Hobocode Polarization Monitor
How we identify polarizing narratives, test them against their own logic, and watch the ones drifting toward hate
This is the operating manual for the polarization side of our monitor: the gate a narrative has to pass before we code it, the logic test every coded claim then faces, and the tripwires that tell us a protected argument is moving toward the territory of our hate methodology. Everything measured here is protected political speech. That sentence is load-bearing, and this page explains why.
This section animates one worked example in eight steps: the trigger event, the capture, the gate, the canonical claim, the prediction check, the verdict, the mirror, and the score. Each step's full text is in the captions that follow.
The method in practice
Follow one narrative through the method
The event, the frames, and every number below are invented for illustration; no real company or campaign is described. Both frames in this walkthrough are protected political speech, and they stay protected at every step.
The event
An AI company publicly asks the federal government to take roughly a 5% investment stake in it. Coverage is descriptive, the frames are still about the merits, and no camp has claimed the event yet. On the Charge Model this is C1: a trigger has landed on a loaded fault line (§3).
AI firm asks Washington to take a 5% equity stake, citing "shared national interest" in the technology.
The capture
Within a day, two frames appear whose subject is no longer the ask itself. One reads the event as proof of what the company really wants; the mirror reads it as proof of what the government really wants. The event has become evidence about an opposing camp, which is the capture moment, C2, where cards open (§3).
Two prongs, both required
Frame A passes the object check, because its primary object is the company's hidden intent rather than the merits of the ask, and it performs a coded operation, motive attribution. An argument that the stake is simply bad policy, at any temperature, fails the first prong and stays uncoded. Heat alone never codes (§5).
The claim, statable in plain form
To found a card, the coder has to fill one template honestly: the event shows that the actor is, wants, or intends something. If filling it would require adding an imputation the text does not carry, the item is commentary and stays uncoded. Here the sentence assembles from the frame's own words (§5.3).
What else would have to be true?
If loss-shifting were the motive, the world should show two things: a much larger ask, since a 5% stake absorbs 5% of losses and rescues no one, and an ask shaped as downside protection, guarantees or backstops, rather than equity that shares the upside too. We check both against the observable shape of the ask, and both fail (§6).
The verdict and the plain reading
The checks we authored contradict the claim, at moderate confidence, and the card also records our plain reading: we read an ask this small and equity-shaped as seeking legitimacy and policy alignment, not a bailout. The verdict is about the narrative's internal logic, never about the underlying policy question, which remains entirely legitimate to contest (§6).
The mirror gets the same test
Before either verdict is published, the mirror frame, "government equity means state-controlled AI," gets the identical four steps: claim, predictions, plain reading, verdict. If no mirror existed, we would record and report that absence as a measured asymmetry. Symmetry of method, never false balance: lopsided measurements get published lopsided (§1).
The score, the watch, the write-up
Five anchored factors combine into a Charge Priority Score of 2.3, tier W3: reviewed weekly, thresholds set. The seven escalation markers all read clear, so ignition status is cold. What we publish is the card: both verdicts, the measured carrier set, and the standing statement that everything on it is protected political speech (§9, §7).
That was one narrative through the whole pipeline. The full methodology holds every rule we used and the reasoning behind each one.
Section §01
Standing principles
Every principle of our hate methodology applies here unchanged, and five more sit on top, all pushing in the same direction: this instrument touches protected mainstream speech, so its bars are higher.
The Counter-hate methodology answers one question with high precision: does this content collectivize, dehumanize, or incite against a group, past a documented bar? Most of the culture war never gets near that bar, and that instrument is built to say so and drop it. This methodology measures the part it leaves unmeasured: the camp-versus-camp narratives that stay protected speech from end to end. Its verification standard, silent dropping of non-hits, human-gated adjudication, and no-publish-without-review rules all carry over in full. Five additions are specific to this instrument:
Every published polarization finding states explicitly that the coded material is legitimate political expression being described, never flagged. The word "hit" does not exist on this side of the monitor; the unit is a narrative card.
