated under its F.10 rule. If the spatial evidence permits both outcomes, report the limit. Use an evidence status or a local acceptance result when its own rule applies.
Incident restoration duration
The promise content states restoration within 60 minutes per in-scope incident. Each incident-handling Work has observed start and restoration events. A separate evaluation Work occurrence applies the declared event and subtraction rule; its application binds those timestamps and returns the result. A playbook may be the selected evaluation MethodDescription when its edition changes that rule, but it is not the Work or proof of the duration.
Anti-patterns & remedies
Extended worked examples
CDN latency by region
Promise content: p95 end-user latency ≤ 200 ms per region per month. Delivery Work population: delivery occurrences per region in that month. Evidence: response-time observations tagged by region and path. If probes measure the promised characteristic directly, use C.16 and A.10 without a proxy. If they indicate a distinct user-experience characteristic, name the exact probe-to-user relation, its defining or testing pattern, and last-mile loss.
Evaluation Work applies the regional rule to its observations. A numerical p95 needs a nonempty population with the required measurement basis. If the population was empty, use the contract’s specified empty-region treatment. Missing observations alone do not establish an empty region. A global all-regions statement is a separate logical aggregation of those results or statuses, not a property of a table row.
Stroke care door-to-needle
Promise content: at least 90% of in-scope ischemic-stroke episodes achieve door-to-needle ≤ 30 minutes in the quarter. Delivery Work population: patient-episode care occurrences. Evidence: observations of defined door and needle events. Evaluation Work binds those events and divides the qualifying count by the eligible count when the denominator is positive. For an established empty eligible population, use the stipulated empty-population treatment or report that the percentage is undefined.
If missing triage tags or ambiguous events leave both threshold outcomes possible, identify what remains unresolved. A formal evidence status, requirement status or local acceptance result needs its own applicable rule.
Cold-chain warehouse
Promise content: product temperature remains in [2,8] °C for at least 99.5% of each day. Delivery Work: the daily storage occurrence or defined population. Evidence: calibrated thermistor observations. First ask whether the measurement model directly concerns product exposure. If the measurement concerns another characteristic, such as air temperature at the sensor, establish its indicator relation to product temperature and account for stratification loss. Return missing-governor if no rule defines or tests that relation. Otherwise apply the rule with the available coverage.
Apply the promise to the entire in-scope day. For equally weighted minutes, the known in-band count is a lower bound. Adding all unclassified minutes gives an upper bound. Of 1,440 minutes, 1,433 in band suffice for 99.5%. Eight confirmed out of band already establish failure. With 1,430 confirmed in band and ten unclassified, two further in-band minutes would give 1,432 and four would give 1,434. If both completions are compatible with the evidence and the applicable rule, both outcomes remain possible. An independent bound establishing that at least three of the ten minutes were in band already settles this threshold. Use the contract's own treatment of missing readings and assess a separately promised trace on its own terms. Retain relevant indicator limits in the result, status and reliance claim.
SaaS incident restoration duration
Promise content: restoration within 60 minutes for each in-scope incident. Delivery Work: each incident-handling occurrence. Use the start event named by the contract and the observed restoration event. If notification at 10:00 starts the clock and service returns at 11:15, the interval is 75 minutes. Starting repair at 10:40 gives a different, 35-minute interval. Apply the declared duration rule per incident. Quarterly reporting aggregates those results or their separately warranted statuses.
Reasoning about promise fulfilment
- Match scope. Confirm that the promise content covers the exact delivery Work or population and keep A.2.3 promise use, delivery, and fulfilment distinct.
- Name the window. Make time, batch, phase, and exclusions explicit.
- Test direct measurement first. Confirm whether each observation and its measurement model directly concern the promised characteristic; if so, use C.16 and A.10 and add no proxy.
- Recover an indicator only when needed. When another characteristic stands in, name both participants, the defining or testing pattern, coverage, and loss. Use C.16.P to recover the indicator. Return A.6.RCD
missing-governorwhen its defining or testing rule is missing. A known rule can also show that the indicator fails for this use. - Check values. Name characteristic, scale, unit, aggregation, and uncertainty.
- Perform the evaluation. Apply the declared rule to its selected inputs and recover the result on its scale. Identify the performing System, dated Work and exact A.6.1 application when that performance account is claimed or required; keep the actual admission and binding grounds. Cite a MethodDescription edition when it changes the result or replay.
- Use evidence directly. Recover the direct evidence-use claims, the descriptive A.10 path and its bounded-reliance disposition. Enter B.3 only for an actual named assurance claim, and E.13 only when a proxy is optimized or drives a decision, gate, incentive, release argument, reputation signal, or repair.
- Keep the result on its declared scale. Boolean, trichotomous, graded,
N/A, andInconclusive-including scales are examples, not defaults. - Map status separately. Report the result that the acceptance rule and the evidence establish. Apply the corresponding F.10 rule before assigning a requirement or evidence status.
- Create a verdict episteme only on demand. Use C.2.1 only when another use needs a durable assertion about the result or status.
- Aggregate explicitly. Population-level results and statuses follow the promise's stated quantifier; they are not inferred from a few green cases.
- Preserve history. New promises, monitors, evaluation methods, or scales create new evaluations rather than changing old ones silently.
