Sessions
—30-minute idle boundaryEVIDENCE, BEFORE APPLAUSE
Show the work.
Name the unknowns.
The receipt checks that same-origin release metadata, test totals, source inventory, and a provenance digest agree. It does not authenticate the deployed code or prove that people learn more from it. This page keeps those claims visibly separate.
No efficacy study has been completed. This release exposes source provenance, release integrity, and this browser’s narrowly defined activity. It deliberately withholds claims about improved learning, retention, or transfer until a preregistered study supports them.
THIS BROWSER ONLY
The local record
Active study time
—Idle gaps are not countedCards learned
—Completion, not retentionReturns completed
—Completion, not correctnessPractice hands
—Completed, not proficiencyRecall checks due
—Due is not attemptedThe Pocket Thesis outcome ledger stays on this device. It uses no analytics SDK, stores no typed learner answers, and creates no account or analytics identity. The authored self-check is the complete learning path and sends nothing. The optional AI path is intentionally disabled in this release; the Apply move stays page-only.
Read data flow and provider details
Hosting security may set essential cookies. After one lesson, your recall and Apply move can appear in a temporary Pocket Thesis Card held in this open page's memory. Making that receipt sends nothing new and adds neither field to local storage, the outcome ledger, or a server. If you requested the optional AI check, the receipt says how many recall drafts Pocket Thesis tried to send to its server; a draft may have reached that server or its configured provider. Dismissing the card clears the receipt, while Copy or Print makes a copy only because you asked. Leaving an unfinished lesson also clears its recall and Apply drafts and invalidates a pending response in the interface, though a provider request that already started cannot be unsent. The optional AI path is disabled in this public release. If enabled in a future release, choosing it would make the server first turn the transport IP into a keyed hash for minute and daily abuse counters; it never stores the raw IP or answer in that quota record, ignores expired windows, and deletes expired rows opportunistically. The server then sends the answer and card rubric to exactly one configured evaluator: Lovable AI Gateway, direct Google Gemini, or direct Anthropic. There is no automatic retry to a second paid provider. Direct Gemini requests explicitly disable request storage. Do not submit sensitive or personal information. The app stores only the result locally. Processing and retention follow the configured services' terms: Lovable privacy, Google Gemini API terms, and Anthropic commercial terms. This browser record is user-modifiable, not tamper-evident, and not an audited research dataset.
MEASUREMENTS WITH EDGES
What each number can say
Every metric is narrow on purpose. Engagement is never relabeled as learning.
- Active study time
- Sum of gaps of 60 seconds or less between local activity events, capped at two hours per session.
- Session
- A local activity period. A 30-minute idle gap, a hidden/closed tab, or a clock change starts a new session.
- Card learned
- A card whose Learn step emitted a completion event on this device. It is not evidence that the idea was retained.
- AI-checked recall
- Optional path, disabled in this release. An AI provider result that passed the app's structural, exact-evidence, and concept-facet gates. Correct means result=pass; partial and retry stay in the denominator. This is not human adjudication or proof of truth.
- What is excluded
- Only verified semantic-provider results are eligible. Unverified results, self-checks, and human grades are excluded from AI-checked recall-rate metrics.
- Delayed recall
- A recall check scheduled for one or seven days after the most recent recorded in-app lesson or source-bearing feedback exposure. While a check is pending, the app hides that card’s teaching prose in its library and source map; opening in-app lesson or feedback prose restarts every pending delay. Reading the external source or another browser tab cannot be detected. A private attempt can complete a due return, but carries no score or correctness claim.
EVIDENCE CHAIN
Source to retained decision
The product is instrumented through the first three links. The final link still needs a study. Open the 90-second learning-loop walkthrough →
- 01Source lockedAvailable
Claims point back to a versioned primary source and a visible locator.
- 02Recall observedAvailable
A content-free attempt completes a due return without storing the learner’s answer or claiming it was correct.
- 03Decision practisedAvailable
A scenario records completion and calibration—not a claim of real-world skill.
- 04Retention comparedPlanned
A randomized, assessor-blinded seven-day comparison must establish whether the format helped.
PRIMARY SOURCE 01
Situational Awareness
Open the official essay PDFPRIMARY SOURCE 02
NIST AI RMF 1.0
Open the official NIST PDFIN-PRODUCT INSPECTION
Card-level provenance
Inspect source notes in the deckTHE TEST THAT MATTERS
A randomized, assessor-blinded seven-day study
The study will ask a harder question than “Did people swipe?”: did they remember and use the ideas better one week later? The in-product return card is personal feedback—not this experiment. Open the return card →
A human scores recall without seeing the learner’s condition. Report the difference, uncertainty, and raw participant count.
A novel scenario tests whether learners can use the idea. It supports—but never replaces—the seven-day result.
Report missing, late, and withdrawn learners plus any lesson or feedback reopening, separately for each condition. A copied or printed card, an outside source, or another tab cannot be detected here, so the preregistration must say how outside exposure will be collected and analysed.
MINIMUM RECRUITMENT TARGET
≥30completed participantsThis is an operating floor, not a statistical power claim.- Design
- Random assignment with condition-blinded outcome scoring.
- Control
- A static summary using the same source material and time budget.
- Primary outcome
- Human-adjudicated free recall seven days later under a preregistered rubric.
- Transfer check
- A novel decision scenario scored without revealing study condition.
- 1Lock the plan
Preregister hypotheses, rubric, exclusions, and analysis before recruitment.
- 2Randomize fairly
Give both groups the same source and study time; change only the format.
- 3Check twice
Measure immediate recall, then invite the same learners back after seven days.
- 4Score blind
Hide condition labels from scorers and adjudicate uncertain grades.
- 5Report the misses
Publish attrition, missing data, uncertainty, and every preregistered outcome.