There is no polarization roster, no watchlist, and no "polarizer" designation for any account. Claims attach to a narrative and its measured carrier set. Individuals are identified only at the public-figure tier the hate methodology already uses, and private accounts appear only in aggregate.
Before any verdict publishes, we look for the opposing camp's narrative about the same event and run the identical test on it. When no mirror exists, that absence gets recorded and reported as a measured asymmetry. This is symmetry of method, never false balance: lopsided measurements get published lopsided.
This methodology codes essentialization of camps, so its own prose never essentializes a camp. "The left claims" and "the right believes" are banned constructions; the subject is always the measured carrier set. A camp is not a monolith and our language must not make it one.
The logic test exists to test narratives, never to debunk one camp. When an imputed motive survives its prediction checks, the verdict is supported_by_checks and we publish it as plainly, and as prominently, as we publish contradicted_by_checks.
Section §02
Vocabulary discipline
The terms this instrument runs on, each anchored in the polarization literature, and the bars our own prose has to clear before using the loaded ones.
Dislike and distrust of the opposing camp as people, distinct from disagreement on issues1. Our Intensity factor measures the affective kind; issue disagreement alone never codes, because a healthy polity disagrees.
The othering, aversion, and moralization triad2: the theoretical frame for what the Intensity anchors are climbing.
Partisans overestimate the extremity of the other camp, by twenty points and more on issue after issue3. Load-bearing twice here: nutpicking manufactures perception gaps, and our own collection oversamples the loudest voices, so every prevalence estimate we publish is a ceiling (§8).
Content hostile to the out-group reliably outperforms content positive about the in-group: posts about political opponents earn roughly twice the shares4, and each added moral-emotional word raises diffusion by about a fifth5. Platform incentives push narratives up our Intensity scale without any coordination at all.
Camp members adopting a frame because trusted elites voiced it: the mechanism behind our Elite-adoption factor, and the reason convergence is expected here rather than suspicious (§11).
Presenting an opposing camp's fringe voice as representative of the whole camp. A coded operation at the gate (§5) and, when both camps do it to each other in a tightening loop, the accelerant our Lock factor measures.
Mobilization by hostility to the out-group rather than affinity for one's own; the engine sorted identities hand to conflict entrepreneurs6.
An actor whose reach or revenue depends on keeping a fault line hot7. Descriptive vocabulary only; it is not a roster category and never becomes one.
Our operational term for the mobilized, visible side of a divide: the accounts, outlets, and figures actually carrying a narrative. Never the electorate or demographic that side claims to speak for.
2.1Terms that have to clear a bar before we print them
The hate methodology's cautionary table ("both sides," "coordinated," "went viral," and the rest) applies unchanged. These are added on top:
| Term | Bar to meet |
|---|---|
| "Polarizing" | Requires a named narrative, at least one coded operation, and measured spread. Never a synonym for "controversial." |
| "Culture war" | Vernacular, fine inside quotes; in our own prose we identify the specific fault line instead. |
| "Bad faith" | Only as a paraphrase of a contradicted_by_checks verdict with the prediction checks cited; never free-standing. |
| "Propaganda" | Requires a coordination verdict per the hate methodology's evidence order, not just reach or repetition. |
| "The left / the right [verb]" | Banned in cards and reports; the claim attaches to the measured carrier set. |
| "Dangerous" / "pre-hate" | Never applied to a polarization narrative or its carriers. The escalation instrument (§7) reports named markers, not menace. |
Section §03
The Charge Model: a polarizing narrative's lifecycle
A polarizing narrative does not want violence. It wants to become the lens you see the next event through, and that ambition has a measurable lifecycle.
