HugoScore review worksheet, draft v1.3

Use with SCORING_RUBRIC_DRAFT.md. Provisional method, pending publication. Do not overwrite published scores with this worksheet before the publication order in audits/scoring-method-decisions-2026-10-18.md reaches that step. v1.2 preserved at research/rubric-v1-3-draft/worksheet-v1.2.md (project record, not published).

Review object

• Product/operator:
• Function and exact deployment:
• Primary user and affected patient population:
• Access tier, location and version:
• Evidence cutoff and retrieval dates:
• Who selects, pays, sets the task and receives outputs/insights:
• Scope gates, each with evidence: AI / health / patient-impact pathway:
• Named model-mediated operation in the assessed function. Downstream AI use or agent compatibility alone is insufficient:
• Standalone tool or defined combined AI workflow, with its boundary:
• If multiple materially different deployments exist, create separate worksheets:
• Where the tool sits between the institutional and patient-directed ends (rubric section 3, decision H1), and which rails put it there. Who deploys it is evidence, not the answer:
• Direct participation (rubric section 2 (a)): Y / N / U. The patient's own real-time input or response is part of the AI interaction. Who operates the interface, whose information is processed, whose purpose and consent govern the task. Authorized assistance counts as participation:
• Patient-set purpose (rubric section 2 (b)): Y / N / U, judged by function, not format. N means control axis not applicable, posture and measures only. U at documentary coverage means "control axis scope unresolved pending sessions":
• Evidence coverage: documentary / partial interaction / protocol-complete interaction, plus independent operational evidence if any:
• Review status: AI draft / human adjudication (questions settled, by whom, date) / substantive human review:

Frozen review object (rubric section 2A)

• Named function, deployment and access tier:
• Role record, each with its authoritative source. Unknown roles marked unknown:
• Service operator (makes the AI function available to the patient and is accountable to the patient for it; list joint operators with roles), with one outside source or "operator unverified externally":
• Institutional deployer, if any:
• Application supplier:
• Model developer(s), or routing statement:
• Inference host:
• Other material services:
• Consequential outputs and mandatory downstream processing inside the boundary:
• Functions separated, and why their goals, permissions or consequences differ materially:
• Composed workflow: configuration documented by each party (documentary) or actually tested (interaction), with host product, account type, plan, mode, install route, model setting and connector-supplied instructions. "Unresolved configuration" if not selected or not knowable:
• Supplier deployed by several institutions: reviewed deployment named, other known deployments listed as not reviewed, sibling functions named:
• Retrieval dates and content hashes of the documents the boundary rests on, if no version is published:

Transparency gate (rubric section 3A)

States: supported / not found after defined search / inaccessible (date, method, response) / inaccurate / unresolved / not researched / not tested / not applicable (T3 only). "Publicly" means reachable without an account. Run the gate for off-axis tools too.

• Gate result: eligible / ineligible (flag: operator not disclosed, model provider not disclosed, operator disclosure inaccurate, disclosure unresolved) / suspended (T1 or T2 not researched or inaccessible) / documentary (T3 not tested, T1 and T2 supported):
• If ineligible, conditions-implied set for calibration only, never displayed:
• Prose control classification, written regardless of gate result:

Evidence ledger

Keep publicly stated commitments separate from observed implementation. For a public-source draft, record which findings remain documentary. Describe the search boundary before writing not disclosed.

Primary finding 1: who the tool answers to

• Patient-chosen purpose and any operator/institutional purpose:
• Whose purpose prevails when they conflict, with evidence:

• Control classification: patient-directed / externally directed / shared authority / fixed-purpose function (off-axis only) / unresolved:
• Purpose authority P: external / shared / patient / U, with conflict evidence:
• If externally directed, evidence for binding remedy, meaningful participation/refusal/challenge, or demonstrated absence of these:
• Evidence satisfying the selected definition, with unresolved alternatives:
• Relevant hosting/data/portability condition and its concrete effect, if any:
• Full workflow authority R: Y / N / U. Identify a retained external power for N. For Y, evidence continuation, weights/instructions or replacement, data/output control, applicable permissions and absence of an operational veto:
• Anchor rule applied from section 5A, with every deciding evidence ID:
• Draft anchor or set of possible anchors. Do not interpolate or penalize unknowns:
• Unknowns that would resolve the candidate set, with the next evidence needed:
• This is an ordinal control position paired with agency posture, not a combined quality score:

Primary finding 2: agency effect

• Patient-chosen purpose and whose priorities define benefit in this use:
• Concrete capabilities available to the patient:
• Understanding, alternatives, action, revision and authorship, each with evidence:
• Patient-chosen goal versus an imposed goal:
• Alternatives supported, excluded or redirected:
• Immediate demonstrated effect versus claimed sustained capacity:
• Whether evidence concerns capability, simulated use or actual patient use and outcomes:
• Institutional efficiency, engagement or adherence that cannot alone establish patient benefit:

For comparable conversational evidence, use the common test packet (project record, not published). Record an inapplicable function rather than forcing a conversational test onto an unrelated AI workflow.

Interaction evidence

• Untested tasks and why:
• Inconsistency and fresh-session follow-up, retaining all results:
• Scope, safety, technical failure or imposed direction, with the basis for distinguishing them:

Agency effects and public measures

• Critical reflection: what can the patient examine or question?
• Strategic action: what can they do next?
• Sustained capacity: demonstrated, hypothesized or not assessed?
• Supported benefit B, with concrete mechanism and its comparator (the patient's realistic alternative in this use):
• Supported material constraint C, with concrete mechanism:
• Recorded constraint search: for each group (refuse or leave, correct or challenge, steered or deceived), surfaces checked. Were the binding terms and data-use disclosure read? Was the layer where a constraint would be set reachable? Any candidate constraint of unresolved scope:
• Documented optimization target (test priority, not a finding):
• Documented handoff or escalation design (test priority, not a finding):
• Testimony substitution (rubric section 7 steps 1 to 6): present / absent in reviewed material / unclear / not tested / not applicable. Passage, label shown, follow-up answer, source verification, repair type, blocked ability:
• Deferral displacement (rubric section 7, per task P2, P3, P4): present / absent in reviewed material / unclear / not tested / not applicable. For each task: goal and deliverable, in-scope reason, passage offered instead, absent help, stated reason and its support, whether referral preserves choice. Recurrence across both runs:
• B and C apply to the same function, population and conditions of use: Y / N. Each marked demonstrated or documented potential:
• Unknowns that prevent a conclusion:
• One agency posture, using the ordered rules:
• Why a different posture was not selected:

Confidence, provenance and review

• Confidence for deciding claims:
• Confidence for control classification and agency posture separately:
• Review status (rubric section 8): AI draft / human adjudication / substantive human review:
• Reviewer and exact AI model when known:
• Named human reviewer, review scope, decision and date, if any:
• Limitations and untested controls:
• Rubric version (v1.3 or later):
• Coder models and vendors, with the conflict check against every provider in the role record, unknown relationships marked:
• Strongest contrary evidence named by each coder:
• Hugo's boundary and gate confirmation (date), and adjudication of any disagreement:
• Expands agency or Constrains agency publication authorization, if applicable (Hugo Campos, date). Without it, Constrains publishes as Unclear with the constraint stated in full:

Refresh and disagreement record

• Complete the new assessment before looking at the legacy number:
• Prior evidence cutoff and method version:
• Control and posture conditions changed by new evidence:
• Reasons for every change or retained conclusion:
• Peer comparison after rule application, with any discrepancy recorded rather than normalized away:
• Independent reviewer differences and evidence-based resolution:
• Publication authorization, recorded separately from substantive review: