Compare a backlog with its stated priorities

Load CSVs → compute soft alignment, coverage vs. target prior, entropy, nudges → export results. Local files and calculations.

Try the fictional backlog, inspect a close match, then change a target share. Text and tags estimate an allocation of effort across themes; they cannot establish actual strategic value.

Match weights are normalised lexical similarities, not calibrated probabilities. A work item with no matching words or tags remains unallocated. Target edits change the comparison, not the matches. Candidate moves are independent suggestions for discussion, not a jointly optimised plan.

Method, limits and original design idea

The English-oriented tokenisation ignores word order. The model uses term-frequency/inverse-document-frequency vectors and cosine similarity, boosted by tag overlap. Top-K retains positive matches, including all ties at the boundary; selected scores are exponentiated and normalised (softmax). Coverage divides attributed effort by all effort, so unmatched work remains visible rather than being redistributed. Entropy describes the allocation weights, not the quality of the work. Unmatched work has no allocation distribution: its entropy is N/A and it is excluded from the entropy histogram and ranking.

The original Intent Field Navigator imagined a live intent codex, work harvesting, embeddings, a graph field solver, integration with GitHub/Jira/Notion and adaptive portfolio steering. Those remain design hypotheses. This working example reads local CSV files and uses a transparent lexical heuristic. Its graph roughness is a stated weighted squared-difference diagnostic, not a solved field or an optimal investment decision.

Optional outcome updates use a hand-built standardised score: outcome minus half lead time and half defect rate. A declared realised intent takes precedence; otherwise the top lexical match is used. One update per reset is allowed, and zero target shares remain zero. There is no causal attribution or reinforcement-learning guarantee.

Based on Intent Field Navigator, source revision 9d9c253, reviewed 2 October 2026. See the authors’ introduction to text weighting. All scripts are local; nothing is uploaded. Keep the source CSVs to retain your inputs.

Load the demo or choose your CSV files.

1) Data

2) Parameters

3
0.08
1.30
0.10

Exports

CSV Schemas

IDs must be unique. Supply every active target weight or leave all blank for equal shares. Missing effort defaults to 1; zero effort stays zero. Quote commas, quotes and line breaks. Limits: 2,000 rows per file, 30 active intents, 5 MB per file and 5,000 distinct words/tags. Numeric inputs must be finite, nonnegative and at most 1 trillion. Defect rate is a fraction from 0 to 1. Legacy export columns p_top/probability contain allocation weights, not calibrated probabilities. Unmatched items have a blank entropy field (N/A).

intents.csv: intent_id,name,description,tags,target_weight,active
work_items.csv: item_id,title,description,tags,type,status,squad,effort,url
outcomes.csv (optional): item_id,outcome,lead_time,defect_rate,realized_intent

Dashboard

Explainers

Attributed Effort vs Target Shares

Coverage Gap (attributed − target)

Allocation entropy (matched items)

Graph roughness proxy

Weighted squared difference between top allocation weights and target shares. A comparison aid; no validated threshold or diagnosis.

Intent Weight Sliders (editable)

Changing a slider renormalises target shares and updates charts, candidate moves and exports immediately. Reset target shares restores the loaded example.

Totals

Most Spread Allocations (matched items)

ItemSquadEntropyTop-2 weightsTags

Candidate Moves

ItemFrom → ToΔgap (est.)ΔpEffort
Chart values and method diagnostics
Ranked Matches (per item)

        

          

            
ItemTop IntentTop weightTop-3TitleSquadTags