Our hate methodology's Ember Model treats ambient charge, its S0, as a precondition to be mapped before an event arrives. The Charge Model is the instrument that does that mapping. Three assumptions found it. The unit is the frame, the same as the Ember Model: an event spawns competing frames, and a polarizing frame is one whose compressed claim is about the opposing camp rather than about the event. Success is sedimentation, not violence: a hate narrative's endpoint is mobilized harm, while a polarizing narrative's endpoint is becoming a standing lens that pre-frames every future event. And the feedback loop is the object: each narrative that sediments raises the charge on its fault line, which lowers the activation energy for the next capture. Measuring one narrative in isolation misses the accumulation, and the accumulation is the thing we exist to watch.
Five stages and an exit ramp
click a stage for its signals, its measurable indicator, and what we do there
3.1The exit ramp: ignition
At any stage, when escalation markers (§7) fire past threshold, or any item inside the narrative actually clears the hate methodology's Track A or Track B, that narrative, or that branch of it, hands off to the Ember Model at the matching stage and is scored there. The polarization card records the handoff and keeps tracking whatever part of the narrative remains below the bar. The reverse path exists too: a decayed hate narrative's embers commonly survive as sediment on a fault line, and the inventory links to them.
3.2The fault-line inventory
A slow-moving registry, reviewed quarterly, listing the active fault lines: immigration, tech and AI, gender and family, religion in public life, policing, public health, culture and entertainment, and whatever the data actually shows. Each entry carries the sedimented standing lenses on both sides with their source cards, the reignition history, and a one-paragraph rationale. This is the polarization analog of a roster: a map of terrain, never a list of people.
Section §04
What gets collected
Collection starts from events and narratives, never from a list of people, because no such list exists on this side of the monitor.
A polarization pass is event-anchored and narrative-anchored: it starts from a trigger event or an active card, then samples the discourse around it across the same open-platform collection stack the hate methodology documents. Where account sampling is needed for a prevalence estimate, we use lane-tier accounts only, under all of the hate methodology's cross-lane comparison rules. Its confirmed tier, accounts admitted because of a past rubric-clearing act, is never a valid polarization sample: it was selected on conduct, and using it would poison both the prevalence estimate and the protected-speech framing at once. No new roster infrastructure exists for this methodology, by principle, not by omission.
Section §05
The polarization gate
Two prongs, both required, and a template that has to fill honestly. Heat alone never codes, and documented motive claims never code.
The hate methodology's gates run first on every item, always. An item that clears Track A or Track B belongs to that methodology and is never additionally coded as a polarization exemplar; it can enter a polarization card only as ignition evidence (§7). The polarization gate runs on what stayed protected:
- The object check. Is the item's primary object the opposing camp, its motives, character, or intentions, rather than the merits of the event or policy at hand? An item arguing that a policy is wrong, harmful, or stupid, at any temperature, fails this prong and stays uncoded. An item arguing that the policy reveals what its proponents really are or really want passes it.
- The operation check. Does the item perform at least one coded polarizing operation from the table below? Profanity, mockery, and vigor are not operations.
An evidence-backed motive claim fails the gate by design: reporting that an actor said or did a documented thing, with the source attached, is journalism, no matter how damaging. The gate catches imputation, not documentation.
5.1The eight polarizing operations
| Operation | Definition | Detection handle |
|---|---|---|
motive_attribution | Asserting the opposing camp's hidden motive without evidence | Imputed intent with no cited act or statement carrying it |
essentialization | "This is who they are": an act recoded as camp character | A trait predicate on the camp; the behavior-to-trait move applied to a political camp |
stakes_inflation | Ordinary politics framed as existential | End-of-country framing on routine policy contest |
guilt_by_association | A camp indicted through its worst adjacent actor, absent a documented tie | Association claim with no organizational or endorsement link |
nutpicking | A fringe voice presented as camp-representative | Prominence mismatch between the quoted voice and the "they" it stands for |
bad_faith_default | The opponent's stated reason presumed pretextual as a rule | Stated rationale dismissed without engagement or evidence |
symmetry_collapse | Whataboutism: unlike things equated to void a specific charge | "But they did X" substituted for answering the claim at hand |
purity_policing | In-camp enforcement: insufficient hostility treated as betrayal | Attacks on own-camp moderates for engaging the other side |
5.2The canonical-claim test
To code an item or found a card, the coder must state the narrative in one form:
<event> shows that <camp/actor> <is / wants / intends> <predicate>.If the template can only be filled by adding an imputation the text and context do not carry, the item is commentary, and it stays uncoded. The filled template becomes the card's canonical claim and the input to the logic test (§6). This is the same kind of precision tool as the hate methodology's articulability test, doing the same job one instrument over.
