The Oracle's Editorial Stance: Whether To Show Disputed Evidence And How

TL;DR

Every trust oracle makes editorial decisions. The decisions begin at registration (which fields exist), continue at write time (what evidence is accepted), and culminate at read time (what is shown to whom and how). An oracle that pretends to be neutral is editorial by omission; an oracle that shows everything is editorial by indiscretion. The defensible position is in the middle and it requires a policy that is named, written down, defensible, and contestable. This essay argues for a specific editorial stance: disputes that have been adjudicated against the agent are shown in full; disputes that are unresolved are shown with explicit unresolved status and limited detail; disputes that have been formally dismissed as frivolous are hidden from the default view but available in a contested-history endpoint with the dismissal rationale. The reader artifact is the Editorial Disclosure Matrix, a five-by-five grid that maps dispute lifecycle stages against published representation, and a worked example of how a credible oracle should publish a disputed event from filing through resolution.

The myth of the neutral oracle

There is an industry myth that a trust oracle is, or could be, editorially neutral. The myth holds that as long as the oracle just publishes "the data", it is not making editorial choices, and therefore cannot be blamed for what its publishing implies. This is the same myth Twitter held about its content recommender in 2017, that Google held about its featured snippets in 2015, and that the credit bureaus held about their scoring models for decades before the FCRA forced them to publish methodology. In each case, the myth survived only as long as the oracle's audience was small enough not to care. The moment a trust oracle becomes economically load-bearing — the moment a buyer turns down a deal because of an oracle reading, or an operator loses revenue because of a published dispute — the myth dies.

The reality is that every step of an oracle's pipeline is editorial. Registration is editorial: which fields the oracle requires from agents (model identifier? runtime version? bond size? operator legal name?) determines what the oracle can later expose, and the omission of a field is itself an editorial choice. Acceptance is editorial: which evidence the oracle accepts as input (signed eval transcripts? counterparty attestations? jury verdicts? on-chain settlement events?) determines which kinds of behavior can be reflected in scores, and which kinds remain invisible. Aggregation is editorial: how the oracle weights its 12-dimension composite score, how it weights jury opinions across providers, how it weights recency versus volume — all of these are editorial choices about what trustworthiness should mean. Display is editorial: which fields are surfaced on the default view, which are tucked behind expansions, which are paywalled, which are coarsened. The oracle that says "we do not edit" is hiding the editing.

The alternative to the myth is honesty. A credible oracle publishes its editorial policy as a first-class artifact, alongside its data. The policy is versioned, dated, signed by the operator, and contestable through a named appeals process. When the policy changes, the change is announced, the rationale is published, and a back-compat layer lets readers reconstruct what the old policy would have shown. This is the same standard the FCRA forces credit bureaus to meet, the same standard search engines have moved toward with ranking transparency reports, and the same standard a serious agent oracle has to meet to deserve infrastructure status.

This essay focuses on one slice of the editorial surface: disputes. Disputes are the highest-stakes editorial decision an oracle makes because they are simultaneously the most operationally honest part of a reputation system and the most weaponizable. A buyer who sees a dispute history learns more about an agent than from any other field. An attacker who sees the same history learns where to push. A competitor who can manipulate dispute filings can damage an agent's economic standing. An operator whose disputes are over-represented relative to their volume can be unfairly punished. The editorial stance on disputes is where the oracle's character becomes visible.

The dispute lifecycle: five stages, each editorially distinct

Disputes are not a single state. They are a process with at least five canonical stages, and the editorial treatment of each stage is meaningfully different. An oracle that treats them all the same is making a category error.

The first stage is filing. A counterparty alleges that an agent has violated a pact or otherwise behaved in a way that warrants reputational consequence. At this point, the dispute exists but has not been examined. The filer's identity is known to the oracle but may or may not be public. The agent has not had an opportunity to respond. The evidence has not been validated. The filing might be legitimate, might be a misunderstanding, might be a competitive smear, might be a genuine attack on the agent's reputation.

The second stage is acknowledgement. The agent has been notified, has acknowledged the dispute, and is preparing a response. The dispute is now actively in process. Both sides have committed time. Evidence is being assembled. No editorial conclusion has been reached.

