Illinois Workers’ Comp Notice: Choose a 2-Event Hybrid Screen

An interval between extracted dates can be correct arithmetic and still be the wrong basis for a notice flag. That is an illustrative interval, not a verified Illinois notice deadline. The supplied Illinois General Assembly excerpts do not substantiate the headline’s asserted threshold or explain how to choose the events that start and stop the calculation. The evidence gap belongs on the intake screen, not behind an apparently definitive result.

From Natalie Fletcher’s legal-informatics perspective, the central problem is the promotion of an extracted date into a legal conclusion. Choose a two-event hybrid screen: capture the candidate injury event and candidate employer-notice event separately, retain their supporting text, and calculate an interval without labeling the claim timely or late. Software handles extraction and arithmetic; qualified legal review resolves ambiguous events and the governing notice rule.

That review must also verify the effective statutory text. The legislature warns that its database can lag recent enactments or display amendments before they take effect. A useful screen therefore distinguishes source-backed dates, unresolved event classifications, and an unverified legal threshold. Its promise is disciplined triage—not an automated notice ruling unsupported by the supplied law.

Build a 2-Event Notice Screen, Not an Automated

A notice screen should select evidence, not manufacture a verdict. For Chicago workers’ compensation intake, the critical design boundary is between a candidate injury-trigger event and a candidate employer-notification event. Store them separately, even when the same sentence describes both. Document creation, scanning, and upload timestamps belong to the record’s processing history; explicitly prohibit the extraction model from substituting them for either event date. A precisely timestamped upload does not cure an uncertain account of when something happened.

Each candidate event should carry its own identifier, event description, original date expression, normalized date, document identifier, and page or message location. Use a consistent format for resolved calendar dates, but preserve the source wording alongside the normalized value. Normalization changes representation, not evidentiary certainty. If a date expression cannot be resolved without assumptions, leave the resolved date unset or retain explicit bounds rather than silently supplying missing information.

Constrain the NLP component to extracting event descriptions, speakers, notification recipients, and supporting quotations. Distinguish the person recounting a notification from the person who allegedly delivered it, and preserve the recipient’s stated identity and role without inferring legal authority. Require every candidate to point to a passage that supports that event—not merely a nearby date. The model should return these structured candidates, not an elapsed interval or a prose assessment of timeliness.

For example, consider a hypothetical Chicago intake message saying, “I told my supervisor the next morning,” in a document uploaded later. The notification candidate should retain that quotation, its speaker, the described recipient, and its message location. “The next morning” remains unresolved unless the record supplies a supported anchor. If the anchor is identifiable, preserve the link to that separate passage as part of the derivation. Neither the upload date nor the message’s composition date automatically supplies the missing event date.

Deterministic code should consume the selected event dates, perform calendar arithmetic, and apply the configured screening threshold. Its output should identify the selected event identifiers, the calculation inputs, and the reason for an attorney-review flag. Missing anchors, conflicting accounts, or an apparent notification before the selected injury event should produce an uncertainty explanation rather than a coerced interval. An interval outside the screening window warrants review; it does not authorize automated rejection. An unflagged result likewise must not become an affirmative legal conclusion.

Keep extraction provenance separate from attorney interpretation. Preserve source passages and extraction versions as an immutable record; store event selections, reviewer annotations, and screening runs as linked records. When an attorney chooses a different trigger or notification candidate, create a new selection and calculation while retaining the original automated reasoning. Before deployment, test that replacement operation: the reviewer should be able to change the selected event, inspect the revised flag, and recover the exact quotations and inputs behind the earlier result. That makes the screen auditable without giving automation ownership of the legal conclusion.

Ground the Threshold in the Governing Statute, Not an Unverified

The governing notice provision, not a model’s confidence score, supplies the legal basis for ordinary accident-notice screening. The supplied excerpts of the Illinois Workers’ Compensation Act, 820 ILCS 305/, do not establish the employee-notice deadline. Verification must establish the applicable timing requirements before they are used for screening. For Chicago intake, a deterministic comparison therefore screens a verified statutory boundary; it cannot establish that notice satisfied every legal requirement. An apparently late notification warrants attorney review, not a software-generated conclusion that the claim is barred.