5.3From items to narratives
Clustering is a documented step, never an intuition. Items cluster into one card when they share a canonical claim about the same trigger event on the same fault line; wording differences do not split a cluster, and a genuinely different imputed predicate does. A card requires either two independent carriers or a single carrier at official or major-outlet tier; below that, an item stays logged and unclustered. Research lanes are collection beats, the places we look; fault lines are terrain, the divides narratives land on; a week's fault-line list is the inventory plus whatever the data itself surfaces. Both camps receive the same retrieval effort in every lane, a mirror is searched before any card records an absence, and cluster-boundary disputes go through the same human review gate as verdicts.
Section §06
The logic test and the plain reading
Every coded claim gets tested against its own internal logic, and the act it interprets gets our best ordinary explanation, voiced as our read.
This is the instrument that makes the polarization monitor more than a taxonomy. Four steps, recorded on every C2-or-later card:
- State the claim in canonical form, including the imputed motive.
- Check the predictions. If the imputed motive were true, what else would we expect to observe? List the concrete predictions and check each against what is actually observable, with sources. This is the load-bearing step, and predictions must be things the world can show or fail to show.
- Give the plain reading. The most ordinary available explanation of the act, from the actor's stated reasons and the observable shape of the act itself. Always voiced as our read: minds are not directly observable, and the write-up never pretends otherwise.
- Record the verdict with confidence. The names describe what we did, never a property of the narrative:
supported_by_checks(the motive survives the checks we ran),mixed_evidence(the checks land on both sides, or on one reading of the frame and not another),contradicted_by_checks(the predictions we drew from the claim are contradicted by what is observable), ornot_testable(the motive generates no checkable predictions from available evidence, recorded as such and never rounded to any of the other three).
The predictions are our own operationalizations of the frame, not the speaker's stated test, and reasonable readers may formulate different checks; every published verdict carries that disclosure. Before running any check, the coder records the exact source claim verbatim, the minimal interpretation of the frame, the strongest reasonable interpretation, the predictions that follow from each reading with a mark for which reading they test, the evidence that would support the claim as well as weaken it, and a confidence that the test faithfully represents the frame. A check that contradicts only the strongest reading leaves the minimal reading standing, and the verdict weighs both. Without this, an instrument like ours can drift into formulating hostile frames broadly and preferred frames narrowly, so the second-coder review in §14 attacks the test design itself, not just quotes and dates.
The walkthrough at the top of this page runs one invented example end to end: a 5% government-stake request, the loss-shifting frame it spawned, the two predictions that frame makes, and the contradicted verdict both of them earn.
Expect not_testable to be common. Many imputed motives generate no checkable predictions, and that is itself a finding: unfalsifiable claims are load-bearing in polarization. It does limit how often this instrument produces a satisfying verdict, and we say so in §15 rather than pretending otherwise.
Section §07
Escalation markers and ignition
Seven observable tripwires, each keyed to a construct the hate methodology already defines, and a categorical status that never inflates into menace.