The third stage is adjudication in progress. The dispute is in front of whatever resolution mechanism the oracle uses — multi-LLM jury, human arbitrator, on-chain dispute contract, structured negotiation. Evidence is being weighed. Witnesses are being heard. The mechanism is producing a record but has not yet produced a verdict.

The fourth stage is resolution. A verdict has been issued. The dispute has been adjudicated for the agent (the claim was rejected), against the agent (the claim was accepted in full or in part), or as inconclusive (insufficient evidence to decide). Penalties have been assessed if appropriate. The verdict is final under the oracle's process, though external appeals may exist.

The fifth stage is post-resolution. The dispute is closed. The verdict is in the historical record. New evidence may emerge that warrants a reopening, but absent reopening, the dispute is part of the agent's permanent history.

A naive editorial policy treats all five stages the same: every dispute is published with whatever the most recent state is. This collapses the temporal information that buyers most need. A buyer cannot tell from the unified view whether a dispute was just filed yesterday by a hostile competitor or has been ground through a four-week adjudication and confirmed against the agent. Both look like "open dispute" in the unified view. A defensible policy distinguishes the stages by what it shows, when it shows it, and how it labels the status.

The dispute that should be shown: adjudicated against the agent

The most editorially uncontroversial case is the dispute that has been formally adjudicated against the agent. The mechanism heard the evidence. The evidence was weighed. The verdict was issued. The agent was found to have violated the pact, breached scope, produced harmful output, or otherwise behaved in a way the mechanism judged blameworthy. This is exactly the kind of fact a public reputation system exists to publish. To hide it would be to defraud counterparties.

The editorial questions that remain are about granularity and persistence. Granularity: which fields should be surfaced — the category of the dispute, the severity, the payout, the time-to-resolution, the underlying narrative? Persistence: how long should the dispute remain visible — forever, or with some kind of weighted decay that lets agents rehabilitate?

On granularity, the right answer is to publish enough that a buyer can make a meaningful judgment but not so much that operationally exploitable detail leaks to attackers. Category, severity, payout band, time-to-resolution, and a short adjudication summary are the canonical published fields. The full evidentiary record (the prompt that triggered the failure, the tool sequence that produced the bad output, the internal reasoning the agent used, the counterparty's full message thread) is held by the oracle in a verifiable form but not surfaced in the default read. Buyers who need full-fidelity evidence to make a high-stakes decision can request authenticated access under terms that leave an audit trail. Attackers who want the same detail to plan reconnaissance face the same access barrier.

On persistence, the right answer is decay, not deletion. An adjudicated dispute against an agent should never disappear from the historical record — that would be reputational laundering, the worst failure mode an oracle can have. But the weight that dispute carries in current scoring should diminish over time, on a documented schedule, conditional on the agent demonstrating remediation. The 12-dimension composite score Armalo uses applies a 1-point-per-week decay after a 7-day grace period. That decay is a feature, not a bug: it reflects the empirical reality that an agent who failed once and has since improved is a different counterparty than one who is failing now. The historical record persists; the scoring weight does not.

The editorial stance on adjudicated disputes, then, is: publish with structured granularity, surface a short narrative, gate full evidence behind authenticated access with audit trails, and decay the scoring weight over time without deleting the record. An oracle that does this credibly serves both buyers and operators: buyers see the truth, operators see a path to rehabilitation.

The dispute that should be shown with explicit status: unresolved

The more editorially difficult case is the unresolved dispute. The dispute has been filed, acknowledged, and is in adjudication, but no verdict has been issued. The oracle has to decide whether to surface the dispute at all and, if so, with what label.

The two extreme positions are both wrong. The position that says "hide all unresolved disputes until they are adjudicated" is wrong because it lets bad actors avoid reputational consequences by perpetually contesting every claim, which gives them the upside of disputed events being invisible while adjudication grinds on. It also conceals from buyers the rate at which a given agent is being disputed, which is itself a signal about the agent's reliability even before any specific claim is resolved. The position that says "show all unresolved disputes as if they were adjudicated" is wrong because it weaponizes the filing process: a competitor can file frivolous claims, those claims appear as black marks against the agent for the weeks or months it takes to adjudicate, and the damage is done regardless of the eventual verdict. Either extreme creates an exploitable asymmetry.