According to that same subsection, notice may be oral or written, and a defect or inaccuracy in notice does not bar proceedings unless the employer proves undue prejudice resulting from it. This matters when translating intake records into structured evidence: an absent written report is not equivalent to absent notice, and an imperfect account is not necessarily legally ineffective. The express protection concerns defective or inaccurate notice; it does not support a blanket assertion that complete failure to notify is automatically excused. Whether the evidence establishes notice, and whether any deficiency caused legally relevant prejudice, remain attorney questions.

Peoria County Belwood Nursing Home v. Industrial Commission supplies a particularly important constraint on injury-date extraction. According to the Illinois Supreme Court’s decision, a repetitive-trauma injury’s manifestation involves when both the injury and its causal relationship to employment would have become plainly apparent to a reasonable person. The inquiry is not simply “When did symptoms first appear?” For a Chicago worker describing gradually developing symptoms, a dated symptom entry may document discomfort without establishing awareness of its work relationship. Source-linked extraction should preserve that distinction rather than silently promoting the earliest symptom into the legally operative accident date.

The separate application-filing periods require verification against the governing statutory text; the supplied excerpts do not establish their duration. A filing-deadline field therefore cannot substitute for a notice-screening field: it represents a different procedural act, uses different periods, and can depend on compensation-payment history rather than employer notification. An application deadline still in the future does not resolve whether accident notice was timely. Keeping those fields semantically distinct prevents a correct filing calculation from supplying a false answer to the notice question.

The supplied Illinois General Assembly excerpts do not reproduce the notice provision and cannot independently substantiate its threshold. Before publication, verification must reach the governing statutory text, applicable effective dates, and subsequent treatment of Belwood. According to the Illinois General Assembly’s database notice, compilation is ongoing; recent laws may be missing, while amendments may appear before becoming effective. Check the relevant source notes against the underlying Public Acts, and check whether later decisions limit or clarify Belwood’s manifestation analysis.

That verification establishes legal support for the screening logic—not measured NLP accuracy or demonstrated reductions in missed notice issues. The defensible design preference is traceability: extracted dates retain their sources, and deterministic comparisons expose their legal assumptions. Potential lateness or material uncertainty triggers attorney review; neither a favorable comparison nor an unfavorable one authorizes software to declare notice legally sufficient or reject the claim.

Choose the Hybrid Screen

The hybrid screen wins as a governance choice, not yet as a demonstrated performance winner. For Chicago workers’ compensation intake, source-linked NLP extraction followed by deterministic screening and attorney review makes evidence capture, calculation, and judgment separately inspectable. That separation supports a more auditable workflow than an unsupported timely-or-late label; it does not itself prove fewer missed notice risks.

Evaluate the architecture on an attorney-adjudicated, de-identified notice-intake test set. Preserve source relationships and relevant date intervals during de-identification. Attorneys should establish which files require review without seeing the competing screens’ outputs, using a documented protocol for resolving disagreements. Compare manual intake and the hybrid against the same reference standard on the same files; keep evaluation files separate from rule development.

The primary safety outcome is the missed-routing proportion: reference-standard cases requiring review that the screen fails to route, divided by all reference-standard cases requiring review. Overall classification accuracy is not a substitute: correctly leaving straightforward files unflagged can obscure failures on the cases the screen exists to catch. The reference label concerns escalation, not whether notice is legally sufficient.

Measure review burden separately: flagged files that attorneys determine did not require escalation, divided by all flagged files. For example, a Chicago intake narrative with conflicting notification dates may warrant escalation even if an attorney ultimately finds notice sufficient. That favorable legal conclusion does not retroactively make the flag unnecessary. Judge escalation against the intake evidence available, not the eventual claim outcome.