Polarization below the boiling point is protected speech and stays protected speech. But hate narratives ignite out of loaded fault lines, and this checklist instruments that upstream region: the specific, observable markers that a polarizing narrative is moving toward the hate methodology's territory. Marker counts produce a categorical status. There is no invented continuous "distance to hate" number, because the markers are ordinal evidence, not a scale.
| # | Marker | Keys to |
|---|---|---|
| M1 | Pronoun collectivization: discourse shifts from named actors and policies to a generic "they / these people" | The Ember Model's collectivization tell |
| M2 | Motive-to-essence recode: an imputed motive on one act hardens into a standing trait claim about the camp | The recode ladder's behavior-to-trait move |
| M3 | Guilt spread: the frame starts attaching to people whose only link is camp membership | The proxy logic of Track A and B, inverted |
| M4 | First dehumanizing metaphor inside the carrier set: disease, vermin, or machine framing on people | Track B's metaphor check |
| M5 | Lexicon coinage: a derogatory coined camp label enters and spreads in the term set | The Ember Model's new-term-emergence tell |
| M6 | Eliminationist drift: end-state language shifts from winning the argument to removing, imprisoning, or destroying the opponent | Track B's eliminationist register |
| M7 | Mirror-lock acceleration: each camp's narrative cites the other's worst voices as its primary evidence, in a tightening loop | The Lock factor trending up, with nutpicking on both cards |
7.1Status
cold no markers observed. warming M1, M2, M3, or M7 observed, none of M4, M5, M6. near_ignition three or more markers, or any single M4 or M6 in mid-tier-or-wider carriers. ignited any item inside the narrative clears Track A or Track B.
Ignition status is orthogonal to the priority score and never feeds it, on the same two-number logic the hate methodology borrows from vulnerability scoring: how big a narrative is and how close it sits to the bar are different questions, and we keep them separately answerable. A near_ignition status escalates the card's review cadence, never its published framing; "dangerous" and "pre-hate" stay banned words.
7.2The handoff
On ignition, the clearing item is scored and reported entirely under the hate methodology. The polarization card records the handoff reference, keeps its history intact, and continues tracking whatever below-bar remainder exists. The hate-side card links back. One item is never reported as both a hate hit and a polarization exemplar: the linkage is the record, not double-counting. A near_ignition card that produces no clearing item within its review window decays back to warming with history retained, and that decay is calibration data for the forecast quality we track in §14.
Section §08
Reach, prevalence, and trajectory
Reach and trajectory come from the hate methodology unchanged. Prevalence is this instrument's addition, and every prevalence number we publish is a ceiling, said out loud.
Reach. The tiered lookup, the estimate schema, the labeling rules, and the "not measurable, say so" discipline all apply unchanged. No polarization-specific reach machinery exists.
Prevalence, penetration within the camp, is the new measurement. It inherits every constraint the hate methodology puts on cross-lane comparison: a penetration estimate requires a lane-tier sample over a stated common platform footprint, and the published figure names its residual biases every time, prominence-conditioned inclusion, snowball bias toward the network core, and the footprint bounds. One more bias is named on top, specific to this instrument: the outrage premium4 means visible discourse oversamples the hostile end of every camp, so penetration estimates are ceilings, never point estimates, and the write-up says so each time. An instrument that forgot this would not just be imprecise; it would actively widen the perception gap it draws on3.
Trajectory. The τ machinery applies to polarization cards unchanged: same bands, same confidence rules, same platform-footprint caveats, feeding the same post-hoc multiplier position in the score.
Section §09
Scoring: the Charge Priority Score
Five anchored factors, a geometric mean, and watch tiers lettered so that no shared dashboard can ever confuse this queue with the hate queue.
The structure deliberately matches the hate methodology's severity score: five factors anchored 1 to 5, combined by geometric mean, damped-compensatory for the same stated reasons and with the same honesty about what a geometric mean is and is not. A narrative must be big on multiple axes to rank high.
