Build a grant scoring rubric that scores 500+ applications consistently. Anchored criteria, weighted dimensions, citation-level AI evidence.
A scoring rubric is the closest thing to objectivity a funding process has. You define criteria, weight dimensions, describe what a strong application looks like at each level. Then twelve reviewers walk into the same cycle and produce twelve slightly different interpretations of what "demonstrates clear community need" actually requires as evidence. This is not a training failure. It is a language problem.
Rubric criteria written in evaluative language leave interpretation to the reader. "Strong," "compelling," "demonstrates understanding" all describe how the reviewer feels rather than what is present in the application. The result is a scoring process where drift is structural, calibration is temporary, and the shortlist reflects who reviewed what, when, and in what order as much as it reflects the merits of the applications.
Most grant review rubrics sit closer to evaluative language than to observable evidence. The fix is not more training. The fix is rewriting the rubric so the same evidence produces the same score regardless of who is reading.
Reviewer 12 at essay 80 is not the same evaluator as Reviewer 1 at essay 1. Late-session scoring runs 15 to 20 percent more lenient on average. Quality of reasoning drops measurably after 8 to 10 proposals in a single session.
Cognitive load research, multiple studies 2018-2024.
No two humans interpret "demonstrates community engagement" the same way. Criteria that seem clear in training diverge in practice, often by the second day of a multi-day review. The training session anchors interpretation. The next session resets it.
Brown University Sheridan Center research on inter-rater reliability.
Polished writing from well-resourced applicants scores higher regardless of substance. Reviewers unconsciously reward familiarity of tone and structure. Applicants whose first language is not English, or whose programs serve communities that produce different rhetorical patterns, score lower for reasons unrelated to program quality.
Documented in foundation pilot data; published systemic reviews vary.
The first three applications a reviewer reads set their mental benchmark. All subsequent scoring is anchored relative to that opening sample rather than to the rubric. A weak cohort scored first inflates the entire pool. A strong cohort scored first compresses the middle.
Anchoring effect, Tversky and Kahneman 1974; replicated across review contexts.
Score variance per criterion across 12 reviewers reading the same proposal. Variance widens late in a session.
Score variance per criterion when a single AI applies one anchored rubric across the entire applicant pool.
The sections below cover how to build rubric criteria that produce consistent scoring at scale, what holistic and analytic rubrics actually mean for grant review, which scale granularity fits which program type, how the rubric becomes an AI instruction set, and how to write evaluator instructions that survive contact with the actual review process.
Sopact Sense translates your existing rubric into AI-ready analysis prompts and scores your full applicant pool against every criterion. Update a weight or add a criterion mid-cycle and the pool re-scores automatically.
The most consequential decision in a grant scoring rubric is the language of the anchors. Anchors are the descriptions of what each scoring level looks like. Unanchored rubrics use evaluative adjectives that each reviewer interprets privately. Anchored rubrics describe observable evidence: what must be present in the application narrative, the budget, or the attachments. The same evidence produces the same score from any reviewer or AI applying an anchored rubric.
Consider two versions of the same criterion. The unanchored version asks the reviewer to judge. The anchored version asks the reviewer to confirm what is in the application.
Applicant demonstrates a strong understanding of community need and presents a compelling case for funding.
What happens in review: each reviewer defines "strong" and "compelling" privately. Twelve reviewers produce twelve interpretations. Forty-seven applications score 3.8 because reviewers cluster toward the middle when criteria are vague. Discriminating power collapses.
Application names a specific geographic area, cites a data source for the identified need, and identifies at least one gap in existing services. All three elements must be present in the narrative or supporting documents.
What happens in review: any reviewer can confirm whether the three elements are present. AI can confirm the same. Scores are comparable across the pool. A score of 3 instead of 5 is diagnostic: one of the three elements is missing or unverified.
Replace evaluative adjectives with observable nouns. Not "strong methodology" but "names a specific implementation approach, lists at least three milestone events with target dates, and identifies the evaluation method." Adjectives invite interpretation. Nouns invite confirmation.
Require evidence to be locatable. Every anchor should describe where in the application the evidence can be found: "in the narrative section," "in the budget detail," "in an uploaded letter of support," "across narrative and budget combined." A reviewer who cannot find the evidence in the location specified scores the criterion accordingly. AI applies the same rule.
