feat(junior-profile): predict-before-reveal loop with evidence-based adaptive fading

[DanielPodolsky]

What

Reworks the Junior profile from "hand-type every line" (outdated) to "evaluate everything" — without falling into the passive-observation dependence trap. The junior now commits a prediction (approach / data structure / control flow / edge cases) before the AI reveals code, then an AI judge grades it against a rubric, names the specific gap, and grounds the feedback. Scores are logged as a judgment-growth curve.

Why it's where it is

• The loop lives in /own:feature Phase 6 (the command that always runs), not /own:guide (rarely run) — so the gate is hard, not skippable under fatigue.
• Friction applies only to Implementation-group tasks; Setup/Verification flow.
• Gate 1 (Ownership) in /own:done now uses the prediction transcript as evidence, not vibes.

Evidence-based (not a hunch)

Validated against two streams (full report: docs/research/junior-profile-predict-before-reveal-validation.md):

• Learning science (adversarially-verified deep research): Tucker et al. 2024 RCT — predicting code beats producing it on learning, transfer, and motivation; reinforced by the pretesting / errorful-generation literature.
• Industry practice (real GitHub artifacts): ownership = accountability for understanding, not typing (kornia, psimm, github/docs); one unrelated project (hotaru) independently reinvented the same model; vapor reserves good-first-issues for human learning.

The research surfaced one gap — adaptive fading — now implemented (P0): the gate fades per-dimension after sustained MATCHes (expertise-reversal effect), driven by the Prediction Scorecard we already log. Cautious thresholds: 4 consecutive MATCHes → faded/spot-check; PARTIAL/MISS resets; ownership floor never removed.

Tested

Manually run end-to-end on a real project (HR voice-screening prototype). The loop, the gate, and per-dimension fading behave as designed; first real prediction is logged.

Known debt (tracked, not blocking)

• Octocode grounding is retrieval, not a quality oracle — prevalence ≠ correctness. Source-vetting deferred.
• P1 (worked-example escape hatch for cognitive-overload tasks) and P2 (self-explanation at the "Own" step) are recommended in the research but not yet built.

Files

• profiles/junior.md — The Implementation Loop + Adaptive Fading
• .claude/commands/own/feature.md — Phase 6 (hard gate, reads scorecard, fades per-dimension)
• learning/LEARNING_REGISTRY.md — Prediction Scorecard
• docs/research/… — the validation report

[claude]

Review — predict-before-reveal loop with adaptive fading

The substance here is strong, and I want to be clear up front: this does not weaken Gate 1 — it arguably strengthens it. The old "type every line" junior model conflated keystrokes with ownership; this reframes ownership as evaluation, not transcription, which is exactly the Active Typist protocol read correctly (comprehension, not typing volume). The Predict → Gate → Reveal → Judge → Own loop forces a committed position before the reveal, and /own:done Gate 1 now consumes the prediction transcript as real evidence and probes MISSes. That is a P2 upgrade, not a regression. The design is also genuinely evidence-grounded (P3): the validation doc cites a domain-matched RCT, names its own limits, and even records a killed claim — that intellectual honesty is the founder walking his own talk on ownership.

A few things I would want tightened, none touching the soul of the product (details inline):

• P2 / OWN step — the ownership-evidence step asks only for acknowledgment, while the research doc you shipped in this same PR (§C2 P2) recommends a why-explanation. Implementation and its own evidence base disagree at the one step that matters most. Highest-leverage fix.
• P4 scoping — gating the loop to profile.type = "junior" leaves Interview Prep (the audience that most needs to defend judgment live) and Career Switcher out, framed as "they do not need the gym." Make that a deliberate, softenable choice.
• Adaptive fading floor — the "all four faded → flows like Setup" path is your max P2 exposure; the floor you wrote is load-bearing, and the 4-MATCH threshold / soft snap-back trigger deserve a "this is a conservative v1 number" caveat (your own §C3 lists it as open).

What blocks merge — shipping discipline, not substance:

• D2 (CHANGELOG) — required. This changes the entire junior implementation experience and there is no CHANGELOG.md entry. The project rule is non-negotiable: every PR adds one, and a new user-facing capability is a MINOR version bump (2.6.0 → 2.7.0). A feature this significant cannot merge undocumented.
• D1 (README/guide sync) — required. Juniors are now told not to run /own:guide ("the gym is built in"), but the README Commands table (line 107) still presents /own:guide as the implementation-guidance command with no note that it is bypassed for the Junior profile. Discoverability is a product feature — reconcile the two.
• D3 (commit scope) — nit. feat(junior-profile) / docs(research) use scopes outside the valid set; these should be feat(profiles) / docs(...). Cheap to get right next time.

Fix the two required doc items (a CHANGELOG entry + a one-line README/guide note) and I will flip this. The philosophy and architecture are approved — I am holding only on the paper trail the project insists every change carries.