Relations
Builds on:
- Use F.1 and F.0.1 to recover exact sources and local claims, and F.2, F.3, and F.17 only when expressions or durable addresses are needed.
- Use F.5 for clear designations, F.9 only for an actual relation between distinct local meanings, F.10 for separate EvidenceStatus and RequirementStatus uses and windows, and F.11 to keep Method, MethodDescription, Work, and output distinct.
- Use A.2.3 for exact promise content, PromiseContentUse, delivered outcome and fulfilment; A.15.1 for admitted delivery Work and any claimed evaluation Work; and A.6.1 when the exact evaluation-operation application and result binding are current.
Uses direct subject patterns. Use C.2 and C.16 for observations, characteristics, scales, units, and measured values. When the measurement does not directly concern the promised characteristic, use C.16.P to recover the distinct indicator relation and cite the pattern that defines or tests it; use A.6.RCD missing-governor when no such rule exists.
Use A.10 for descriptive provenance and bounded reliance, B.3 only for an actual named assurance claim, E.13 only for optimized or decision-driving proxies, and the appropriate direct pattern for kind, control, or transformation claims.
Constrains: Reporting and assurance preserve the promise content, delivery Work, observations, measured values, relevant window, evaluation rule and result scale. Any claimed evaluation Work, exact application, verdict episteme, EvidenceStatus, RequirementStatus, evidence use, material reliance, indicator relation and F.9 relation retains its own meaning and grounds. A relation-specific CL or loss is reported with that relation rather than folded into the result or status.
Migration notes
- Promise revision. Keep the old promise-content identity, evaluation results, and status assertions; evaluate the new claim separately.
- Monitor change. State whether the new observation model directly measures the promised characteristic or needs a separately defined indicator relation; preserve past evidence identity.
- Scope correction. Retire a result or status assertion about the wrong Work or population and issue a corrected evaluation rather than redefining the promise.
- Scale and unit change. Apply the direct conversion and measurement relations; use F.9 only when local meanings also differ.
- Population refinement. Treat per-region, per-zone, or per-episode changes as explicit promise or evaluation changes.
- Indicator retirement. Prefer direct measurement when available; keep prior indicator-dependent results, status assertions, and evidence uses with their original limits.
Acceptance tests
Static conformance
- SCR-F12-S01 (actual subjects). Every evaluation names exact promise content, delivery Work or population, observations and measured values, window, and rule; cells are optional addresses only.
- SCR-F12-S02 (scope match). Promise, Work, evidence, population, and window align.
- SCR-F12-S03 (evidence). Observations concern the promised outcome of the judged Work.
- SCR-F12-S04 (evaluation explicit). The selected inputs, applied rule, supported result, characteristic, scale, unit, aggregation, threshold and exclusions are recoverable as required by the acceptance claim. Any claimed or rule-required performing System, evaluation Work, enacted Method and exact A.6.1 application satisfy their direct admission and binding conditions.
- SCR-F12-S05 (indicator boundary). Direct measurement adds no proxy. For a distinct indicator, identify both participants and apply the defining or testing rule. Return A.6.RCD
missing-governorif no such rule is available. Limit the acceptance conclusion to the evidence the rule supports. - SCR-F12-S06 (direct relations). Promise use, delivery, fulfilment, measurement, indicator, evidence use, assurance or material reliance, evaluation, status, and any verdict assertion use their defining or testing patterns.
- SCR-F12-S07 (result and status). The operation result stays on its declared scale; RequirementStatus and EvidenceStatus are mapped separately, and evidence insufficiency is never implicit target falsity.
- SCR-F12-S08 (optional episteme). Identify a durable C.2.1 verdict episteme when a receiving use needs that assertion, keeping it distinct from result and status.
- SCR-F12-S09 (no generic Bridge). F.9 is used only for a real local-meaning relation and establishes none of the other relations.
- SCR-F12-S10 (temporal identity). No timeless or retroactively rewritten result or status assertion appears.
Regression
- RSCR-F12-E01 (relation update). A changed indicator, evidence, or semantic relation affects only evaluations that depended on it.
- RSCR-F12-E02 (edition change). Source-local meaning remains tied to the edition used by each evaluation.
- RSCR-F12-E03 (population drift). New population definitions create explicit new evaluations.
- RSCR-F12-E04 (window partition). Weekly and monthly results and statuses remain distinct; any roll-up states its aggregation.
- RSCR-F12-E05 (indicator retirement). Direct measurement changes future evaluations without silently rewriting prior indicator-dependent results or assertions.
Didactic distillation
“Name the exact promise, the delivery Work it covers, the promised characteristic, the observations and measured values, and the relevant window and population. First ask whether the measurement is direct; if another indicator stands in, establish its relation or leave that basis unresolved. Apply the acceptance rule to its inputs and return the supported result on its declared scale. Recover an evaluation-Work or exact application identity when the claim or rule needs it. Use the acceptance result to assign RequirementStatus when its rule applies. Assess EvidenceStatus from the evidence for the stated use. Identify a durable verdict assertion when another use needs one. Plainly: met, not met, or cannot judge.”
F.12:End
Last Updated: 2026-11-22 — upstream FPF commit 7a93d8f8 (github.com/ailev/FPF)