Any effort to reduce unnecessary flags must report its effect on missed routing alongside its effect on burden. Otherwise, suppressing difficult cases can masquerade as efficiency. An apparent notification outside the applicable window must trigger attorney review, never automatic claim rejection or a declaration that the claim is barred.

Before claiming superiority over manual intake, require a comparison report disclosing test-set size, reference-standard review-case counts, disagreement-resolution method, and confidence intervals for each outcome and the between-approach differences. Explain how intervals account for evaluation on shared files. If uncertainty leaves the direction of the safety difference unresolved, retain the governance-winner label rather than asserting demonstrated outperformance.

What the Data Doesn’t Tell You

A clean intake dataset can be clean because the hardest notice cases never entered it. Retrospective records from opened Chicago workers’ compensation files may exclude people screened out during an initial call or preliminary document review. That selection process can remove precisely the apparent notice problems the proposed automation should detect. Performance on retained files therefore cannot establish performance across everyone seeking intake, even when every extracted date links correctly to its source.

The useful audit question is not merely “How many files were tested?” but “Who could never appear in this test?” Document the dataset’s entry point and whether pre-opening screening records are represented. If those records are unavailable, identify that missing population explicitly rather than treating opened files as representative. Otherwise, an apparent absence of missed notice risks could reflect prior human exclusion, not a reliable automated screen.

Aggregate extraction accuracy also needs to be separated by intake condition. For translated accounts, assess whether dates retain their relationship to the event described and whether uncertainty in the original account survives translation. For handwritten notes, examine omitted dates, ambiguous numerals, and incorrect links between dates and events. For low-quality scans, check unreadable passages and character substitutions separately. These conditions can overlap; a translated handwritten account should not disappear into whichever category makes reporting easiest.

Consider a hypothetical Chicago intake record containing a clear typed summary and a faint handwritten correction. Extracting the summary faithfully while missing the correction is not equivalent to extracting the complete account. Report condition-specific omissions and wrong-event associations, alongside the amount of material actually evaluated. Where examples are sparse, the defensible finding is limited evidence for that condition—not equivalence with clean, typed records. Deterministic screening cannot repair an input omission it never sees.

Attorney annotations introduce a different uncertainty: reference judgments can be contested. Distinguish disagreements about what a document says from disagreements about its legal significance. Preserve unresolved attorney disagreement rather than silently converting a majority label into ground truth. A model that differs from a contested legal annotation has not necessarily committed an extraction error, although that disagreement does not vindicate the model either. Likewise, an extracted notification outside the apparent window does not authorize software to label the claim barred. Potential lateness or material uncertainty belongs in attorney review, not automated rejection or a declaration of legal sufficiency.

A model confidence score should remain an unvalidated ranking signal unless calibration has been tested on comparable notice-intake records. Specify what the score purports to predict: correct transcription, correct event attribution, or agreement with an annotation are different targets. Calibration for one does not establish calibration for another. Fluent, high-confidence output cannot establish that an account contains every relevant communication or that its legal significance is settled.

For a small practice handling few notice-screening files, careful structured manual review may be the lower-burden choice. The comparison must include extraction upkeep, validation, exception handling, and monitoring—not just time saved on an individual file. Before committing to infrastructure, record those local burdens alongside the effort required for source-linked manual review and attorney escalation. Without measured local costs, no volume break-even point is defensible; automation’s maintenance burden is justified only where the practice can demonstrate a corresponding benefit.

Work the Intake Example

The later written record is not necessarily the decisive notice event. In this synthetic Chicago workers’ compensation intake file—not a reported case or an empirical research finding—an employee describes a discrete accident, a later conversation with a supervisor, and a subsequent written message. The screening task is to preserve both candidate communications and expose what remains unknown, rather than let the more explicit document displace the earlier conversation.