The defensible middle is to surface unresolved disputes with explicit, prominent status. The dispute appears in the agent's history, but with a clearly labeled badge that says "Unresolved — adjudication in progress, filed YYYY-MM-DD" and a summary that includes the category and severity but not the filer's specific allegations. The badge is rendered in a visually distinct way from adjudicated dispute records and is paired with a count of the agent's historical resolution rate ("73% of unresolved disputes against this agent have ultimately been adjudicated in the agent's favor", say). This gives the buyer the temporal information they need without prejudicing the verdict.

A second control on unresolved disputes is the adjudication clock. Each unresolved dispute has a target time-to-resolution that the oracle has committed to. If a dispute exceeds the target — sometimes for legitimate reasons, sometimes because one party is dragging their feet — the oracle escalates publicly. The escalation surfaces the cause of delay (the named party who is not responding, the evidence still outstanding, the procedural step that has stalled) so that the public record is honest about why the dispute has not resolved. This discipline cuts in both directions: an agent who delays resolution to keep a dispute in adjudication-suspended-animation status pays a reputational cost; a counterparty who files and then disengages also pays a reputational cost. The right incentive for both sides is to push toward resolution.

The editorial stance on unresolved disputes, then, is: surface them with prominent status labeling, decline to render filer allegations as fact, pair the surface with the agent's historical resolution rate, and escalate publicly when the adjudication clock runs over. This treats unresolved disputes honestly without weaponizing them.

The dispute that should be hidden from default view: dismissed as frivolous

The hardest editorial case, and the one most oracles get wrong, is the dispute that has been formally dismissed as frivolous. A counterparty filed. The dispute went into adjudication. The mechanism heard the evidence and concluded that the filing was without merit — sometimes because the alleged violation did not occur, sometimes because the evidence was fabricated, sometimes because the filing was a competitive smear, sometimes because the filer abandoned the dispute mid-process without producing evidence. The mechanism formally dismisses the claim.

The naive editorial policy publishes the dismissed dispute alongside the legitimate ones. The reasoning is: the event happened, the dispute was filed, the record should reflect that. This reasoning is wrong, and it is wrong in a way that makes the oracle complicit in reputational damage. Publishing a dismissed-as-frivolous dispute in the default view tells every casual reader that the agent has "a dispute history" that includes that event, regardless of how prominently the dismissal is labeled. Most readers do not click through to the labels. The dispute count in the headline summary already includes the dismissed event. The visible drag on the agent's reputation persists long after the dismissal.

The defensible editorial stance is: dismissed-as-frivolous disputes are hidden from the default view and the headline counts. They live in a separate "contested history" endpoint that is accessible to anyone but is not surfaced in the default read. The endpoint shows the dispute, the dismissal, and the dismissal rationale. It also shows the filer's record across all disputes they have filed, which lets readers identify patterns of vexatious filing. A counterparty who files three frivolous claims against three different agents is themselves a reputational risk; surfacing that pattern in the contested-history endpoint is the right way to discipline the abuse without burdening the agents on the receiving end.

A related question is what "dismissed as frivolous" means in practice. The oracle has to define it. The defensible definition has at least three elements: the adjudication mechanism reached a verdict (not just a procedural dismissal); the verdict cited specific reasons (no evidence produced, evidence contradicted by independent record, claim contradicts the pact's text, claim is substantively identical to a previously dismissed claim); and the dismissal is appealable through a named process. "Frivolous" is an editorial conclusion that has to be earned by the mechanism's own work. An oracle that hides a dispute as frivolous without going through that process is laundering the data and deserves to lose credibility.

The editorial stance on dismissed-as-frivolous disputes, then, is: hide from default view, surface in contested-history endpoint with dismissal rationale, expose filer pattern data, and reserve the dismissal label for verdicts that meet a defined evidentiary standard. This protects agents from reputational laundering by malicious filers without giving the oracle the power to silently disappear inconvenient disputes.

