35.1Who the target public is
Two populations must be named before any claim of representation can be assessed: the population actually consulted, and the population the model is intended to serve. The first is fixed by elicitation—who was asked, and how answers were gathered. The second is fixed by intended use. Merging the two lets a description of consultation quietly answer a question about deployment.
Once the two are kept apart, the accountability question at this transition is narrow: whose input entered the record? Elicitation procedures alone do not identify whose responses actually entered the record; the invited population and the responding population must be distinguished. Conversely, naming the served population requires a specification of use that no elicitation procedure supplies. A well-identified consulted population is therefore necessary for a claim of consultation, but not sufficient for a claim about who benefits.
Collective Constitutional AI shows the distinction under narrow evidence. Huang et al. (2024) report public input carried into principles, training and evaluation, and lower evaluated bias than a developer-principles baseline with comparable language, math and helpful-harmless performance (Huang et al., 2024). The reported comparison does not establish that every other training factor was held fixed; it contrasts two training conditions, not two beneficiary populations. "Comparable" covers those three measures, not capabilities the evaluation did not test. Recruitment procedure is not in view at this level of evidence, so the consulted population cannot be fixed more precisely.
Suppose, conditionally, that a project recruited a defined sample yet trained a system for a wider market: consultation would then be bounded by the sample while service is bounded by deployment. Whether the two coincide is contingent, and training results do not settle it; a match would have to be argued from the deployment specification and the recruitment record, taken separately.
Fixing the consulted population does not, by itself, fix whose endorsement has been recorded. Agreement obtained from participants licenses only the claim that those participants endorsed the formulation under the elicitation conditions used. It does not license the claim that a wider public endorses the principles. Membership in a named public is not sufficient for endorsing a particular wording, and absence from the consulted group supplies no evidence either way. Silence from an unobserved group is not consent; nor is it dissent. It is simply a gap in the record.
A second limit concerns what a tally of contributions can show. A count of submissions compares the volume of input received across some period or channel. It does not compare agreement rates, representativeness, or the strength of objections, and it cannot reveal how many people in the intended population held a view but never submitted one. Treating that count as a measure of public support confuses what was counted with what is being claimed.
Consider a conditional supposition: suppose a future consultation drew participants from every region the model is meant to serve. The evidential base would widen, and claims about the range of perspectives heard would become better supported. It would remain possible, however, that unobserved members of those populations dissent from the resulting principles. Broad geographic coverage is helpful, but it is not sufficient to turn sampled agreement into universal endorsement.
The practical consequence is that two scopes must be reported separately. The principles may govern a model intended to serve everyone; the authorization for those principles extends only as far as the people actually consulted. That divergence is a question for the later evaluation and accountability stages, not something the consultation record resolves. Claims about endorsement should therefore name their limited population, and any wider application should be carried forward as an open question rather than inferred from participation.
35.2Collecting principles
Collecting a contribution and accepting it into the set used for a later stage are two different acts. Collection records what a participant put forward, in the terms and conditions in which it was offered. Acceptance decides which formulation the next stage will actually receive. The first is accountable to the elicitation: was the contribution captured without repair? The second is accountable to selection: what rule brought this formulation into the working set, and what did that rule change?
The invented editorial example shows why the distinction is not clerical. A participant proposes to explain a recommendation when it affects access to a service; a rewrite offers to always explain every output. Suppose the rewrite is the version accepted for a later stage. It is not the original with clearer wording: it drops the service condition and widens the target from recommendations to every output. The scope change is carried out by wording alone, so the accepted item and the collected contribution are not the same object.
That gives the distinction its test. Keeping the service condition is necessary for faithful collection, but it is not sufficient for acceptance: an accurately recorded contribution can still be excluded, and the ground for exclusion is a separate decision that must be stated. Nor is acceptance a finding that the formulation is endorsed by the target public; it establishes only that this is the formulation a later stage will see.
