Reconciliation

Read this once, before the first book. Then only when a rule surprises you.

protocol.md was written at 14:44 on 2026-08-30. evidence/reading.md was written at 15:01 and evidence/reading-with-ai.md at 15:10 — seventeen and twenty-six minutes later. The evidence files contradict the protocol in fourteen places and acknowledge none of them, because they were written after it and never folded back in.

This file folds them back in. Where they conflict, the evidence wins — that was decided explicitly, and the reason is in the protocol’s own appendix:

No study tests this pipeline. The components are evidenced; the assembly is design. Treat it as a well-reasoned bet. — protocol.md:1063-ish, “What is reasoning rather than evidence”

A well-reasoned bet loses to a measured finding. Every row below cites both sides by line number so you can check the call rather than trust it.

One thing this file does not do: it does not rewrite protocol.md. That file stays as the record of what was originally designed. Where the two disagree, this file is operative and the prompt library in web-chat/prompts/ already reflects it.

The fourteen, at a glance

#	Conflict	Winner	What it changes in the machine
1	No closed-book recall stage	evidence	A required `## Recall` section + `<Recall>`
2	“No cap” on actions vs “choose ONE”	evidence	One `committed` + `tier: now` per book
3	Whole-book upload vs one chapter	evidence	P0/P3 rewritten; Stage 1 changed
4	P5’s template triggers sycophancy	evidence	P5 restructured, question-first
5	P0’s profile maximises sycophancy	evidence	P0 trimmed to the constraint set
6	P3 clarifies every chapter	evidence	`clarified:` per chapter; P3 conditional
7	Reread the recap vs retrieve it	evidence	`/review` serves the question, not the answer
8	Four-week ceiling vs 30-day gap	evidence	`nextEligible` on the book page
9	Throughput arithmetic vs volume-as-failure	evidence	No counters anywhere in the chrome
10	No mental contrasting	evidence	`obstacle:` required on a committed action
11	No knowledge/execution diagnosis	evidence	`gap:` required in `book.json`
12	Solo private site vs human in the loop	neither	Stated as a limitation, never simulated
13	Rule 11 exempts fiction, but 3 CORE books are novels	internal	`kind: lifelong`
14	Three divergent frontmatter shapes	internal	One schema

Rows 13 and 14 are internal contradictions — the protocol against itself, or against universe.md. They are here because they are settled in the same place.

Rows 2, 4, 5, 6 and 12 were amended on 2026-09-14 by his own instructions, after he read a complete sample book. The rows below stay as written; the amendments are at the foot of this file.

1 · There is no closed-book recall stage, and it is the most important one

Protocol. The twelve stages run GENERATE → READ I → EXPLAIN → READ II → DIALOGUE (protocol.md:16-41). Nothing between READ I and EXPLAIN asks you to produce anything from memory.

Evidence. evidence/reading-with-ai.md:485, Stage 5:

AI off: write a one-page summary from memory, book closed, before any AI contact. This step is the learning. Then answer every question before checking. — 61% vs 40% (Roediger & Karpicke); 89% vs 73% with LLM-generated questions (An et al.)

And :542, Guardrail 4:

Never let AI do the retrieval. Attempt from memory first, every time. This one rule prevents most of the crutch effect.

Call — evidence wins. A required ## Recall section lands before the explanation layer, carried by a <Recall> component that reads as a wall rather than a note, for the same reason <Attempt> does upstream: a reader who scrolls past it should feel that they skipped something.

/validate refuses to mark a chapter status: complete with an empty Recall. This is the single largest structural change the reconciliation makes, and it is the one that makes the difference between reading and having read.

2 · “No cap” on actions against “choose ONE behaviour”

Protocol, in three places — :187 (Stage 8), :624 (P6), :1007 (Rule 4):

Extract every genuinely distinct, relevant action this chapter supports. No cap.

At twelve chapters that is plausibly dozens.

Evidence. evidence/reading.md:443 is a section titled “One behaviour per book”, and :568 step 11:

Choose ONE behaviour. One active change at a time. Adding intentions doesn’t raise the ~53% conversion rate; it divides attention.

evidence/reading-with-ai.md:487, Stage 7: “Choose one. One implemented beats five considered.”

Call — both, and the distinction is the point. Extraction is uncapped, because an action is data and a ledger row costs nothing. Commitment is capped: exactly one action per book may carry committed: true together with tier: now. The rest are logged at next, later or reference and are not being attempted.

audit.mjs warns — never refuses — above one. The protocol’s own hedge already pointed here: “three simultaneous new behaviours is usually the ceiling before none of them survive.”