State the threshold explicitly. "All three elements must be present" is testable. "Substantially addresses the criteria" is not. The threshold makes the difference between a 5 and a 4 something a reviewer can defend in one sentence rather than something they have to argue.
Pick analytic for any decision that funds or denies, and any process that owes feedback to applicants. Holistic rubrics assign one overall score from general impression and work for first-pass screening of very high-volume programs. Analytic rubrics score each criterion independently with weighted totals, produce higher inter-rater reliability, enable bias detection across criteria, and provide the diagnostic detail required to give applicants useful feedback. Analytic rubrics are also far more compatible with AI scoring because each criterion becomes a discrete analysis instruction.
Holistic rubrics earn their place in two contexts. The first is eligibility screening at the very top of the funnel, where the question is binary: does this application meet the basic threshold to advance to substantive review? A pass / partial-meets / does-not-meet decision is faster with a holistic rubric than with an analytic one. The second is volume triage in programs receiving 1,000 or more applications, where the goal of the first pass is to identify clear non-fits rather than to score for award decisions. Substantive scoring still happens analytically further down the funnel.
Outside those two cases, analytic rubrics win. The cost of slower scoring per proposal is paid back many times over in defensibility, in bias detection, in applicant feedback quality, and in compatibility with AI scoring.
In practice, most foundation review processes use a hybrid: a holistic eligibility screen at intake, then an analytic substantive review for everything that passes. The eligibility screen often happens in the application form itself through required fields, not in a separate scoring step. The analytic review is where the rubric becomes the operating document. The two stages have different rubrics, different scales, and different reviewers. Conflating them is a common source of confusion in rubric design.
A 5-point analytic scale with 4 to 6 criteria works for most grant programs. Brown University Sheridan Center research confirms inter-rater reliability decreases as quality levels increase beyond 5. The NIH 9-point scale suits research funding where a fraction of a point determines a multimillion-dollar award. A 3-point scale fits eligibility screening. The 2025 NIH framework update treats different criterion types with different granularity intentionally: 1-9 for merit, binary for credentials, because the credential criterion is where institutional bias would otherwise inflate scores.
Three quality levels. Simple to train, fast to apply, coarse in discrimination. The reviewer is choosing among three options, not seven or nine, so the scoring decision is faster but produces less differentiation across the pool.
Best for: eligibility screening at intake. Binary-ish decisions where the program is asking whether the application clears a basic threshold. Sufficiency assessments where the credential is either acceptable or it is not.
Not appropriate for: final funding decisions or any context where applicants will receive feedback. The diagnostic value is too low.
The most common choice in grant review. Five quality levels provide enough granularity for meaningful differentiation without overwhelming reviewers with false precision. Brown research finds reliability peaks at 4 to 5 levels and degrades thereafter.
Best for: any final review decision. Any program owing feedback to applicants. Any context where bias detection across criteria matters. The 5-point scale is also the most AI-compatible: each level maps to a distinct evidence profile that the AI can confirm or deny.
Recommendation: default to this scale unless there is a specific reason to deviate. Pair with 4 to 6 weighted criteria.
The NIH peer review scale runs 1 (exceptional) to 9 (poor) for research grant scoring. High granularity is appropriate for research funding where small differences in scoring determine multimillion-dollar awards across thousands of competing proposals. Requires extensive reviewer training and detailed anchor descriptions at each of the nine levels.
The NIH 2025 framework update: retains 1-9 scoring for "Importance of the Research" and "Rigor and Feasibility" criteria, but switches "Expertise and Resources" to a binary sufficient / insufficient assessment. The change is deliberate: the credential criterion is where institutional reputation bias would otherwise inflate scores, so the granularity is removed where it does the most harm.
Best for: research funding. Federal scientific peer review. Programs awarding 50+ million dollars per cycle where review investment is justified by stakes.
Increasing the number of levels does not increase the accuracy of scoring. It increases the number of choices the reviewer must defend without making the underlying evidence any more or less present. A program that cannot defensibly differentiate a 4 from a 5 on a 5-point scale will not defensibly differentiate a 7 from an 8 on a 9-point scale. The scoring will move but the reliability will not.
The reliable signal is the difference between adjacent observable evidence states. Three present out of three required is a 5. Two present is a 3. Zero present is a 1. Reviewers and AI can both confirm those states. Reviewers and AI both struggle to defend the difference between "very strong" and "extremely strong" when the anchors do not describe distinct evidence.
Legacy grant management platforms treat the rubric as a scoring template that humans consult. Sopact Sense treats the rubric as an instruction set that directs AI analysis. The difference is architectural, not cosmetic. A template asks reviewers to read every proposal end to end and convert their judgment into a number. An instruction set asks AI to read every proposal against every criterion, propose a rubric-aligned score, and attach the specific sentences from the application that support each score. Reviewers validate the AI in two to three minutes instead of building the analysis from scratch in twenty-five to thirty.
When the rubric becomes an instruction set, the work of grant review changes shape. The reviewer's role is no longer to perform criterion-by-criterion analysis on every proposal. The reviewer's role is to validate AI analysis, add the judgments AI cannot make (community trust, local innovation, political feasibility, portfolio balance), and refine the rubric based on what the pool reveals.
This is what enables genuine rubric iteration. On legacy platforms, the rubric is frozen at cycle start because changing it would invalidate every score already produced. In Sopact Sense, the rubric is editable throughout the cycle because every applicant is re-scored automatically when the rubric changes. The team learns from the first quarter of applications and refines the rubric for the rest, without losing the work already done.
"Rubric criteria written in evaluative language leave interpretation to the reader. Anchored criteria describe observable evidence. The first version produces a different score from every reviewer. The second produces the same score from any reviewer or AI."
The grant scoring pipeline in Sopact Sense has four steps: define the rubric with analysis prompts, AI reads every application for meaning, AI proposes scores with citation evidence, humans validate and adjust. The pipeline is the same whether you have 30 applications or 3,000. The work that scales linearly with applicant count on legacy platforms (reading, criterion-checking, evidence-extraction) is performed once by AI for the entire pool. The work that does not scale linearly (judgment, validation, contextual override) is what humans actually spend their time on.
Build the rubric the way you would on any platform: criteria, weights, quality levels, anchor descriptions. Then add one new element to each criterion: an AI analysis prompt. The analysis prompt is a one-paragraph instruction telling the AI what evidence to extract and how to evaluate it. The prompt is written in natural language by the program team, not in code.
Example for Methodology and Approach (weight 30%): "Extract the methodology section. Identify the specific approach (cohort model, train-the-trainer, direct service, other). Check for implementation timeline with named milestones, evaluation methodology, and budget allocation for evaluation. Flag if evaluation budget is zero or below 5% of total budget."
When applications arrive, Intelligent Cell processes each one against every criterion. The AI is not scanning for keywords. It is reading for meaning. Does the methodology section describe specific activities, or only general intentions? Does the budget align with the proposed scope (a proposal claiming to serve 500 participants with a 20,000 dollar budget raises a feasibility flag)? Are stated outcomes measurable and time-bound? Does the team description connect individual qualifications to specific project roles?
This step is the most time-expensive in legacy review and the most consistent in AI review. Every proposal receives the same depth of reading regardless of submission order or session timing.
For each criterion, AI proposes a score and shows the evidence behind it as quotes pulled from the application text. Example for the methodology criterion: the AI proposes 4 of 5 (Good) and cites the 12-week cohort model description from page 7 of the narrative, the four named milestones from page 9, the pre / post survey evaluation plan from page 14, and the 3.5% evaluation budget allocation from line 18 of the budget. The reviewer sees not just the score but exactly which sentences produced it.
This is the difference between a score and an audit trail. Every score is defensible because every score is traceable to specific applicant text.
The reviewer reads the AI proposal in 2 to 3 minutes. They confirm whether the cited evidence supports the proposed score. They adjust up or down if they disagree. They add reviewer notes capturing the judgment AI cannot make: community trust, innovation in local context, political feasibility, fit with portfolio balance, conflicts of interest. The final score reflects human judgment informed by AI analysis, not AI judgment alone.
Reviewers who override AI scores leave a reason. Across cycles, the pattern of overrides becomes a calibration signal: where humans systematically disagree with AI, the rubric anchors or analysis prompts get refined.
The compression is not from cutting corners. It is from removing the work that does not require human judgment. AI does the reading, the criterion-by-criterion evidence extraction, and the score proposal. Humans do the validation, the contextual judgment, and the rubric refinement. A 500-application review that took 250 reviewer-hours on legacy platforms takes 40 to 60 hours in Sopact Sense, with better consistency and full evidence trails.