| Factor | Anchor at 1 | Anchor at 3 | Anchor at 5 |
|---|---|---|---|
| Reach (R) | Same anchors and tiered lookup as the hate methodology's Reach factor | ||
| Intensity (I) | Merits argument with camp framing at the edges | Motive attribution or bad-faith default as the running lens | Essentialization of the camp, existential stakes, purity policing. Dehumanization is past the top of this scale; that is marker M4, in §7 |
| Penetration (P) | Scattered voices, no pickup | A recognizable current, recurring across unrelated carriers | Camp orthodoxy; deviation policed; lane-tier sample majority, with §8's ceiling caveat |
| Elite adoption (E) | Fringe and anonymous accounts only | Partisan media and mid-tier influencers | Officials, candidates, or major outlets voice the frame as their own |
| Lock (L) | One-sided; the opposing camp ignores it | The opposing camp responds; occasional cross-citation | A locked pair: each camp's narrative runs on the other's as primary evidence |
The worked example's factors
the walkthrough narrative, scored
CPS=min( 5, (R·I·P·E·L)1/5×τ )
9.1Watch tiers
Logic test and mirror kept current; escalation markers checked every pass.
Card worked weekly.
Reviewed weekly, thresholds set.
Logged, not written up. Writing up a small narrative is amplifying it.
CPS is an operational priority score for our own review queue, never an empirical measurement of how polarized anything is, and published prose says so. Public surfaces show it to one decimal and treat smaller differences as noise; nearby scores are a tier, not a ranking. Factor values appear wherever a score appears, because Reach, Penetration, and Elite adoption travel together in practice (elite carriage lifts all three), and our §14 audits watch whether they collapse into one signal. Any composite across narratives on a shared surface is the median across all scored cards with the tier distribution, never a mean conditioned on the top tiers.
Ordinary partisan argument, satire, and heated merits debate should fail the gate entirely or score W4, and that is the instrument working, not failing. One floor: any card at near_ignition takes a W2 floor regardless of computed score, recorded on the card as a floor application, so the escalation watch can never starve because a narrative is small.
Section §10
The narrative card
One card per narrative, machine-checkable, with the logic test and the mirror recorded as first-class fields rather than prose afterthoughts.
Cards live in a schema of their own, with ids prefixed pol- so they can never collide with hate-side records. Item-level machinery, reach estimates and classifier provenance, is reused verbatim from the hate methodology's coding schema. The card's fields:
Narrative id, title, the fault line it sits on (linking the §3.2 inventory), and the trigger event with description, date, and source.
Origin camp and target camp, both identified by carrier set, never by demographic; the canonical claim as the filled template; the imputed motive; and the coded operations present.
Charge stage C0 through C4, and the sediment record: recurrence events and whether the frame has become a standing lens.
The exact source claim verbatim, the minimal and strongest reasonable readings of the frame, then every prediction checked, each marked for which reading it tests, with what was observed, its source, and whether it holds, fails, or lands mixed; the plain reading, voiced as our read; the verdict; confidence; and a stated confidence that the test faithfully represents the frame.
Status (locked, answered, unanswered, or none found), the mirror card's id when one exists, and the date checked. An unanswered mirror is data, not a gap.
Observed markers M1 through M7, each with evidence and date; the categorical status; and the handoff reference once any item clears a hate gate.
The five factors, τ, the computed CPS, the watch tier, and whether the W2 ignition floor was applied.
Top carriers at the nameable tier, the carrier pattern, and the penetration basis: the lane-tier sample and footprint behind the P score, or "judgment, no sample" stated plainly.
Linked narratives, the coordination verdict under the hate methodology's rules, classifier provenance, and the rationale paragraph every card owes its reader.
Section §11
Coordination
The hate methodology's coordination rules apply here unchanged and in full, and its central guardrail matters more on this side, not less.
The signal order, the shared-semantic-payload standard, the bridge requirement, and the overt-coordination anchor all carry over exactly. Above all: convergence is not coordination. Polarization narratives will trip the convergence pattern constantly, because that is what elite cueing looks like when it works. Thousands of accounts voicing the same frame within hours of a trusted elite voicing it is the expected mechanics of the discourse, not evidence of a campaign. A mirror pair locking is never, by itself, evidence that either side coordinated anything.
Section §12
Running both monitors: the interlock
Every research pass scores against both methodologies, both feed the weekly report and the dashboard, and each can produce a complete report without the other.
