The Someris platform
Software in this market is sold in adjectives. Someris publishes measurements: every pillar below ships with numbers taken against coordinator-signed assessments and written up in an internal report, including the results that went against us. The clinician stays at the center of every one of them; the platform multiplies what one coordinator can carry, with judgment and signature staying human.
[01] Extract
Someris reads every page of the chart: EHR feeds, discharge summaries, therapy evals, and the scanned or faxed packets no text pipeline can see. Deterministic form readers run first at zero model cost; vision models read what the parsers cannot; and every extracted item carries its source document, page, excerpt, and date.
The receipts
HIPPS billing-code position accuracy against coordinator-signed assessments: the 39 signed billing codes in a 47-patient audited corpus.
Accuracy scorecard, Aug 2026
marginal cost of the deterministic floor, which matched or beat the LLM pipeline on the Nursing and NTA payment components in a 17-patient A/B.
Deterministic-floor A/B report, Jul 2026
queries routed to the right page (top 3) by vision retrieval inside a 42-page image-only scanned packet, where text search found 0 of 12.
Vision-retrieval spike, Aug 2026
Scribes transcribe the visit that is happening. Record systems store the chart that already happened. Someris reads the existing chart into the payment assessment itself.
[02] Verify
A draft you cannot score is a draft you cannot trust. Someris scores every drafted item through a verification chain: was the evidence found, is the value consistent with it, is the code valid. Items that fail are flagged for human judgment with measured precision, and an independent second pass escalates the genuinely ambiguous ones. One layer that measured below its bar stays switched off, and we publish that result too.
The receipts
precision of the wrongness flags at the operating threshold: 2.6 times the base error rate, measured over 10,764 labeled items on attested gold assessments.
Shadow-calibration report, Aug 2026
items flagged per case, so clinician review lands where the risk actually is instead of everywhere.
Shadow-calibration report, Aug 2026
truly wrong selected items correctly escalated to human review in the live second-pass probe.
Shadow-calibration report, Aug 2026
"Mostly autonomous" is an adjective. A published precision number on a published corpus is a measurement. We sell the measurement.
[03] Optimize
Each coding decision is priced in per-diem dollars before attestation. Someris surfaces only revenue the documentation can defend: over-code guards, template-text guards, and computed-score integrity checks run against our own output first. When we disagree with a signed assessment, we argue both sides the way an auditor would, record a verdict with a confidence score, and keep score in public.
The receipts
per day of defensible upside upheld against a signed assessment: a swallowing complaint two clinicians documented in-window that the assessment missed. Verdict confidence 0.82.
Adjudication report, Jul 2026
the adjudication score on the last disputed cohort. The coordinator was right six times, Someris once, two unresolved. We publish our losses.
Adjudication report, Jul 2026
bed-days of payment governed by a single completed assessment. That is why one defensible correction compounds.
Pricing model, Aug 2026
Revenue software that cannot show its evidence is an audit finding waiting to happen. Every dollar Someris suggests ships with the chain that survives the audit.
[04] Submit
The assessment goes out the door already checked: CMS item-consistency edit rules cleared, the HIPPS code recomputed against the CMS grouper logic itself, and the submission XML in the exact format the live CMS validator has already accepted. Submission snapshots are immutable, so what was filed is always answerable.
The receipts
response from the live CMS validation utility for our generated iQIES XML, with zero structural findings.
iQIES verification, Jul 2026
CMS item-consistency edit rules enforced pre-submission, with a self-check that verifies the rule set actually loaded.
CMS edit engine, in production
the CMS grouper version our payment logic is audited against, with every fix cited to the CMS source.
FY2026 engine audit, Jul 2026
Everyone claims CMS compliance. We kept the receipt from the CMS validator.
[05] Defend
When the ADR letter, the surveyor, or the ransomware arrives, the response is already built. Every item carries its pre-assembled evidence packet. Every access to patient data is logged and retained. Backups live in an isolated vault under a locked retention policy that even our own administrators cannot override, and the restore path is drilled with a stopwatch, not a policy document.
The receipts
measured recovery time for a full production database restore, with row counts verified before teardown.
Restore drill, Aug 2026
the HTTP status our own org administrator receives when attempting to delete vaulted backups. The retention lock cannot be removed, by design.
Tabletop exercise, Aug 2026
audit-log retention, with every access to patient data written to a dedicated access log.
HIPAA safeguards, in production
Compliance as a marketing layer versus compliance you can drill. Our disaster answer is a stopwatch reading.
[06] Sovereign
Every client owns the model tuned on their facility and can leave with it: weights, training manifest, evaluation harness, all signed and hash-verified. A tuned model is promoted only if it beats or ties the previous one on that client's own audited assessments, with zero payment-item regression. Before any model ships, we attack it ourselves with a membership-inference audit, and a failing privacy score blocks promotion unconditionally.
The receipts
residual patient-name leaks across 3,053 documents in the latest de-identification audit, with canary scans at every storage boundary.
Sovereign model architecture, Aug 2026
checks passed by the latest model promotion gate on client gold data. A privacy-attack score above threshold refuses promotion, no override.
Training run report, Aug 2026
every promoted model ships with a cryptographic signature and a manifest answering: trained on what, under what authority.
Sovereign model architecture, Aug 2026
The anti-lock-in that locks in: clients stay because leaving is easy and staying compounds.
About these numbers: every figure on this page comes from a written internal engineering report, measured against coordinator-signed assessments or verified live against the named external system, on the dates shown. Nothing here is projected, modeled, or presented as "up to". Corpus sizes are stated with each figure. Methodology and reports are available to customers and partners under NDA. Someris provides clinical decision support; a licensed clinician reviews and attests every assessment.
Bring one real case to the demo and watch every pillar run on it: extraction, verification, priced decisions, and the evidence packet, live.
