SOURCELOCK R10 — EARNINGS-COMPARISON ADMISSION RULE
PROSPECTIVE DEPLOYMENT RECORD v1 (Science Trust addendum; registered 2026-10-14)

Parent registration: SCIENCE TRUST PROTOCOL v1 (umbrella governance, protocol_id be8b34040c2e).
Class: methodological admission change, prospective only. No statistical endpoint; no run is ever
recorded against this entry. Research and education only. Not investment advice.

1. WHAT CHANGES
SourceLock Guard is the deterministic validator applied to every model-extracted filing event before
the event can be accepted. From the deployment date (section 5) it enforces rule R10
(earnings_comparison_support): an extraction classified EARNINGS_BEAT or EARNINGS_MISS is admitted
only when the cited excerpt itself states the comparison in the classified direction against an
expectations benchmark (expectations, consensus, estimates, forecast) within one sentence, with no
negation and no opposite-direction term in that sentence. When the extraction also carries a
structured comparison object, that object must be complete (actual, expected, metric, period,
currency, basis, both text spans, benchmark source, benchmark publication time earlier than the
result publication time, both timezone-aware), its values must be finite, and it must not contradict
the classified direction. Rule R5 additionally refuses a non-finite confidence or a confidence outside
[0.5, 1.0]; rules R3, R4 and R5 refuse boolean values. The neutral event type EARNINGS_RESULT is added
to the vocabulary for an earnings result stated without an expectations comparison.

2. WHAT THE RULE IS NOT
The check is lexical and structural. It is not factual entailment, not an independent verification of
analyst expectations, not a check of the benchmark's authenticity and not source authentication. An
admitted event is at most "textual comparison stated" or "structure consistent"; neither is a truth
claim about the comparison.

3. SCOPE IN TIME AND WHAT IS PRESERVED
The rule applies to extractions admitted on or after the deployment date. No previously accepted event
is reclassified, relabelled or removed; no previously reported estimate, table, ledger root, registry
chain line, packet or public statistic is modified. Previously accepted EARNINGS_BEAT or EARNINGS_MISS
events whose excerpt does not state the comparison are flagged descriptively on the current Why
surfaces (excerpt_comparison_checks: UNSUPPORTED_COMPARISON) and retained as historical records.

4. EXPECTED EFFECT ON REGISTERED COUNTS
Fewer EARNINGS_BEAT and EARNINGS_MISS admissions after the deployment date, and a possible shift of
such extractions to EARNINGS_RESULT or to no event. Per-type accepted-event counts that feed Evidence
Coverage v1, Evidence Lifecycle v1, the completeness profile and the anytime observer are affected
prospectively from the deployment date. Any registered read that pools accepted events of these two
types across the deployment date must disclose this rule change and its date.

5. DEPLOYMENT
The rule is deployed in the production extraction path when the release carrying this record is
published and the production checkout moves to that release; the deployment date is that release's
push-to-main timestamp in the release journal, not this registration date. Until then the production
extractor runs the prior rule set (R1 to R9). This registration precedes any new evaluation window
that could be affected; it does not open, close or re-date any enrolled window.

Method hash = SHA-256 of this file as registered in the protocol chain.