12.1The dual pass, per item
- The hate methodology's gates run first, always: Track A, Track B, the coded-language scan.
- Items that clear a hate gate are scored and reported under the hate methodology only. If such an item belongs to an active polarization narrative, it lands on that card as ignition evidence, never as a polarization exemplar.
- Items that stayed protected run the polarization gate. Items passing both prongs feed narrative cards; items failing stay uncoded, with the same aggregate screened-count discipline the hate side keeps.
12.2Separation rules
No CPS factor reads a hate-side score, and no hate-side factor reads a CPS. The only cross-reference in either direction is the ignition handoff record and the card linkage.
Card ids prefixed differently, watch tiers lettered W against the hate side's P, lifecycle stages lettered C against its S. No shared surface can conflate the queues even by accident.
In any joint artifact, polarization findings and hate findings live in separately labeled sections, and the polarization section carries the protected-speech statement. Appearing in the polarization section can never read as a hate accusation, structurally, not just by footnote. Any joint surface showing both scores states prominently that the two are scored on separate anchors and are not comparable across columns.
12.3What each report looks like
The joint weekly and the dashboard carry both queues in separate sections; the polarization section's natural shape is the fault-line map, the top narratives with their logic-test verdicts and mirror status, and the escalation watch reported as status changes only. A polarization-only report builds entirely from cards plus shared infrastructure, citing any ignition handoffs as dated events reported separately. And the hate-only report is exactly what this project publishes today, unchanged, with the option of citing a hate narrative's upstream polarization history as lifecycle context.
Section §13
Ethics
Everything in the hate methodology's ethics section applies, and the deltas all run one direction: this instrument touches protected mainstream speech, so the bars go up.
Individuals are identified only at the public-figure and institutional tier the hate methodology already uses, and only as carriers of a narrative, never with a "polarizer" label. Below that tier is aggregate-only, always, with no "unless notable" exception.
No polarization roster exists, no per-account polarization history is compiled, and the fault-line inventory contains terrain, not people. A capability to answer "which accounts are most polarizing" must not exist, parallel to the hate side's "watch the hate, not the hated."
Publishing logic-test verdicts on live partisan narratives is itself a move in the discourse being measured7. The mandatory mitigations: the mirror obligation, report-level verdict symmetry, equal prominence for supported_by_checks and contradicted_by_checks verdicts, and the symmetry disclosure from §14 on every report.
We identify and quote the minimum needed to make a finding checkable. A W4 narrative gets logged, not written up, because writing up a small narrative is amplifying it.
Section §14
Cadence, measures of our own accuracy, and the gold set
The instrument gets measured too: false positives first, symmetry disclosed every week, and a forecast score that tells us whether the escalation watch actually forecasts.
Cadence. Daily, the W1 cards get worked and escalation markers get checked on anything warming or above. Weekly, W2 cards get worked, W3 reviewed, W4 pruned, and the report's fault-line map refreshed. Quarterly, the fault-line inventory gets reviewed, sedimentation calls get audited against actual recurrence, and the score distribution gets the same drift audit the hate side runs.
14.1What we measure about ourselves
- False positives, first-class. A curated gold set of vigorous, heated, uncoded-by-design merits argument from both camps that the gate must not code. Recall gained by regressing here is rejected.
- Coding symmetry, disclosed. The distribution of coded narratives and verdicts by origin camp, published in every report. Symmetry is not assumed and not enforced; it is measured and shown.
- Verdict reproducibility, including test design. A sample of verdicts gets double-coded, and the second coder attacks the test itself: would they state the same claim, draw the same predictions, and mark the same readings? A verdict two coders cannot apply consistently is a collapse-or-clarify candidate.
- Score structure. A factor-removal check, whether dropping any single CPS factor materially reorders the queue, and cross-week factor correlations, watching whether Reach, Penetration, and Elite adoption collapse into one signal.