Whether the altered scope affects anything downstream is possible rather than inevitable. A comparison that claims it does must fix the task and inputs and vary the encoding alone; otherwise the two principles differ in more than wording. The set used for a later stage counts the formulations made available; it does not measure how many contributions were gathered or how many conditions were preserved. Treating one as a measure of the others would hide the transition.
Not every principle entering a collection survives into the set used for training. In the editorial example, the participant's condition—that a recommendation affecting access to a service should be explained—may be set aside in favor of the broader rewrite, "always explain every output." The dropped version still bears on an audit: it records what was collected and why it was not carried forward.
That claim is weaker than saying every collected item must affect the model. Exclusion from the training set is neither necessary nor sufficient for erasing an item's evidentiary role. It is not sufficient, because a rejected item may still show which alternatives were considered and what the final set gave up. It is not necessary, because an item that does reach training can still have been collected under uncertain scope, or can have lost its conditions at a later rewrite. What an audit needs is a reconstructable path from collection to use, including the steps at which items fell away and the reason given at each step.
Merging raises a different question again. The merged statement may be the one carried into training while both originals remain necessary to show which conditions each contributed. If the service condition disappears in the merge, a reader consulting only the trained version sees a broader obligation than the participant proposed. That expansion is a property of the rewrite; it does not by itself show that anyone intended to widen the scope, nor that widening was inevitable at every later stage.
The consequence is a record-keeping requirement rather than a training requirement: preserve what was collected, what was rejected or merged, and the stated reason in each case. As a conditional supposition only, if such records are absent, an auditor comparing outcomes with the final principle set may find that the set alone cannot show which narrower alternatives were available and discarded. That failure of reconstruction, not rejection as such, is what weakens the audit.
35.3Selecting and rewriting them
Selection and rewriting change different things. Selection decides which proposals survive into the training material; rewriting decides what wording carries a surviving proposal forward. Because a rewrite may alter specificity, applicability conditions, and strength at once, these properties should be tracked separately rather than treated as a single measure of fidelity. In the invented editorial example, the step from the participant's condition that a recommendation affecting access to a service should be explained to the broader rewrite "always explain every output" both drops the service condition and widens the object from recommendations to every output. Selection produces a comparable effect by omission: a dropped proposal leaves the represented set narrower than the elicited one, although the wording of the survivors is untouched.
The resulting rule is broader, but breadth is not consensus. The original states a case in which an explanation is owed: a recommendation affecting access to a service is a sufficient trigger for the duty as recorded, not a claim that no other case could ever warrant an explanation. The rewrite replaces that trigger with one that holds for all outputs. That is a possible editorial policy; it is not the inevitable consequence of the participant's proposal, and its strength cannot be attributed to the participant without a record supporting the change.
Diagnosis of such changes also requires a fixed comparison. The task must be the same — which outputs demand an explanation — judged over the same input cases and the same understanding of service access. Otherwise the two wordings are being measured over different decision problems, and disagreement between them is uninterpretable. A count of outputs covered likewise compares editorial scope, not the number of principles participants endorsed.
So the accountability question at this transition is which property changed — scope, condition, or strength — and which editorial act produced it.
Because no checked before-and-after record is supplied, the rewrite here is explicitly hypothetical. In the invented editorial example, a participant proposes to explain a recommendation when it affects access to a service; a rewrite says, “Always explain every output.” The second sentence is not simply a stronger version of the first. It removes the service-access condition and universalizes the trigger. Under the first wording, an effect on access is sufficient to require an explanation; it need not be the only sufficient trigger. Under the second, the trigger is every output. Thus the rewrite changes the set of cases to which the principle applies, while also discarding the reason attached to that set.
Suppose, only for this comparison, that the downstream task is to generate explanations for recommendations. Holding the task and inputs fixed, the two wordings direct different behavior: the first asks for an explanation in a bounded class of cases, the second asks for one across the entire output space. What is compared is the scope of the obligation, not the quality or length of the explanations produced. A later training procedure might turn either wording into examples, preferences or objectives, so the rewritten text is a possible source of changed behavior, not by itself proof that behavior changed.