3 · Whole-book upload against one chapter at a time

Protocol. :74-:76, Stage 1:

Verify the upload before doing anything else. … Large books: split by part and upload sequentially.

Evidence. evidence/reading-with-ai.md:348:

Uploading a whole 300-page book and asking a broad question is the worst case for this effect. Upload one chapter at a time and ask specific questions.

:445 Rule 3 and :540 Guardrail 2 repeat it, citing Liu et al. on lost-in-the-middle: >30% degradation for mid-context material.

Call — evidence wins. One chapter per upload. P0 and P3 are rewritten; Stage 1 ACQUIRE drops whole-book verification and verifies the chapter map instead.

4 · P5’s dialogue template states a conclusion in every field

Protocol. P5’s template, :577-:584, has the reader fill in MY CONFUSIONS, WHERE THE EXPLANATION CONTRADICTED WHAT I CONCLUDED, MY PUSHBACK and MY QUESTION. Every field supplies a belief before the model answers.

Evidence. evidence/reading-with-ai.md:422, the top sycophancy mitigation:

Never state your conclusion before asking. Ask “what does this chapter argue?” not “this chapter argues X, right?” — Sycophantic agreement is triggered by the user conveying a belief. Withhold the belief and the trigger is absent.

The same file notes that all four practitioner guides it reviewed miss sycophancy entirely, and that self-help is “the exact input condition under which sycophancy is strongest.”

Call — evidence wins. P5 is restructured into two messages: the question first, alone, then your position in a separate turn once the model has committed. ## Dialogue on the page records both in that order, so the sequence is auditable months later — if the page shows your position first, the answer that follows it is suspect.

5 · P0 maximises personalisation; personalisation maximises sycophancy

Protocol. P0 — MASTER SETUP (:250) opens with a WHO I AM block (:258) — roughly 700 words of personal profile, pasted at the top of every book conversation. Its rule 4 is “PERSONALISE, NEVER GENERALISE.”

Evidence. evidence/reading-with-ai.md:412:

Personalisation and memory increase sycophancy. Benchmarking across five frontier models found supplying memory profiles or interaction history raised agreement sycophancy up to +45% for one model.

And :577: models “progressively mirrored users’ viewpoints.”

Call — evidence wins, narrowly and precisely. P0 keeps the constraint set — time budget, country and language context, what a usable answer looks like. It drops the self-narrative: values, aspirations, the story of who you are. Those move to Stage 7 DECIDE, where you are choosing an action and the model is not being asked to judge you.

P0’s rule 5 — “BE DIRECT. Don’t soften a correction into a compliment” — is the half with evidence behind it and it stays, strengthened.

6 · P3 clarifies every chapter, on every book

Protocol. P3 runs unconditionally, and instructs: “LENGTH IS NOT A CONSTRAINT. If clarity needs more words than the original chapter, use more.”

Evidence. evidence/reading-with-ai.md:252:

Practical rule: on unfamiliar material, let it explain freely. On familiar material, go back to the original instead.

That is the expertise reversal effect. The same file’s “worst uses” list names “summaries in place of reading”, and its §5.3 makes the sharper argument:

A self-help book is usually one idea plus the examples, repetition and specificity that make it actionable. Summaries preserve the idea and delete the machinery.

Call — evidence wins, conditionally. book.json carries familiarity per book, and each chapter carries clarified: true|false. When false, ## The clarified chapter is omitted from the page and READ I is the original text. The protocol’s own Rule 8 — “open the original for the passages that matter” — is upgraded from a hedge to a branch.

7 · Reread the recap, against the finding that rereading loses to retrieval

Protocol. :1010, Rule 7: “Reread the previous recap before every new chapter. Thirty seconds.”

Evidence. Rereading is rated low utility by Dunlosky et al.; retrieval practice is rated high. The protocol’s own appendix cites Roediger & Karpicke — the retrieval finding — and then concludes “this is why the recap card is reread before every chapter,” which inverts what the study found. That is an internal misreading of its own citation, not merely a conflict.

Call — evidence wins. The previous chapter’s recap is answered from memory first, then opened. /review serves the question and withholds the answer until you have committed to one. Thirty seconds either way; only one of them is practice.

8 · A four-week ceiling against a thirty-day gap

Protocol. :1006, Rule 3: “Four-week ceiling. Beyond that, it’s the wrong book.” Combined with Rule 2 (one book at a time) and Stage 0’s “When: between books”, the next book starts immediately.