Keep each extracted date attached to its supporting source and event meaning. Here, the employee’s intake account supplies the accident date and the recollection of speaking with the supervisor; the written-message record supplies the later communication and its receipt date. Those references support different propositions. The employee remembers discussing pain during the conversation but cannot yet say whether that discussion connected the pain to the workplace accident. The later written message expressly describes the accident. Its clearer content does not establish what the employee did—or did not—tell the supervisor earlier.

For an illustrative calculation, document the counting convention and calculate separate intervals from the described accident to the conversation and to the written-message receipt. Both calculations should use the same accident date and counting convention; neither should substitute the message’s receipt date for the conversation date. Without supported dates, no numerical totals can be supplied here.

The statutory threshold must be verified before it supplies the screening benchmark. The supplied excerpts do not establish that threshold, so this example cannot classify either communication as falling before or after it. The calculation is illustrative: counsel must resolve any controlling legal time-computation question, along with the legal significance of either communication.

The unresolved variable is therefore content, not arithmetic. Selecting only the later written record would suppress potentially material evidence of an earlier communication; it would not resolve whether that communication conveyed the workplace connection. Conversely, selecting only the earlier date would hide the content uncertainty. Neither a late-looking document nor an unsupported NLP “timely” label warrants rejecting the claim or declaring notice legally sufficient.

Retain separate review references: earlier supervisor conversation—employee intake account, workplace connection unresolved; later written-message receipt—written-message record, accident expressly described. Ask counsel to review both and identify what further clarification is needed about the conversation. Worked output: attorney review required—earlier communication content unresolved.

How to Choose Well

A notice screen's output vocabulary is a governance decision, not a UI detail. A Chicago workers' compensation intake pipeline that emits "timely" or "late" has already rendered a legal conclusion it cannot support. One that emits "priority review," "recipient verification," "framework selection," "notice evidence missing," or "no automated notice flag" has instead sorted files by the next human action required. The second vocabulary is the one that survives contact with the Illinois Workers’ Compensation Act, 820 ILCS 305/, and with the attorney who signs off.

The reason is structural. Source-linked extraction yields dated events; deterministic screening yields intervals between candidate pairings. An interval is a fact about the record. Whether that interval satisfies the statute — and whether an exception, a substitute recipient, or a different trigger event applies — is a fact about the law applied to facts the intake record usually lacks. Collapsing the two produces a confident label with no defensible basis, which is the exact failure mode the hybrid design exists to prevent.

Route to the priority attorney-review queue when every plausible event pairing yields an interval beyond the applicable notice limit. The quantifier "every" is doing the work: if even one defensible pairing falls inside the limit, the flag weakens and the file belongs in ordinary review. What you must not do is let the flag become an automated refusal of representation. A screening result is a routing instruction, never a disposition.

Route for recipient verification whenever the purported recipient's employer affiliation or authority is unknown — even when the calculated interval looks comfortably within the limit. Interval arithmetic assumes you know who received the notice and in what capacity. Notice logged to a staffing-agency coordinator, a parent entity's safety officer, or a coworker with no supervisory role may or may not satisfy the statute's recipient requirement. A clean interval built on an unidentified recipient is a clean interval about the wrong event.

Stop the ordinary accident-notice screen when the account suggests occupational disease rather than a discrete accident. Repetitive-trauma and exposure claims have no single trigger event for the pairing logic to anchor on, so the screen will either find nothing or find an arbitrary date. Route the matter for selection of the governing legal framework instead; that choice determines which notice provision applies at all.

Assign "notice evidence missing" when the source materials contain no identifiable employer-notification event, and route for attorney-directed clarification. The discipline that matters: absence from the intake record is not evidence that notification never occurred. Intake records are built by whoever happened to document the call, and the hardest notice cases arrive with the thinnest documentation.

Retain "no automated notice flag" when the screening inputs are supported and no configured risk condition triggers, then continue ordinary attorney intake review. Never promote that result into a legal assurance that notice requirements were satisfied. "No flag" describes the screen's behavior on the inputs it had — not the claim.