The 80% time saving is real. It is also not the most important outcome. The most important outcome is that every application receives the same rigorous analysis against every criterion. No proposal skipped because a reviewer was tired. No criterion overlooked because a reviewer focused on the narrative and ignored the budget. No score inflated because the reviewer recognized the applicant's institution.
Most grant programs invest weeks in rubric design and almost nothing in evaluator instructions. The rubric tells reviewers what good looks like at each scoring level. Evaluator instructions tell reviewers how to actually work the rubric: where to find evidence, what to do when applicants present mixed evidence, when to override an AI proposal, how to record reasoning, and how to flag conflicts of interest. Without explicit instructions, every reviewer interprets the workflow differently and scoring drifts even when the rubric itself is anchored.
For each criterion, name the application section where the evidence lives. Significance evidence might be in the narrative pages 1-3 and the supporting data attachment. Methodology evidence is typically in the narrative pages 4-8 and the budget detail. Sustainability evidence is in the narrative final pages, the partner letters, and the multi-year budget projection.
Naming the evidence location does two things: it speeds up review by directing the reviewer to where the answer lives, and it sets a uniform expectation. A reviewer who looks at only the narrative and ignores the budget cannot score methodology accurately. The instruction makes this explicit.
Applicants rarely present clean evidence states. The narrative may name a specific geographic area but cite no data source. The budget may include an evaluation line but at 2% of total. The methodology may describe the approach but skip the timeline. Evaluator instructions describe the decision rule for mixed evidence: which elements are essential, which are nice-to-have, and how to weight partial fulfillment.
Example rule: "If two of three required elements are present, score 3 (Satisfactory). If all three are present but one lacks specificity, score 4 (Good). If one of three is present, score 2 (Needs improvement). Score 5 (Excellent) only when all three are present with specificity AND the application names a source for each."
Reviewers using Sopact Sense receive AI-proposed scores with citation evidence. The validation instruction set covers: confirm that the cited evidence actually appears in the application as quoted, confirm that the cited evidence supports the proposed score, decide whether other evidence in the application changes the score, and either accept the AI score, adjust by one point with a reason, or override completely with a reason.
Adjustments and overrides leave a reasoning note in 1 to 2 sentences. Over time, these notes become a calibration signal: criteria where reviewers systematically adjust the AI score in one direction reveal anchors that need rewriting.
Every score gets a reasoning note. Not a paragraph. One to two sentences capturing what was decisive. "Methodology described clearly with named milestones; evaluation budget at 3.5% is below the 5% threshold for this criterion's level 5." Reasoning notes serve three purposes: they make the reviewer's score defensible if challenged, they provide raw material for applicant feedback, and they give the program team a window into how the rubric is being applied in practice.
Reasoning notes do not replace AI citation evidence. They sit alongside it. AI provides the evidence trail. The reviewer provides the judgment trail.
Evaluator instructions must include a clear flag mechanism for two situations. First, conflicts of interest: prior employment with the applicant organization, personal relationship with named staff, financial interest in the proposed work. The flag is mandatory; reviewers do not score the application. Second, out-of-expertise applications: a reviewer trained in early-childhood education cannot reliably score a clinical research proposal. The flag triggers reassignment to a more appropriate reviewer.
The flag mechanism should be one-click and visible at every stage of review. Hidden flagging is a failure mode: reviewers who feel uncomfortable but cannot find the mechanism either score uncomfortably or quietly disengage.
The applicant proposes a 10-week training cohort for early-career foundation program officers. The narrative describes the cohort structure on page 6 and names three milestone events (kickoff retreat, midpoint check-in, final showcase) with target dates. The budget allocates 4,200 dollars to evaluation, which is 3.5% of the total program budget. The evaluation plan in the narrative mentions pre / post surveys but does not name a survey instrument or describe analysis methodology.
The AI proposes a score of 3 (Satisfactory) on Methodology with these citations: cohort structure (narrative p.6), three milestones with dates (narrative p.6), evaluation budget at 3.5% (budget line 18, below the 5% threshold), survey-based evaluation without named instrument (narrative p.11).