Evidence. evidence/reading-with-ai.md:489, Stage 9: “30 days minimum before the next book.” evidence/reading.md:568, step 19: “Only then start the next book” — after the week-6 keep/adapt/drop decision, so six weeks at least.

Call — evidence wins. A thirty-day gap is the default. The book page carries nextEligible, computed from finished. It is a default, not a lock: you can start earlier and the site will say you did.

9 · Throughput arithmetic against volume as the failure mode

Protocol. :965-:967 computes “~3.75 weeks per book → 12–15 books a year” and “The 73 🔴 CORE books become a five-year path.” A whole Part is given to this arithmetic.

Evidence. evidence/reading.md:491 is a section titled “Reading the next book instead of doing the last one”:

The genre’s central risk is that consuming it feels like doing it, and this is the most insidious failure mode because it is indistinguishable from productive effort while it is happening.

Call — evidence wins, and the protocol already agrees with itself at :994 and :1032: “This is the metric. Not books read. Not chapters published.” · “The ledger is the metric. The count is vanity.”

So: no book counter, no chapter counter, no word count, anywhere in the site chrome. /shelf counts running actions — rows on the ledger still marked day30: running. The arithmetic stays in protocol.md as a planning estimate and never becomes a display.

10 · Mental contrasting is absent, and it is the strongest negative finding in the corpus

Protocol. P6’s STEP 3 attacks the draft commitment on five criteria, none of which is the inner obstacle. No stage forbids visualising the outcome.

Evidence. evidence/reading.md:569, step 12: “Do NOT visualise success.” — and :479:

It is one of the more robust findings in motivation research — and it directly contradicts a technique promoted by a large fraction of the self-help genre. Any book recommending that you visualise success as already accomplished is recommending something the evidence contradicts.

evidence/reading-with-ai.md:488, Stage 8: “Ask me for the inner obstacle. Refuse a vague one.”

Call — evidence wins. obstacle: is a required field on any action carrying committed: true. /validate flags one under six words as probably vague. And nothing anywhere on this site ever asks you to picture the outcome — not a component, not a template, not a prompt.

11 · The knowledge-gap / execution-gap diagnosis is missing from Stage 0

Protocol. Stage 0 SELECT (:44) asks four questions. None is this one.

Evidence. evidence/reading.md:122 states it as the question that comes before everything:

“Is my problem a knowledge gap, or an execution gap?”

and :127: “This single distinction predicts most of the disappointment people report with the genre.” It is grounded in Marrs (1995) — :25 — which found the genre works for knowledge gaps and largely does not for execution gaps.

Call — evidence wins. gap: knowledge | execution is a required field in book.json. A book chosen for an execution gap gets its action stages weighted and its explanation layer trimmed: more reading will not fix a doing problem, and the machine should stop pretending otherwise.

12 · The accountability the evidence wants, this site cannot provide

Protocol. The whole apparatus is a solo, private, gated website. No human appears in any of the twelve stages.

Evidence. evidence/reading.md:607: “(Plus one structural condition: someone who knows what you’re attempting.)” evidence/reading-with-ai.md:546, Guardrail 8: “Keep a human in the loop.” Harkin’s meta-analysis (N ≈ 20,000) finds d = 0.40, larger when progress is reported and physically recorded.

Call — neither wins, and that is the honest answer. A website cannot supply a person, and a private MDX ledger satisfies neither of Harkin’s moderators. This is stated as a limitation on /ledger in words, and never simulated. No streak, no badge, no encouraging copy. The machine does not get to pretend it is accountability.

If you want the effect the evidence describes, tell someone. That is outside this repository, and saying so is more useful than a progress bar.

13 · Rule 11 exempts all fiction, but three 🔴 CORE entries are novels

Internal contradiction, protocol against universe.md.

Protocol. :1014, Rule 11:

♾️ books are exempt from all of this. Meditations, the Gita, the Tao Te Ching, the Dhammapada, and all fiction. No pipeline, no explanation layer, no actions, no page.

Universe. universe.md:382 🔴 S **The Course of Love** · :780 🔴 S **Siddhartha** · :790 🔴 L 🇳🇵 **Karnali Blues** — all novels, all CORE. And “73 🔴 CORE books cover all 16 domains” counts them. If all fiction gets no page, three CORE books and most of Domain 15 cannot exist on the site.

Call — a third book shape resolves it. kind: lifelong gives a book one page and a dated log: no chapter pipeline, no explanation layer, no actions, no rating. Rule 11’s intent — do not apply the completion instinct to these — is preserved. Its literal claim, that they get no page, is dropped, because a page that only holds what you noticed on each pass costs nothing and is the only place those notes could live.