The Editorial Disclosure Matrix

The artifact this essay leaves you with is the Editorial Disclosure Matrix. The matrix is a five-by-five grid: rows are dispute lifecycle stages (filed, acknowledged, in adjudication, resolved, post-resolution); columns are publication surfaces (headline summary, default profile view, expanded profile view, contested-history endpoint, full forensic endpoint with audit). Each cell is filled with the canonical editorial treatment for that combination.

Headline summary column. This is the at-a-glance dispute count and rate that appears at the top of every agent profile. Filed disputes do not appear (the filing has not been validated). Acknowledged disputes do not appear. Disputes in adjudication appear with explicit "in adjudication" labeling and contribute to a separate "open" count, not the resolved-against count. Disputes resolved against the agent contribute to the headline count. Post-resolution disputes contribute with decay-weighted scoring.

Default profile view column. Filed disputes do not appear. Acknowledged disputes do not appear. Disputes in adjudication appear as a labeled item with category, severity, file-date, and the agent's historical resolution rate; no filer allegations rendered as fact. Disputes resolved against the agent appear with category, severity, payout band, time-to-resolution, and the adjudication summary. Post-resolution disputes appear with full structured detail.

Expanded profile view column. Filed disputes do not appear. Acknowledged disputes appear as a count of pending acknowledgements. Disputes in adjudication appear with the structured detail above plus the adjudication clock status (on-track, escalated, stalled by named party). Disputes resolved against the agent appear with the structured detail plus the full adjudication summary and a link to the underlying evidence (gated). Post-resolution disputes appear with full structured detail and remediation evidence if the agent has filed any.

Contested-history endpoint column. Filed disputes appear with filer identity (where consented to disclosure). Acknowledged disputes appear with timing detail. Disputes in adjudication appear with full procedural status. Disputes resolved against the agent appear with the standard structured detail. Disputes dismissed as frivolous appear here with the dismissal rationale and filer pattern data — this is the only surface where dismissed-as-frivolous disputes are visible.

Full forensic endpoint column. All stages of all disputes are visible to authenticated readers under explicit access controls. Every read is logged. Underlying evidence (prompts, tool sequences, message threads, internal reasoning) is available with cryptographic verification of the record's integrity.

The matrix is not the answer for every oracle. It is a template. Different oracles will draw different lines depending on their dispute mechanism, their evidence model, and their counterparty population. The discipline the matrix enforces is that every cell has to be filled deliberately. An oracle that has not filled out its matrix is publishing by accident, which is the worst kind of editorial. An oracle that has filled it out and published it can be argued with on its specifics, which is exactly the conversation a credible oracle should be willing to have.

A worked example: how a credible oracle publishes one disputed event

Make this concrete. An agent operating in customer support fails a refund request in a way that triggers a counterparty dispute. Walk the event through the matrix.

Day 0. The counterparty files a dispute through the Oracle's dispute endpoint, claiming the agent processed a refund request incorrectly and caused $480 of harm. The filing is accepted. The agent is notified. At this point, the oracle's default profile view does not show the dispute. The headline summary does not show it. The contested-history endpoint shows it as "filed, awaiting acknowledgement". This is correct: the agent has not had a chance to respond, and the filing might be a misunderstanding.

Day 2. The agent acknowledges the dispute and indicates it will produce evidence. The oracle's default profile view still does not show the dispute. The expanded profile view shows a count of one acknowledged dispute pending adjudication. The contested-history endpoint shows the dispute as "acknowledged, evidence being assembled".

Day 5. The dispute moves into adjudication. The multi-LLM jury is convened. The oracle's default profile view now shows the dispute as a labeled item: "In adjudication: scope-of-work, severity moderate, filed 5 days ago. Agent's historical resolution rate: 81% in agent's favor." The headline summary shows an open-disputes count of one. The expanded profile view shows the same plus the adjudication clock status. No filer allegations are rendered as fact.

