SPARC dynamic instruction browser acceptance: October 12, 2026
Status: local acceptance is partial. This walkthrough used the existing localhost:3200 hotfix application, an authenticated admin in Chrome, configured live AI providers, private uploads of both new ZIPs, and synthetic answers. No production deployment, commit, push, practice reset, or destructive history change occurred. The original lesson packages were preserved.
Observed cases
| Case | Observed result |
|---|---|
| Upload and opening | Both new dynamic confidence-interval variants uploaded successfully. Each opened with the existing canned question and seven expectations / three misconceptions. |
| Feedback before choice | A synthetic specific-interval probability misconception received substantive corrective feedback before the four student options. Selection did not create another assessed answer. |
| Covered-expectation progression | The initial student implementation repeatedly offered E2 despite coverage 1. The corrected projection excludes covered expectations from automatic selection. A subsequent answer advanced from E2 to unfinished A; answering A advanced the eligible set, and answering C later led to B. |
| Explanation and question | Student instruction on interval construction A displayed explanation and question in one tutor box, with one diagnostic. The selected button remained immediately before that instruction. |
| Switching KCs | After answering A, the student used Move to the next KC before requesting rich instruction. This produced instruction on C rather than exercising every move repeatedly on E2. Navigation itself produced no additional learner assessment. |
| Worked example | C produced an explanation, a complete t-interval example with formula and numerical reasoning, and an isomorphic question with new inputs. DOM inspection confirmed one tutor bubble, three components, and question-mark counts 0 / 0 / 1. |
| Question alone | After answering C, the student selected Question on B. The generated question concerned margin of error for sample sizes 25 and 100, with no instructional explanation appended. |
| Choice history and reload | Reload restored the worked-example instruction and its saved choice. Saved navigation and instructional selections remained disabled selected buttons at their original locations in the transcript. |
| Agent selection and figure | The agent recognized interval-construction evidence, selected unfinished E2, and generated explanation / figure / isomorphic question. Diagnostics showed expectancy logit -1.39, assistance 2.77, target logit 1.39, temperature 0.1, and rounded selection probability 1. One tutor bubble contained three components and one figure; question counts were 0 / 0 / 1. Reload restored that composition. |
Defects found and repaired
- New student projection inherited eligibility for covered expectations from the older student policy. New dynamic automatic selection now excludes them; explicit next-KC navigation retains the authored cycle. Older policy behavior remains intentional and separately tested.
- Separately persisted primitives were presented as separate tutor boxes. The dialogue projection now assembles the primitives of a selected move into one tutor bubble in component order, preserving their durable identities for retry.
- Consuming the menu removed the learner's selection from visible history. Saved decisions now project disabled selected buttons between the preceding feedback and resulting instruction. Navigation selections are also retained. A failed instructional request still permits retry of only the accepted active choice.
- The rich primitive inherited a complete-move instruction to end with a question, causing a learner question inside the figure description and another from the question primitive. The prompt now distinguishes the completed move from each primitive. Explanation, image brief, and figure description contain teaching material only; only the question primitive asks a learner question. Newly generated worked-example and figure cases respected this boundary. Earlier generated content remains in history.
- The first seven-KC agent assessment exceeded the 16 KiB history-extension budget: 16,392 bytes. It bundled initial and updated observations for every KC with assessment. Competitive cycles now save the evidence checkpoint and each KC's derived model observations atomically in separate canonical transitions. A regression reproduces an oversized bundled record and verifies bounded updates and interrupted-write replay without rescoring or duplicate gains. The same live agent case then completed. The budget was not increased and no content was truncated.
Remaining limits
- The new AI-generated repeated-sampling figure contains duplicated or out-of-order sample labels. Figure generation, private delivery, composition, and reload worked; semantic image accuracy is not fully validated. No automated image review or silent replacement was added.
- Long rich moves exceed the chat viewport height and require scrolling. Screenshots of the latest question show only the lower portion of those single bubbles.
- Source edits during the walkthrough caused temporary hot-reload interruptions. Settled-client reload restored accepted history. This does not establish that unsaved in-flight work survives arbitrary source replacement.
- The walkthrough did not exercise final seven-KC graduation in either condition, all misconceptions, every target/move combination, mobile viewports, non-admin access in a live browser, or provider failure/retry in the settled client. Relevant isolated contracts cover completion, rejection, persistence failure, and retry, but are distinct from live evidence.
- A fresh full Meteor integration test was not run; repository instructions require separate authorization for that runner.
Verification and retained evidence
Full application typecheck passed. Lint passed with zero errors and 28 existing bundled-package warnings; server dependency, test-artifact, and 312-surface security checks passed. The isolated SPARC/general/client runner passed 218 tests in 21 suites. No staging or commit occurred.
Screenshots are retained locally with the implementation plan: 05-single-box-explanation-and-choice.jpg, 06-worked-example-single-box.jpg, 07-question-only-and-retained-choice.jpg, and 08-agent-figure-single-box.jpg. Earlier captures document the pre-repair states and are not treated as passing evidence for the repairs.
Authorized Meteor integration verification — October 12, 2026
The user authorized the full Meteor suite and up to three further runs after failures. The first invocation passed 16 harness checks and 795 server tests (15 pending), then exited with Windows process status 0xC0000409 as Chromium client testing started. It reported no failing assertion or completed client result; no source defect was identified from that exit. The second invocation, using the same pinned runner and settings with a fresh disposable CI Mongo replica set, completed successfully: 16 harness checks, 795 server tests (15 pending), and 1,104 client tests (7 pending), with SERVER FAILURES: 0 and CLIENT FAILURES: 0. This includes the new real-Mongo figure-library lease and decision contracts. No further integration retries were used. Test-driver resolutions were cleaned by the supported runner; the disposable test database was removed and the existing localhost watcher was restored. This establishes integration coverage, not deployment or release-image qualification.