VERDICT: REQUEST_CHANGES

[DanielPodolsky]

Addressed the three actionable comments in 9fd7be2:

• P2 (OWN step, junior.md:158) — implemented. OWN now requires the junior to explain why the actual approach is better and where their prediction's logic broke (self-explanation), not just acknowledge the gap. Resolves the implementation-vs-shipped-doc contradiction; applied in both junior.md and feature.md Phase 6.
• Fade threshold + soft snap-back (junior.md:194) — documented "4 consecutive MATCHes" as a deliberately conservative v1 number (research §C3 Q2 open), and added a deterministic staleness nudge: a faded dimension ≥5 tasks stale is re-asked once (lightweight P3 spaced-re-prediction), so re-engagement no longer depends solely on a soft judgment trigger.
• "Don't need the gym" framing (feature.md:49) — reframed as deliberate v1 scoping; names Interview Prep + Career Switcher as the intended next profiles, not a permanent exclusion.

Left as tracked debt (not in this PR): P1 (worked-example escape hatch for cognitive-overload tasks) and the Octocode-quality caveat. Thanks for the catch on P2 — the doc/code disagreement at the ownership-critical step was the right thing to flag.

[claude]

Incremental review — 8e7464e..9fd7be2

The three inline findings are resolved well, no new concerns in the diff itself:

• P2 / OWN step ✅ — OWN now demands why the actual approach is better and where the prediction's logic broke (self-explanation, not acknowledgment), applied in both junior.md:156 and feature.md Phase 6. The highest-leverage fix, and it lands exactly where the shipped research doc said it should.
• Snap-back determinism ✅ — the 5-task staleness nudge turns re-engagement from a soft "on a complex task" judgment call into a scorecard-driven trigger. Faded-but-rusty dimensions now re-engage on their own.
• Fade threshold + framing ✅ — the "4" is now flagged as a conservative v1 dial (§C3 Q2 open), and "they don't need the gym" is reframed as deliberate v1 scoping naming Interview Prep + Career Switcher as next. Both copies stay consistent.

Still blocking — the two required items from the prior review are untouched in this push:

• D2 (CHANGELOG) — still missing. No 2.7.0 entry; CHANGELOG.md still tops out at 2.6.0. Every PR adds a CHANGELOG entry, and a new user-facing capability of this size is a MINOR bump. Your reply addressed the three inline comments but not this one.
• D1 (README/guide sync) — still missing. The Commands table (README line ~107) still lists /own:guide as the implementation-guidance command with no note that the Junior gym now lives inline in /own:feature Phase 6. Reconcile the discoverability story.

Philosophy and architecture remain approved — I'm holding solely on the paper trail. Add the CHANGELOG entry + the one-line README note and I flip this.

VERDICT: REQUEST_CHANGES

[claude]

Incremental review — release cut 8e4e4b3

Both required items from the prior review are now resolved, and the release sweep is otherwise clean:

• D2 (CHANGELOG) ✅ — 2.7.0 entry added, accurate and appropriately scoped as a MINOR bump (new user-facing capability, no schema break). It captures the substance honestly: ownership shifts from typing to evaluation, Phase 6 inline gate, scorecard, adaptive fading + deterministic staleness re-check, the OWN self-explanation, and the /own:guide demotion.
• D1 (README sync) ✅ — Commands table updated for /own:feature Phase 6 and /own:guide, plus a clear Junior-profile note. The discoverability story now matches the behavior.
• T1 (script parity) ✅ — .sh/.ps1 bumped symmetrically across base + project installers; dashboard-data template and README badge all consistent at 2.7.0.

One non-blocking finding (new): the version sweep is incomplete. /own:init (.claude/commands/own/init.md) still hardcodes 2.6.0 in three live spots — the standard-profile manifest write (L618), the Custom-profile meta.version (L642/L1224), and the field-mapping contract assertion "meta.version → Literal 2.6.0" (L1257). Since the installers now write 2.7.0, a freshly /own:init'd project gets a 2.6.0 manifest + dashboard under a 2.7.0 release — an internal inconsistency this commit's own job (the version bump) should have caught. Not in this PR's diff, so I can't inline it; finish the bump in a follow-up before tagging. (The 2.6.0 in DASHBOARD_CONTRACT.md is an illustrative e.g. — leave it.)

• D3 (commit scope) — nit, again: docs(release) uses a scope outside the valid set. docs(install) or chore(core) would pass. Cheap next time.

Philosophy and architecture were approved last round and are unchanged. The two blockers are cleared; I'm flipping off REQUEST_CHANGES. Flagging the init.md version miss as a COMMENT, not a block — but land it before you cut the tag.

VERDICT: COMMENT

[claude]

Incremental review — 8e4e4b3..de68549

This closes the one open item from my last pass: the incomplete version sweep in /own:init.

• D2 / version consistency ✅ — all four live 2.6.0 spots I flagged are now 2.7.0: the standard-profile manifest write (L618), the Custom-profile meta.version (L642) and its dashboard window.PROJECT.meta.version (L1224), and the field-mapping contract assertion (L1257). A freshly /own:init'd project now stamps 2.7.0 end-to-end, matching the installers. The illustrative e.g. in DASHBOARD_CONTRACT.md was correctly left alone.

Nothing else in the diff. Philosophy and architecture were approved two rounds ago; the paper trail is now whole. Ship it.

VERDICT: APPROVE