How Fulcrum verifies and archives claims.

At launch, Fulcrum covers the war in Ukraine. The platform is designed to be region-agnostic and will extend to other active conflicts as moderator capacity allows. Scope expansions will be announced here, not silently rolled out — each new region brings its own source landscape, legal context, and language requirements that must be addressed deliberately.

Fulcrum does not aggregate news about peaceful events, political commentary, or domestic politics outside an active conflict zone. It is a record of what is claimed about events on the ground and how well those claims hold up to evidence.

Every entry on Fulcrum is a claim: a precise, falsifiable proposition about something that happened — a strike, a movement, a casualty count, an official statement. A claim without a precise proposition is not a claim; it is a topic, and we will not score it.

Each claim carries a status that reflects where it stands against the evidence:

unverified

The claim exists, but the evidence on either side is insufficient or contradictory. This is where most claims begin.

partial

Multiple credible sources or pieces of evidence support the claim, but key details remain unconfirmed.

confirmed

Independent, credible evidence supports the claim and no significant contradicting evidence remains outstanding.

debunked

Credible evidence contradicts the claim. The claim and its history remain visible — debunked claims are part of the record, not deleted.

superseded

A later claim or correction replaces this one. The original stays preserved with a link forward to its replacement.

Status changes are not opinions — they are events recorded in an append-only audit log on each claim, attributed to the moderator who made the change and the reason given. See §8 for how to review or dispute a change.

Each claim has a confidence score between 0 and 100. The score is computed from a transparent rule, not a private model:

score = clamp(
  50 + sum(evidence.weight × source.reliability) × 10,
  0,
  100
)

Every claim starts at 50 — neutral, no judgment. Each piece of evidence carries a weight between -1.0 (strongly contradicting) and +1.0 (strongly supporting), multiplied by the curated reliability rating of the source it came from. The sum moves the score up or down, clamped to the 0–100 range.

The score maps to a status bucket as follows:

Every claim page shows the breakdown that produced its score — which pieces of evidence contributed which amount, against which source reliability. If you disagree with the number, you can see exactly where it came from and which assumption you want to challenge. Moderators may override the computed score with a written reason; the override is logged with the same visibility.

Every piece of evidence attached to a claim is captured and preserved on submission. We do not rely on the source URL remaining live — links rot, accounts are deleted, posts get edited.

For each evidence item Fulcrum stores:

• The original asset (image, video, full-page screenshot of a text source) on Cloudflare R2.
• A SHA-256 hash of the original as received. Anyone can compute the same hash on the stored asset to confirm we have not modified what we claim to have captured.
• A second SHA-256 hash of the optimized version we actually serve. Both are visible on the evidence page so the chain from source to served file is auditable end-to-end.
• A fallback snapshot in the Internet Archive Wayback Machine, triggered automatically on capture.

Evidence is never silently removed. When a claim is debunked or superseded, the metadata and links remain in the record; only heavy binaries may eventually be dropped from our own storage after a retention window, with the Wayback fallback continuing to serve as the citable record.

Every source Fulcrum tracks — RSS feeds, news APIs, OSINT analysts, Telegram channels, official statements — is listed in the public source directory with its reliability rating and its political-lean annotation.

Reliability rating is curated by Fulcrum moderators and revised based on observed accuracy over time, not voted on by users. A source whose claims repeatedly fail to hold up loses reliability. The rating is a number, not a label, and every rating is accompanied by a moderator note explaining the basis for it. A full rating-change history is a planned addition.

Political lean is not determined by Fulcrum. We do not classify outlets as left or right ourselves — an in-house algorithm doing that would be an inevitable credibility disaster. Instead, we tag sources using the public ratings from Ad Fontes Media's Media Bias Chart as of May 2026, and cite the date so the snapshot is verifiable.

If a source you trust is missing, or one we list seems misrated, email the moderators at moderators@fulcrum.watch.