There is also a countercondition worth preserving. A universal instruction could be chosen deliberately to cover cases the original condition missed, such as harms outside service access. That expansion may be justified, but it is a different principle. Without the original condition retained alongside the rewrite, the record cannot show whether the wider scope was intentional or an unintended loss. The safe inference is therefore limited: this hypothetical rewrite can expand scope and remove a condition; whether that improves or degrades a system depends on how the training and evaluation stages use it.
35.4From principle to training signal
A training procedure operationalizes principles through particular examples, preferences or objectives. At that transition a principle stops being a statement and becomes a signal: some outputs are rewarded, penalized, or ranked as preferable. The operative question is which instances the procedure treats as carrying the principle, and how it resolves cases where two principles point to different outputs. In the case already in view, Huang et al. (2024) report public input carried into principles, training and evaluation, with lower evaluated bias than a developer-principles baseline (Huang et al., 2024).
Operationalization leaves choice rather than determination. Holding the task and the inputs fixed, one principle can be encoded as preference comparisons, as constraint penalties, or as curated examples; those encodings weight and contrast outputs differently, so trained behavior may differ even though the underlying classification problem is unchanged. Fixing task and inputs is what licenses the comparison, since the variation then belongs to the encoding. Correspondingly, convergence between two encodings would be evidence, not proof, that the principle survived translation into a training signal.
The necessary conditions and the sufficient ones come apart here. A principle may affect behavior through context or instructions, or through prior generalization, so a principle-bearing training signal is not generally necessary; it is far from sufficient. Examples may underrepresent the principle's difficult cases, and an objective may resolve conflicts in ways the principle leaves open. Influence by the principle is therefore possible; faithful adherence is not inevitable.
The quantity carrying the evidential weight is evaluated bias. It compares two training regimes that differ in the source of their principles, as measured on the authors' evaluations. Lower bias therefore speaks to relative standing on those measures; it does not by itself establish universal fairness, causal legitimacy for public input, or comparable performance in every capability.
Once a principle has been written down, the accountability question at the next transition is narrower than it first appears: what does that documentation license a reader to claim about behavior? The presence of a principle in documentation does not alone demonstrate that all generated behavior follows it. A written principle records an intention and a scope of application. Behavior is produced by a trained model in response to particular inputs, and conformity in one context does not transfer by itself to another. That gap is not evidence of failure; it is the space in which the training transition must be examined on its own terms.
Two qualifications sharpen the point. Documentation is not a sufficient condition for principled behavior: a model may depart from the principle in cases the training signal did not address, or where competing principles apply. Nor is documentation a necessary condition: outputs can align with a principle that no document states, because the regularity arose from other features of training. What documentation supports is a claim about intent and scope, not a universal claim about generation. Whether conformity is possible, likely, or inevitable in a particular case is a separate question, settled only by evidence about that case.
Application follows the same discipline. When two encodings of one principle — prose in a constitutional text, for instance, or judgments recorded over candidate outputs — are compared, the task and the evaluation inputs must be held fixed and only the encoding varied. Otherwise the comparison conflates the form of the principle with changes in what the model was asked to do. As a conditional supposition, take a scheme such as Collective Constitutional AI as a prospective design rather than as evidence of outcomes: the principle's presence specifies expectations to be tested, not a retrospective finding that trained behavior conforms.
35.5Comparing trained models
The source supports one bounded comparison of trained models. Huang et al. report lower evaluated bias than a developer-principles baseline, with comparable language, math and helpful-harmless performance (Huang et al., 2024). The two models represent different encodings of the same kind of object: principles carried into training through the public-input process described above, versus principles written by the developer. The evaluation tasks are held fixed; what differs is how the principles were derived and fed into training. That is what the comparison tests.