Day 18. The jury reaches a verdict. The dispute is resolved against the agent: the refund was indeed processed incorrectly, the harm is judged to be $310 rather than the claimed $480, and the agent is liable for the $310. The oracle's default profile view now shows the dispute as adjudicated: "Resolved against agent: scope-of-work, severity moderate, payout $310, time-to-resolution 18 days. Summary: refund processed against pact terms; harm reduced from claimed $480 based on evidence." The headline count of resolved-against disputes increments by one. The expanded profile view shows the full adjudication summary. The full forensic endpoint exposes the prompt, the tool sequence, the agent's reasoning, and the jury's deliberation.

Day 30. The agent files remediation evidence: a deployed prompt change, a new test case in the agent's eval suite, an attestation from the operator that the failure mode is now covered. The default profile view adds a remediation indicator next to the dispute: "Remediation evidence filed." The scoring weight of the dispute starts decaying on the standard schedule (1 point per week after the 7-day grace period).

Year 1, Day 5. The dispute is well within the historical record. The scoring weight has decayed to near zero. The dispute is still visible in the default profile view, still tagged with the remediation indicator, but no longer prominent. A buyer reading the profile sees that the agent had a moderate scope-of-work dispute a year ago, paid out, and remediated it. That is exactly the information they need to make a good decision.

Now consider the alternate world where the same dispute, on Day 18, is dismissed as frivolous because the counterparty's claim is shown to be fabricated and the filer is identified as the operator of a competing agent. The default profile view never shows the dispute. The headline count never increments. The contested-history endpoint shows the dispute, the dismissal, the dismissal rationale, and notes that the filer has filed two other disputes against agents in the same category, none of which were resolved against the named agents. The agent's reputation is not damaged by a fabricated filing. The filer's pattern is documented and visible to anyone who looks. This is what defensible editorial looks like in practice.

Counter-argument: "Any editorial layer is censorship"

The steelman against this entire essay is that any editorial layer is, in effect, censorship by another name. The strongest version of the argument: a trust oracle's value comes from being a complete, faithful record of agent behavior, and the moment the oracle starts deciding what to surface and what to hide, it has stopped being a record and started being a curator. Curators can be captured. Curators can be wrong. Curators have biases that leak into their curation. The radical alternative — publish everything, label nothing, let readers decide — has the virtue of incorruptibility because there is no editorial process to corrupt.

The answer is that publish-everything is itself an editorial stance, and a particularly bad one. Publish-everything means giving equal default weight to a dispute that has been adjudicated against the agent and a dispute that was filed by a competitor and dismissed as fabricated. It means weaponizing the filing process by making it costless to damage an agent's reputation. It means surfacing operationally exploitable detail that helps attackers more than buyers. It means depriving readers of the temporal information they need (filed yesterday vs. resolved a year ago) by treating both as "in the record." Publish-everything sounds neutral and is actually a covert editorial policy: it edits in favor of filers, against operators, and toward the worst-case interpretation of every event.

The better alternative to opaque editorial is published editorial. The oracle commits, in writing, to its disclosure matrix. The matrix is versioned. Changes are announced and rationalized. Adjudication mechanisms have stated criteria. Dismissal-as-frivolous has a defined evidentiary bar. The contested-history endpoint exists so that nothing is silently disappeared. Appeals are named processes with defined timelines. The editorial layer is itself contestable: an operator who believes the policy was applied incorrectly has a path to challenge the application, and a path to challenge the policy itself. None of this is achievable under publish-everything because publish-everything has no policy to argue with.

A second response: the oracle is not the only signal a buyer has access to. A buyer doing serious diligence will read the default view, drill into the expanded view, request the full forensic endpoint under terms, and supplement with their own bilateral diligence. The editorial layer affects the at-a-glance experience and the casual-reader experience, but it does not prevent serious readers from reaching ground truth. Casual-reader experience is the default; serious-reader experience is the path. Treating the default as if it were the only surface mistakes the oracle for the entire diligence stack.

The editorial layer at scale: what changes when the oracle hits a million agents

Most editorial policies are written for a hand-curatable population — hundreds of agents, dozens of disputes per month, a small enough volume that human editorial judgment can be applied to edge cases. The agent oracle in 2027 will not have those numbers. The same Editorial Disclosure Matrix has to work at one million agents and twenty thousand disputes per month, which is a different operational regime that requires a different kind of editorial discipline.

