Pathway, evidence & changes

Read your complete Pathway

Give the connection progress:read. GET /pathway returns the same canonical chapter, section and detailed-skill tree used by Wingman, including descriptions, precise targets, ratings, evidence strength, recent paper links, working observations and baseline progress. Lite includes full agent Pathway access even though its browser view is baseline-only.

Use GET /skill-profile?taxonomyVersion=wm-pathway-2026-10-v1 for the detailed taxonomy's arithmetic evidence facts and watermark. Omitting taxonomyVersion preserves the historical default. Keep versioned targets exact; a broad section result does not assess every narrower skill.

Pathway ratings are explainable practice heuristics, not exam grades or calibrated probabilities. The response includes level explanations and an algorithm version. Recent-work lists are summaries; use paginated evidence reads for a complete account.

Trace the evidence

GET /skill-profile/records returns current attempt revisions, including explicit retraction tombstones. GET /skill-profile/history preserves every recorded revision, including marks, status, feedback, targets, confidence reports, assistance and source provenance.

Both support after, through, limit, attemptId, sessionId, questionId, sourceKind and sourceId. For an exact target use taxonomyVersion with topicId, or specificationId with specificationVersion and reference. Filters combine within your own account and match before or after evidence. Topic filters include descendants in the same explicit taxonomy; narrower filters never match broad-parent-only evidence. Specification triples match exactly.

Follow nextCursor as after and keep the first response's watermark as through across pages. Current/history pages allow up to 100 records. Keep sequence values as decimal strings, including values above JavaScript's safe integer range. Do not count revisions as independent attempts.

GET /skill-profile/insights?sessionId=ID compares a paper's topic results with earlier work. Mapping and criterion reads expose current revisions, effective targets and permitted corrections. These topic comparisons do not replace the exact change ledger.

Correct the skills assessed by each mark

With progress:read, GET /skill-profile/criteria?questionId=ID returns the current rubric, revision, saved criteria overrides, effectiveCriteria and permitted options. Its additive eligibility field gives editable and reason; older responses retain the editable boolean. A saved review remains compatible when its individual allocations match the rubric.

If eligibility.editable is false, use the returned reason: baseline_fixed_mapping keeps reviewed baseline links fixed; manual_review_without_rubric has manual marks without matching allocations; invalid_rubric has no valid source rubric; no_criterion_result has no individual-mark result yet; incompatible_criterion_result means the current result does not match the rubric. Account restrictions, plan access and key scopes are checked separately. Preserve existing reviews and marking when explaining a restriction.

With an authorised correction and mapping:write, POST /skill-profile/criteria accepts questionId, revision, the unchanged rubric and the complete desired criteria override list. Use rubric.criteria[].id as criterionId and exact versioned targets from options. Merge the requested corrections into existing criteria so unrelated overrides survive. To correct a part, identify its rubric partId and correct only the allocations whose requirements support that skill; a displayed part label or repeated M1/A1 code is not a criterion ID.

Read the mappings again after saving to verify the effective targets. A conflict requires a fresh read and reconciliation. Mapping corrections preserve marks, saved reviews, source pins and credit history. Do not remove reviews, rerun paid marking or replace an import to force a skill correction. Question-level topic corrections remain available where permitted but do not establish unsupported individual skills.

See what your paper changed

GET /skill-profile/changes lists exact before/after evidence revisions, observation differences and contribution facts. Filter by sessionId to follow a marked paper, including later reviews and mapping corrections. Pages default to 20 changes and allow at most 50. A 4 MiB soft page budget can return fewer complete events; the first indivisible event is retained even if larger. A short page is not the end when nextCursor is present. Use nextCursor as after and retain the first watermark as through when reading more pages.

GET /skill-profile/changes/{sequence} also reconstructs the affected Pathway ratings before and after that event, using the same timestamp on both sides. ratingImpact.basis is reconstructed: the evidence revisions are exact recorded data, while ratings are calculated using the returned algorithmVersion and asOf. An optional timezone-qualified asOf query makes comparisons across multiple events consistent.

Changes can add evidence, correct marks, move mappings, retract or restore a contribution, or update qualitative working observations. A mapping correction does not create another sitting, and qualitative observations do not add marks. Provisional, pending and excluded evidence remain explicit.

python3 scripts/wingman.py changes --session SESSION_ID --limit 20
python3 scripts/wingman.py change --sequence EVENT_SEQUENCE

Keep saved results and projection readiness distinct

Import completion and authorised result reads can return a profile receipt with evidenceRecorded, mappingPending, sessionId, watermark and changes: {path, after, through}. Use its session-filtered changes path to inspect the recorded impact. Full profile/change reads require progress:read independently of marking permissions.

The assessment is saved before profile refresh. evidenceRecorded=false or mappingPending=null means projection needs recovery; a positive pending count means classification is still running. A later authorised profile/change read can recover the retained feed. Keep the session and receipt if a bounded check is still pending. After paging one fixed watermark, check later changes without the old through boundary so later classifications remain visible. Do not repeat marking or create a replacement import to obtain a profile update.

Ask through the skill

The Python helper supports pathway, profile --taxonomy-version, records, history, changes, change --sequence, insights, mappings and criteria. Its progress reference explains exact filters, pagination and interpretation. Read access does not permit raw score or profile writes.

Show my weakest Further Maths skills and the evidence behind them. Then explain exactly what my latest marked paper changed, separating recorded marks from pending mappings and reconstructed Pathway ratings.