"Lower evaluated bias" therefore names a difference in scores on the bias evaluation used, not a general quantity of bias and not the elimination of biased behavior. It is an outcome of the evaluated samples, which is stronger evidence than the presence of a principle in documentation, and still narrower than a claim about all generated behavior.
"Comparable performance" is similarly a comparison of measured results. It indicates no reported degradation on the language, math and helpful-harmless evaluations that were run. It does not establish statistical equivalence or no degradation on every evaluated component, because capabilities outside those evaluations were not compared. No effect sizes or evaluation details are available here, so the magnitude of either difference cannot be stated.
The inference available is one of possibility, not inevitability: a model trained through the public-input process can score better on a bias evaluation while retaining performance on other measured tasks. The comparison does not make this outcome necessary; a different training change might improve a developer-principles model, and only the particular baseline evaluated was compared. Finally, the score settles how the models differed on these tests, not whether the tests cover the behavior that matters — a question about the evaluation itself.
A narrower evaluation of trained behavior can show that one model produces different outputs under the evaluated conditions. It cannot, by itself, license a model-level gain in public legitimacy. The quantity at issue compares outputs on a fixed task with specified inputs; legitimacy asks whether the people affected had standing in eliciting the principles, whether transitions from contributions to training examples were accountable, and whether evaluation covers the claims being made. Fix the task and inputs before comparing encodings, since encodings that alter what the model receives change what the comparison measures.
Suppose, as a hypothetical illustration, that one trained model performs better than another on a behavior evaluation while both use the same elicited principles. The result would establish a difference in measured behavior, not that the better-performing model is more legitimate. The evaluation may be necessary evidence for one narrow claim, but it is not sufficient for the broader one. Nor would the result show that legitimacy gains are inevitable: they remain possible only if the omitted transitions also support the relevant standing and accountability.
The scope question can be applied recursively. If the task is narrowed again, for example to a single response type or a restricted input set, the measured difference shrinks in claim as well as in size. A quantity that compares output accuracy under restriction says nothing direct about population coverage, principle selection, or the treatment of disagreements during training. Keeping those distinctions separate prevents a behavior result from being redescribed as an institutional result.
35.6A minority whose demand changes
Consider a prospective update to the record used earlier: the minority contributor who supplied a principle now changes one of its conditions. The change is not an extra preference appended to a growing pile of inputs; it revises the terms under which the principle was offered. Because the contributor stands in a minority position, the revision should not be discounted as volatility or averaged away against the greater number of unchanged submissions. Yet the update creates a documentary problem that stands on its own: the record now holds two formulations of the same principle, and the change must be resolved before downstream objects that cite the principle can be checked against a single version.
Two interpretations of the revision have to be kept apart. On one reading, supposing the first formulation misstated the contributor's intent, the later statement corrects the elicitation record and the earlier version should be marked as superseded. On the other, the contributor genuinely changed position; here the earlier version remains a faithful account of what was said at the time, and the new version records a later demand. The same edit to the condition can therefore repair an earlier error or register a real shift, and the accountability response differs: misdescription calls for correcting how the input was captured, while a change of mind calls for updating the proposed record without rewriting its history.
A further qualification concerns what the changed condition governs. It may be a necessary condition for one application of the principle — without it, that application does not follow — while still not being sufficient for any evaluation verdict, which depends on other conditions and cases as well. Nor does inconsistency between the old and new versions inevitably propagate into the derived material; that is possible wherever a documented derivation cites the superseded wording and is not revisited. Distinguishing necessary from sufficient and possible from inevitable keeps the trace of the revision proportionate to what the change actually touches.