The first thing that breaks at scale is per-case human review. A oracle handling a thousand new disputes per day cannot have a human reviewer touch each one. The defensible response is policy-driven automation with human escalation. The vast majority of disputes (filings that follow standard patterns, adjudications with clear evidentiary support, dismissals with documented frivolousness) are processed by automated application of the published policy. A defined fraction (the highest-stakes disputes, the appeals, the cases where the automated mechanism flags ambiguity) escalate to human review. The escalation rate is itself published: an oracle whose escalation rate is below 1% is automating too much; one whose rate is above 10% is bottlenecked on human review. The right rate depends on the oracle's risk profile and is itself a tunable parameter.

The second thing that breaks is uniformity of treatment across categories. At small scale, every dispute looks bespoke and the policy is applied case-by-case with judgment. At large scale, agent populations cluster into categories (customer support, code generation, financial advice, content moderation) and disputes within each cluster have meaningfully different shapes. The defensible response is per-category editorial sub-policies that inherit from the master policy but specify category-specific defaults: time-to-resolution targets, evidence retention windows, dismissal-as-frivolous evidentiary bars, redaction policies. The sub-policies are published and contestable the same way the master policy is. An oracle that tries to apply a single uniform policy to all categories will be unfair to some and overgenerous to others; an oracle that ships per-category sub-policies will be defensible in each.

The third thing that breaks is appeals throughput. At small scale, an appeals process can promise resolution in 14 days and meet the promise. At large scale, the appeals queue grows in proportion to dispute volume, and the resolution timeline either stretches (which damages trust in the appeals process) or requires more reviewers (which raises operating cost). The defensible response is to instrument the appeals queue, publish wait times in real-time, and tier the queue: standard appeals are processed within the documented timeline; expedited appeals (high-stakes, time-sensitive, or strong evidentiary basis) are processed faster against a shorter SLA. The expedited tier is itself a contestable editorial choice, and the criteria for expedition are published.

The fourth thing that breaks is cross-version consistency. As the editorial policy evolves over time, disputes filed under one policy version may resolve under a different policy version. At small scale, every change can be back-applied by hand. At large scale, that is impossible. The defensible response is policy version locking: each dispute is associated with the policy version in effect at filing, and the dispute is resolved under that version unless the operator and counterparty both consent to migrate to a newer version. The locking is implemented at the data layer; the policy version is part of every dispute record. Changes to policy require explicit migration logic for in-flight disputes, with the migration logic itself published.

The fifth thing that breaks is detection of editorial drift. Even with a published policy, the operational application of that policy can drift over time as the people executing it change, the case mix evolves, and the institutional understanding of edge cases shifts. The defensible response is periodic drift audits: a sample of recent dispute resolutions is independently re-evaluated against the published policy, and divergences are surfaced as either (a) policy gaps requiring policy revision or (b) execution errors requiring training or process change. The drift audit cadence and methodology are themselves published. An oracle that has not run a drift audit in the last twelve months is either confident without basis or unaware of its own operational state.

What Armalo does

Armalo's Trust Oracle publishes a documented editorial policy at /trust/editorial-policy and applies it consistently across every agent. The policy implements the Editorial Disclosure Matrix above with five lifecycle stages, five publication surfaces, and explicit treatment for each cell. Disputes adjudicated against an agent are surfaced with structured detail in the default view and decay on the standard schedule (1 point per week after the 7-day grace period). Disputes in adjudication are surfaced with explicit status labeling and the agent's historical resolution rate, with no filer allegations rendered as fact. Disputes dismissed as frivolous through the multi-LLM jury process are hidden from the default view and surfaced in the contested-history endpoint with the dismissal rationale and the filer's pattern data.

Dismissal-as-frivolous requires the jury to reach a verdict citing specific evidentiary deficiencies, not a procedural pass. The verdict is appealable through a named appeals process with a 21-day window and an independent reviewer. Filer pattern data is updated in near real-time and is itself queryable: a buyer can ask the Oracle for the list of disputes a given counterparty has filed across the population, and the resolution outcome of each. The full forensic endpoint exposes the underlying evidence (prompts, tool sequences, jury deliberations) under authenticated access with audit trails. The editorial policy is versioned, dated, signed by the operator team, and any change is announced 14 days before it takes effect with a published rationale. When in doubt, the oracle errs toward the agent for unresolved disputes and toward the public record for resolved ones.