- Capture latency. Time from trigger event to first codable canonical claim, per card. Over time this is the empirical answer to how fast cooption happens.
- Ignition precision. Of cards that reached
near_ignition, the share that ignited within the review window, and of actual ignitions, the share that had prior warning status. This is the number that tells us whether §7's marker list forecasts or needs tuning.
The gold set's composition. Both camps represented; documented-motive journalism that must not code; heated merits argument that must not code; satire that must not code; and known past narratives that did ignite, whose pre-ignition items should code, with markers. Every constant this document introduces, the marker thresholds, the tier cutoffs, the W2 floor, is a convention: tunable, dated on change, and tracked in the same parameter-provenance register the hate methodology keeps.
Section §15
Open questions and limits
Version 1.0 states its own soft spots up front, because an instrument that hides its limits is advocacy with extra steps.
- Camp attribution has no ground truth. Origin camp is coded from the carrier set's observable alignment, and heterodox voices and single-issue coalitions will resist clean attribution. The reproducibility measure in §14 is the check, and we expect this field to be the one that collapses or clarifies first.
- The denominator problem. Penetration wants "share of the camp," and no unbiased sample of a camp exists on this collection stack. Lane-tier sampling with the biases stated is the honest available version, and it is a ceiling estimate, always.
- The perception-gap trap. An instrument that reports the loudest polarizing narratives can widen the very gap it draws on38, by making each camp's worst framings more visible to the other. The de-amplification defaults and the log-only W4 rule are the mitigation; whether they suffice is genuinely unresolved.
- Logic tests on motive claims have a ceiling.
untestablewill be common, because many imputed motives generate no checkable predictions. That is a finding about the narratives, and a limit on the instrument, at the same time. - US and English skew carries over from the collection stack unchanged, and fault lines outside that footprint will be invisible to us.
- The marker list is a first draft. M1 through M7 are keyed to constructs the hate methodology already validates, which is their strength, but no retrospective gold set of narratives-that-later-ignited has been assembled to test them against history yet. Building it is the first calibration task after adoption.
Appendix
References
Every externally sourced claim on this page carries a numbered marker that resolves here.
- Iyengar, S., Lelkes, Y., Levendusky, M., Malhotra, N. & Westwood, S. J. (2019). "The Origins and Consequences of Affective Polarization in the United States." Annual Review of Political Science 22, 129–146. annualreviews.org
- Finkel, E. J., Bail, C. A., Cikara, M., Ditto, P. H., Iyengar, S., Klar, S., et al. (2020). "Political sectarianism in America." Science 370(6516), 533–536. science.org/doi/10.1126/science.abe1715
- Yudkin, D., Hawkins, S. & Dixon, T. (2019). The Perception Gap: How False Impressions are Pulling Americans Apart. More in Common. moreincommon.com (PDF)
- Rathje, S., Van Bavel, J. J. & van der Linden, S. (2021). "Out-group animosity drives engagement on social media." PNAS 118(26), e2024292118. pnas.org/doi/10.1073/pnas.2024292118
- Brady, W. J., Wills, J. A., Jost, J. T., Tucker, J. A. & Van Bavel, J. J. (2017). "Emotion shapes the diffusion of moralized content in social networks." PNAS 114(28), 7313–7318. pnas.org/doi/10.1073/pnas.1618923114
- Mason, L. (2018). Uncivil Agreement: How Politics Became Our Identity. University of Chicago Press. press.uchicago.edu
- Ripley, A. (2021). High Conflict: Why We Get Trapped and How We Get Out. Simon & Schuster. simonandschuster.com
- Bail, C. (2021). Breaking the Social Media Prism: How to Make Our Platforms Less Polarizing. Princeton University Press. press.princeton.edu
The full working bibliography, including the primary-source verification notes behind each entry here, lives in the project's internal research files.
Hobocode Polarization Monitor. A methodology for identifying, testing, and watching polarizing narratives. Companion to the Counter-hate methodology.