AI assistance is not enabled in the current release. All claims and evidence on Fulcrum today are processed by human moderators. The policy below describes the framework under which AI assistance will be introduced, published in advance so the rules are auditable before they apply.

The governing principle is "suggest, don't decide." AI never publishes anything users see. Every AI output that affects what appears on Fulcrum passes through a human moderator who approves, rejects, or edits it. AI handles the tedious work of triage and extraction; humans handle every judgment call that affects the record.

What AI will be used for:

• Pre-filtering incoming source material to identify what is plausibly relevant to a tracked conflict.
• Extracting structured entities (locations, dates, named units, equipment types) from free-text claims for moderator review.
• Suggesting that a new claim may duplicate an existing one, or that new evidence may change an existing claim's status.
• Extracting a neutral English summary from non-English source material, with the original always preserved verbatim alongside the summary and labeled "AI-summarized."
• Drafting event summaries that aggregate constituent claims, always human-edited before publication.

What AI will never be used for:

• Computing or overriding confidence scores. The scoring rule is public and rule-based for a reason.
• Marking claims confirmed, debunked, or superseded without a human moderator's approval.
• Classifying the political lean of a source — see §5.
• Fact-checking claims against "what the model knows." Large language models hallucinate, especially on breaking events outside their training data, which is exactly the material Fulcrum exists to handle. We do not let them do it.
• Generating synthetic evidence — fabricated quotes, descriptions of imagery the model has not actually analyzed, or invented source attributions.

Any field on Fulcrum that an AI step contributed to — even after a moderator has approved it — carries a small AI-assisted indicator with a tooltip explaining which step the AI performed and which moderator approved it. The audit log records the same information permanently.

Documenting conflict requires preserving footage that is distressing to view. Fulcrum's display rules try to balance historical record-keeping against viewer wellbeing.

Every piece of visual evidence carries a severity tier:

none

No imagery that requires a warning.

mild

Damage to property or terrain; no people visible.

graphic

Visible injury, distress, or aftermath including persons.

severe

Death, dismemberment, or material a reasonable viewer would consider deeply distressing.

The display rules are:

• Video and audio never autoplay. A click is always required.
• Visual thumbnails for mild and above are blurred by default. A click reveals the content; the warning copy describes the severity tier so the viewer knows what they are revealing before they reveal it.
• When a moderator flags evidence as graphic or severe, they may write a short description of what the viewer will see. Researchers can decide whether they need to view the imagery directly or whether the written description is enough for their purposes.
• Logged-in users can raise or lower their personal blur threshold. Anonymous visitors get the safest defaults.

At launch, Fulcrum is staffed by a small core of named moderators. Their names and OSINT backgrounds are listed publicly so the people making judgment calls on the record are themselves on the record. Additional moderators are recruited from the OSINT community as the project earns trust; no one moderates anonymously.

Every change is logged. Each claim page exposes an append-only audit log: which field changed, the previous and new value, the moderator who made the change, the timestamp, and the reason given. This includes status changes, score overrides, evidence weight changes, and severity flags. Nothing on Fulcrum is silently revised.

If you believe a claim is wrong, a piece of evidence is mis-weighted, or a severity flag is incorrect, you can submit a correction. Submissions go to the same moderator queue as new evidence. Approved corrections create a visible audit log entry; rejected corrections are still recorded so the disagreement is preserved.

Some commitments are easiest to state by negation. Fulcrum deliberately does not:

• Predict outcomes or accept wagers. Forecasting is a different discipline that rewards different incentives.
• Score people, only claims. Reputation systems for contributors encourage speed over accuracy in exactly the wrong way for this domain.
• Auto-publish anything AI generated. Every AI-touched field has been through a human moderator. See §6.
• Delete the historical record. Debunked claims, superseded claims, and rejected corrections all remain visible. The record is a record.
• Take sides on contested political framings. We cite ratings from third parties for source bias; we do not invent our own.