In the prospective case, suppose a contributor later revises one condition in the principle already submitted. The first object to revisit is the principle record itself, since the revision changes what the record states. Next come any documented principles that cite, paraphrase, or operationalize the same condition. Evaluation cases derived from it also fall within scope: a case that exercises the condition must be re-examined, while a case whose design does not touch it can remain provisionally usable. This asymmetry is the point. Dependence is graded rather than binary. Some derived objects restate the condition explicitly; others combine it with other constraints, so only part of the object requires revision. A case built around the condition alone is more exposed than one where the condition merely narrows a margin. Priority within that scope can follow exposure: objects that encode the condition as a governing rule warrant earlier review than those that inherit it only through a summary.
The revised condition is therefore a necessary trigger for reconsideration of dependent objects, but not a sufficient warrant for claiming that trained behavior has changed. That claim would require evidence drawn from the evaluation cases themselves, not from the edit alone. Nor is behavioral change inevitable: the condition may be redundant with other constraints, in which case its revision leaves outcomes stable. The possibility exists; it is not guaranteed.
A tally of dependent records, if kept, would compare how many derived objects reference the condition. It would not measure how much any trained parameter reflects it, and no one-to-one attribution from statement to parameter should be inferred. The earlier record need not be discarded; it is superseded only with respect to the condition in question. Where the dependence relation is documented, reconsideration has a defined scope; where it is not, the scope remains open rather than assumed.
35.7Traceability without a false mandate
A prospective record chain would keep four kinds of entry distinct rather than merge them into a single provenance claim. A contribution record states the condition offered and the later change to it; an editorial record notes which documented principles were drafted, revised, or left untouched in response; a training-use record marks whether any derived object from that exchange entered a training corpus and in what form; an evaluation record marks which test cases rest on the condition. Each entry is scoped to what its evidence supports. Where no entry exists, the chain records a gap rather than filling it.
Applied to the later contributor's change, the chain answers a narrower question than causal attribution. From the editorial record one can identify documented principles that depend on the condition and therefore need review; from the evaluation record, cases whose expected behavior relies on it. The result is a dependency list of derived objects to reconsider. It is not a claim about individual trained parameters: a parameter-level account would require evidence these records do not supply, and the presence of a dependency does not establish that the change produced any particular parameter state.
The chain is necessary for inspectability, since without it a reviewer cannot tell which transitions occurred or where the evidence stops, but it is not sufficient for the stronger inferences. A nearly complete chain may still omit undocumented exchanges, and a recorded change does not show that it was adopted. Equally, propagation is possible rather than inevitable: whether the revised condition reaches training material or evaluation cases depends on separate editorial and usage decisions, each recorded as its own transition. What the chain yields is therefore a bounded set of review targets, with the evidence for each stated at the step where it exists.
A traceable chain makes transitions inspectable; it does not convert consultation into a general mandate for use. Inspectability is a necessary condition for assigning responsibility across the transitions—target population, elicitation, editorial principles, training, and evaluation—but it is not a sufficient condition for authorizing any particular application. A record showing where a condition changed tells a reader where an accountability question arises; it does not answer that question in advance.
Consider the invented illustration: a later contributor revises one condition in the proposed record. The traceable response is to ask which documented principles or evaluation cases depend on that condition. Dependence here may run many-to-one and one-to-many: several cases may rely on the same condition, and one principle may bear on many cases. Consequently no one-to-one attribution from the revised condition to trained parameters is warranted. The revision makes some outcomes possible and others less likely; inevitability is not established, nor is the size of any difference.
The scope of authorization follows the same logic. Consultation supports applications within the elicited aims and documented principles, not every downstream use to which the material might be put. Where a proposed use falls outside that scope, the traceable chain locates the gap precisely; it does not fill it. A reader who can follow the transitions therefore gains grounds for a narrower claim, not a broader one.
Comparisons of encodings require the same caution. A claim that one encoding preserves or alters meaning holds only when the task and the inputs are held fixed; if either changes, the observed difference confounds the encoding with the change in task. Supposing, hypothetically, that such a difference were observed, it would still need dependency tracing across principles and evaluation cases before any accountability claim could attach to it.