The reviewer reads the AI proposal, confirms each citation appears in the application as quoted, and considers whether to adjust. The mixed-evidence rule for this criterion is: "If three of four elements are present and one is partial, score 3. If three of four are present and all four are specific, score 4." The reviewer reads the elements as: structure (yes), milestones (yes), evaluation method (partial, named approach but not instrument), evaluation budget (no, below threshold). Three of four are present, one is partial, evaluation budget fails. The reviewer accepts the AI score of 3 and adds a reasoning note: "Cohort structure and milestones are concrete; evaluation methodology is the weak point: pre/post is named but instrument is not, and budget at 3.5% signals underinvestment."
Total time on this criterion for this proposal: about 90 seconds. The reviewer reads the AI proposal, scans the cited sections in the application to confirm, applies the mixed-evidence rule, and records reasoning. The reviewer is not searching the 20-page narrative for methodology evidence; that work happened in step 2 of the pipeline.
Most grant scoring rubrics live somewhere between stage 1 and stage 3 of the maturity ladder. Stage 1 rubrics use evaluative language and produce drift. Stage 2 anchors the criteria in observable evidence. Stage 3 adds explicit weighting. Stage 4 turns each criterion into an AI analysis prompt. Stage 5 produces citation-level audit trails from score back to applicant text. The matrix below shows what changes across the five stages for the six design dimensions that most affect scoring consistency. The ★ marks the stage Sopact Sense ships at by default.
| Dimension | Stage 1 Evaluative |
Stage 2 Anchored |
Stage 3 Weighted |
Stage 4 AI-prompted |
Stage 5 ★ Citation-audit |
|---|---|---|---|---|---|
| Criterion language | "Strong", "compelling", "demonstrates" | "Names a specific X, cites Y, identifies Z" | Observable evidence + threshold | Observable evidence + AI analysis prompt | Observable + prompt + evidence locator |
| Quality levels | 1 to 5 with vague adjectives | 1 to 5 with element counts | 1 to 5 with element + specificity | 1 to 5 with element + specificity + source | 1 to 5 with all of the above + cited examples |
| Weights | Equal across criteria, or unstated | Equal across criteria, or unstated | Explicit per criterion, total 100% | Explicit + defendable in one sentence | Explicit + refined cycle over cycle from outcomes |
| Evidence trail | None | Reviewer notes (optional) | Reviewer notes (required) | AI proposal + reviewer notes | Sentence-level citations + reviewer reasoning |
| Mid-cycle change | Effectively frozen | Effectively frozen | Possible but requires re-review | Trigger AI re-score across pool | Trigger AI re-score + diff vs prior scores |
| Time per proposal | 30 to 40 minutes | 25 to 35 minutes | 25 to 35 minutes | 5 to 8 minutes (AI + validate) | 2 to 3 minutes (AI + validate, citations on screen) |
Find the row that describes your current rubric most accurately for each dimension. The dimension where your rubric sits at the lowest stage is the biggest improvement opportunity. A rubric at stage 3 in five dimensions and stage 1 in criterion language will still drift, because evaluative language in even one criterion creates an interpretation gap that absorbs scoring variance from the rest.
Most foundations moving from legacy platforms to Sopact Sense make the jump from stage 2 or stage 3 (anchored, possibly weighted) to stage 5 (citation-audit) in a single rubric refresh. The intermediate stages are not gates the team has to pass through. They exist as descriptions of where existing rubrics sit at the moment of the transition.
Open-RFP foundation grants often sit at stage 2: anchored criteria written by a program officer who has run the cycle before, scored manually by a panel that reconvenes annually. The bottleneck is reviewer time and score defensibility against challenge.
Fellowship and scholarship programs sit at stage 1 or stage 2: criteria written for a personal essay context that does not translate cleanly to evidence locators. The bottleneck is consistency across reviewers who change cohort to cohort.
NIH and federal research peer review sit at stage 3: explicitly weighted criteria, panel scoring, multi-day reviews. The 2025 NIH framework move toward binary credential assessment is an early signal of where federal review is heading.
Accelerator and pitch competitions sit at stage 1 or stage 2: criteria written for fast review of standardized application formats. The bottleneck is volume: 200+ applications scored in two weeks by part-time judges.
Foundation portfolio reviews sit at stage 3 or stage 4: experienced program teams have weighted criteria and may have started using AI for evidence extraction in pilots. The bottleneck is connecting application-stage rubric scores to post-award outcomes, which is what stage 5 enables.