14 · Three frontmatter shapes, none of which agree

Internal contradiction, protocol against itself.

:868 — the Part 4 spec: status: complete # generated | explained | complete
:387 — P2’s book frontmatter emits status: reading, a value absent from that enum
:745 — P8 introduces finished: and a book-level verdict:, present in no spec
P3’s chapter frontmatter omits author, domain, cluster, rating, verdict, actions, committed and read — all of which the Part 4 spec requires

stub was added to this list on 2026-08-31, and its absence is worth recording as a defect in its own right. scripts/new-chapters.mjs:346 has always written status: stub onto every scaffolded chapter and scripts/audit.mjs:159 has always recognised it — but this row listed seven values and md-spec.md listed five. The code was right and both contracts were stale, and this file is one of the ones a web-chat Project uploads, so the stale copy was the one being read. The cost is specific: audit.mjs reports an unrecognised status as “not a known value. It counts as not-complete everywhere, silently”, so a spec that under-declares teaches you to hand-write a value the tooling then quietly discounts.

Note the trade, because it is deliberate: status is not validated as an enum. Enforcing it would turn a typo in a pasted chapter into a build failure, which is the wrong trade for content that arrives from a session that cannot see this filesystem. The cost is that complete, Complete and done differ silently — so scripts/audit.mjs reports them, and that is the only thing that will.

Amendments — his second set of instructions, 2026-09-14

Not a fifteenth conflict between the protocol and the evidence. These are changes he made after reading a complete sample book, quoted in reader.md §10. Where one of them touches a row above, his instruction is operative — the same standing his 2026-08-31 answers already have over the drafts — and the row above stays as the record of what was decided before.

#	Touches	What changed	Where it landed
A	Row 2 · extraction uncapped	Two to ten actions per chapter. Commitment is unchanged: one `committed` + `tier: now` per book. “a conceptual chapter may support none” is withdrawn	`chapter-spine.md` §6c · `P6` · `<Actions />` · `audit.mjs`
B	Row 4 · P5 question-first	The whole exchange is recorded, not one pair. The question still comes first and the position second — that order is the audit and it is untouched. `### How the exchange went` is added between the position and the ending	`chapter-spine.md` §6 · `P5`
C	Row 5 · P0 trimmed to the constraint set	Clarity replaces brevity in the constraint set. `P0`’s “I read fast” is gone; the character sketch stays out, because the reason for removing it — personalisation raises sycophancy — is unchanged	`standards.md` §2b · `P0`
D	Row 6 · summaries delete the machinery	Confirmed, from a third direction. “not oversimplified to the point where meaning is lost. Keep the original meaning intact” is the same rule arrived at from the reader’s side. Nothing about `clarified` changes	`standards.md` §2b a
E	Row 12 · solo private site, stated as a limitation	He now plans to share progress publicly. The limitation is qualified, not removed: a public post is not a person who asks, its effect is untested, and nothing simulates it	`reader.md` §5 and §10f

One amendment touches no row, and is recorded because it reverses an earlier call made in this repository rather than in the protocol: ### Where the author was standing was told on 2026-08-31 not to become biography. At book level, biography is now owed — ## Author context (book-spec.md §6) — and the chapter sub-section keeps only what bears on its own chapter.

What this file does NOT reconcile

Stated so nobody assumes it was handled.

Open	Why it is left open
Books whose unit is not a chapter	The Psychology of Money (19 short chapters), Atlas of the Heart (87 emotions), A Pattern Language (253 patterns), Letters from a Stoic. `kind: reference` covers most of it; the 19-chapter case genuinely breaks the four-week ceiling and needs a per-book decision
96 🔵 REFERENCE books	`kind: reference` gives them a page shape, but nothing tests whether a problem→answer log is the right shape until one is actually used
Reading order ≠ sidebar order	P2 asks whether a chapter should be read out of order; `sidebar.order` is bound to `chapter`. No field carries a reading sequence yet
`verdict: skip` chapters	Whether a skipped chapter still gets a page is decided per book in `book.json` `antiChapters`, but the ergonomics are untested
The 12-chapter assumption	The protocol’s arithmetic hardcodes 12; `weight: L` is defined as 400+ pages, which is routinely 20–30 chapters

Provenance

Every line number above was read on 2026-08-31 against the copies in this directory, which are byte-identical to the originals in learn-anything-in-tech-system/book-context/ (md5 verified at copy time). If a file is edited, the citations move and this file needs re-anchoring — re-anchor them, never delete them.