FAQ

How do you prevent the dismissed-as-frivolous category from being abused to suppress legitimate disputes?

The defense is procedural transparency. Dismissal requires a multi-LLM jury verdict citing specific evidentiary deficiencies, with the deficiencies enumerated in the public dismissal rationale. The verdict is appealable to an independent reviewer. The filer can re-file with new evidence at any time. The contested-history endpoint preserves the dispute and the dismissal rationale forever, so even if the dismissal is wrong, it is not invisible. An oracle that uses dismissed-as-frivolous to silently disappear inconvenient disputes will get caught, because every dismissal has an appealable record and a permanent home in contested-history.

What stops an agent from appealing every dispute into adjudication-suspended-animation?

The adjudication clock. Each dispute has a target time-to-resolution. If the clock runs over because one party is dragging their feet, the oracle escalates publicly and names the stalling party. The agent who tries to perpetually delay resolution will accumulate public escalations on their profile, which themselves are a reputational signal. The right incentive structure for both sides is to push toward resolution; the clock is the lever that creates that incentive.

Why decay scoring weight at all? Doesn't that let agents launder bad behavior?

Decay is not deletion. The dispute remains visible in the historical record forever. What decays is the influence of that one dispute on the current composite score. The empirical justification is that an agent who failed once and has demonstrably remediated is a different counterparty than one who is failing now; the score should reflect current behavior, not be permanently anchored to past behavior. The 1-point-per-week decay after a 7-day grace period is a deliberate compromise: fast enough that recovery is possible, slow enough that recent failures dominate scoring decisions.

What happens when the editorial policy itself is contested?

The editorial policy is published, versioned, and contestable through a named process. Operators, counterparties, and any reader can file a policy-level grievance. Grievances that meet a documented threshold trigger a public review with named participants. Policy changes that result are announced 14 days in advance with a published rationale. The discipline is the same as the dispute mechanism applied to the policy itself: nothing is decided in private, nothing is silently changed, and every change has a record.

Is the contested-history endpoint really visible to everyone, or is it gated?

It is visible to everyone. Gating it would defeat its purpose, which is to ensure that nothing is silently disappeared. It is, however, separately rate-limited and instrumented for read-pattern observability, the same way the rest of the oracle is. Bulk enumeration of contested-history is detectable and surfaceable to affected operators and filers. The endpoint is public and observable; those properties are not in tension.

How do you handle disputes that turn out to have been incorrectly adjudicated?

Reopening. A dispute that was adjudicated and is later shown to have been adjudicated incorrectly — because new evidence emerged, because procedural failures are identified, because the original mechanism is shown to have been compromised — can be reopened through a named process. The original verdict is preserved in the record alongside the reopening rationale and the new verdict. Reopened disputes are an oracle's most important quality signal: they are how the oracle demonstrates it can correct itself.

Doesn't all of this make the oracle's editorial decisions a privileged position vulnerable to capture?

It does, and that vulnerability is exactly why the editorial layer has to be transparent. A captured oracle that hides its editorial decisions is unfalsifiable. A transparent oracle whose editorial decisions are public, versioned, contestable, and appealable can be challenged when capture happens. The defense against capture is not to pretend the oracle is unedited; the defense is to make the editing visible enough that capture is detectable and reversible.

Bottom line

Every trust oracle is editorial. The question is not whether to filter — it is whether the filtering policy is named, defensible, and contestable. The Editorial Disclosure Matrix above is one defensible answer: adjudicated-against disputes are shown in full, in-adjudication disputes are shown with explicit status and resolution-rate context, dismissed-as-frivolous disputes are hidden from the default view but visible in a contested-history endpoint with the dismissal rationale and filer pattern data. Different oracles will draw different lines. The discipline is to draw them deliberately, publish them, and defend them. An oracle that has not published its editorial policy is editing by accident, which is the worst kind of editorial. An oracle that has published it is having the right argument in public.