A grant scoring rubric is a structured tool that defines how reviewers evaluate funding applications. It lists the criteria a proposal must address, the quality levels at which each criterion can be scored, anchor descriptions of what each level looks like in evidence, and weights that reflect how much each criterion matters to the funding decision. The best rubrics use observable evidence descriptions, not evaluative adjectives, so any reviewer or AI applying the rubric arrives at the same score.
Six steps: pick analytic over holistic for final decisions, choose a 5-point scale, select 4 to 6 criteria tied to program goals, write anchored quality-level descriptions using observable evidence not adjectives, weight criteria to total 100%, and write AI analysis prompts plus evaluator instructions. The rubric becomes more useful with every cycle because prior outcomes refine criterion weighting. Most foundations land on 4 to 6 criteria with weights between 15 and 30 percent each.
Holistic rubrics produce one overall score per proposal from a general impression. They are fast but low-reliability and produce no diagnostic information about why a proposal scored where it did. Analytic rubrics score each criterion independently with weighted totals. Analytic rubrics produce higher inter-rater reliability, enable bias detection, and provide specific applicant feedback. They are also far more compatible with AI scoring because each criterion becomes a discrete analysis instruction. Use holistic only for first-pass screening of very high-volume programs.
A 5-point analytic scale with 4 to 6 criteria works for most grant programs. Brown University Sheridan Center research shows inter-rater reliability decreases as quality levels increase beyond 5. A 3-point scale is appropriate for eligibility screening. The NIH 9-point scale suits research funding where small differences determine multimillion-dollar awards. The NIH 2025 update uses 1-9 for merit but switches to binary sufficiency for credentials, deliberately reducing granularity where institutional bias would otherwise inflate scores.
Replace evaluative language with observable evidence. Instead of "demonstrates a strong understanding of community need", write "names a specific geographic area, cites a data source for the identified need, identifies at least one gap in existing services; all three elements must be present." The first version produces a different score from every reviewer. The second produces the same score from any reviewer or AI because it describes what is in the application rather than how the reader feels about it.
Yes, when the rubric is built as an instruction set rather than a passive template. In Sopact Sense, each criterion carries an analysis prompt that directs AI evidence extraction. The AI reads every applicant submission against every criterion, proposes a rubric-aligned score, and attaches sentence-level citations from the original text. The reviewer validates the analysis in 2 to 3 minutes instead of building it from scratch in 25 to 35. Update a criterion mid-cycle and the entire applicant pool re-scores automatically.
Evaluator instructions cover five things: where to find each piece of evidence in the application packet, how to interpret quality-level anchors when applicants present mixed evidence, how to validate AI proposed scores and when to override, how to record reasoning in 1 to 2 sentences per criterion, and how to flag conflicts of interest or applications outside expertise. Without explicit instructions, every reviewer interprets the workflow differently and scoring drifts even when the rubric itself is anchored.
Weights reflect program priorities and must total 100 percent. A common distribution for project funding: significance and need 20 to 25%, methodology and approach 25 to 30%, organizational capacity 15 to 20%, sustainability 15 to 20%, evaluation design 10 to 15%. Research funding weights methodology higher. Equity-focused funding may carry equity as a 20 to 30% standalone criterion. The right weights are the ones the program team can defend in a one-sentence explanation per criterion.
Citation-level scoring means every score is tied to a direct quote from the applicant's submission as evidence. When AI proposes a 4 of 5 on methodology, it shows the specific sentences in the narrative, budget, or attachments that support that score. Reviewers see exactly why the score was assigned and can challenge or accept it on the evidence rather than on a black-box rating. This produces an audit trail from score back to applicant text, which legacy scoring platforms cannot.
On legacy platforms, no. Any rubric change requires re-reviewing all applications already scored. In Sopact Sense, yes: update a criterion definition, change a weight, or add a new criterion, and the entire applicant pool re-scores automatically against the new rubric. The audit trail preserves both the original and revised scoring so the team can compare. This is what enables rubric refinement based on what the pool reveals, rather than locking decisions to the rubric that existed before the applications arrived.
How application-stage rubric scoring connects to post-award outcome tracking, portfolio-level pattern detection, and grantee feedback loops that survive multiple cycles.
Bring your existing rubric and a sample of last cycle's applications. Sopact Sense will show you what anchored scoring looks like in your context, what the AI proposes for each criterion, and where the rubric is producing drift today. Book a working session with the founder.
