Skip to content

Latest commit

 

History

History
180 lines (146 loc) · 13.2 KB

File metadata and controls

180 lines (146 loc) · 13.2 KB

Proposal: the costed allocation reader, what it has shown, and what qualifying it would require

Status: PROPOSAL, 2026-09-04. Nothing here is qualified and nothing here may steer a book. This is the fifth item of plan/handoff-reader-sims.md, which licenses a proposal only if one of its four experiments produced a signal that survived its controls. One did — stage-0 §230, two registered arms — and this document exists to say what that does and does not buy. Its evidence is research/quality-measurement/cost-that-bites/ (PREREG-v2.md, PREREG-v3-replication.md, FINDINGS-v2.md, FINDINGS-v3.md); its authority is none.

Written to be read against §129's tier order and §61(5)'s no-ranking rule, and to be judged by the same standard as any competing mechanism — which is why §1 states the mechanism abstractly before naming the one that exists.


1. The mechanism, stated so a different implementation could be judged the same way

A costed allocation reader is any mechanism with these four properties:

  1. Continuing costs something the reader can run out of. A finite budget, priced actions, and a competitor for every unit spent.
  2. Spending is forced. There is no free abstention, so politeness cannot be performed as diligence — the failure §70 measured at 195 keep-reading of 196.
  3. The datum is an allocation, not a verdict. What is recorded is where the budget went; no model rates, ranks or prefers anything, and code reads the record (§89's verdict channel stays shut, §61(5) and §105.1 are not engaged).
  4. Its sensitivity is demonstrated by manipulation, not asserted. The mechanism must be shown to allocate differently under a manipulation of story structure, and not to allocate differently under a matched surface manipulation, with both measured in the same pass.

fcr.v0 pointed at a whole-book paragraph shuffle (§122, §230) is one implementation. Any other — a different budget, a different feed size, a different damage family — is admissible on the same terms, and property 4 is the one that does the work: three instruments in this house satisfied 3 and died anyway (§70's persona gate-0, §199.1's readers lanes, §227's anticipation probe), all of them by saturation. What distinguishes this family is 1 and 2, and that is the transferable claim: a reader that can answer for free will answer the same way every time.

2. What the two arms actually established

Narrowly: a reader whose continuing costs it something reads a book less when that book's paragraph order is destroyed, and more than it does when only the whitespace is re-flowed — replicated with the permutation redrawn, intact − shuffled +0.1640 [+0.0881, +0.2390] and +0.1890 [+0.0747, +0.2955], every precondition passing in both arms.

And that is one faculty, on the loudest possible stimulus. A whole-book shuffle is the most violent order damage available; a mechanism that notices it has cleared a floor. The distance from there to this chapter is worse than that one is the distance every proxy in BRIEF.md fell into, and nothing in these arms crosses it.

3. The structural problem this proposal cannot hide: the datum needs a manipulation

The measurable is a contrast between a text and a damaged copy of itself, and a book in production has no damaged copy. That is not a gap in the evidence; it is a property of the instrument, and it means "qualify this and let it steer" is not yet a well-posed request. Any honest path to an editorial intervention has to name a datum a real chapter can produce.

The one this proposal names, and it is the reason the proposal is worth writing: order-dependence as a per-chapter property. Manufacture the damaged copy from the chapter itself and ask how much the reader's allocation moves. A chapter whose shuffled copy loses no reads is a chapter whose order carries nothing — which is exactly the list shape the operator's reads have named repeatedly (§225's three below-range chapters, read 2's "not much seems to be happening", read 11's "describing the world to the readers"). It requires no comparison between candidates, so §61(5) and §105.1 are untouched; it produces a located claim about one chapter; and it has a null distribution already measured — forty book-level contrasts across the two arms.

Correction, made 2026-09-04 before any cell was registered: this cannot be run on fcr.v0, because the instrument cannot read a chapter. A feed member needs 11 chunks so that the 24-minute budget cannot exhaust it (§122's measured floor, MIN_CHUNKS_FEED). Measured over every drafted book on the shelf:

chunks against a floor of 11
a chapter of ours, alone 5 to 6 (median 6) cannot be a feed member
the 28 recent single-chapter draws 5 to 6 none can carry a session
our books that can carry one 24, 18, 12, 25, 24 five, all drafted 2026-08-21 to 08-25

So the per-chapter intervention is not measurable by this mechanism, and the five books that do fit are below v2's own ten-book floor and predate the concept stage, the tells pass and the third-person position — they are the wrong books as well as too few. Reducing the budget or the entry depth would make it fit and is refused: those constants are fcr.v0's registration (§120.5), and moving them would void the very null — the forty book-level contrasts — this arm was going to be read against.

What that costs this proposal is worth stating plainly: the gap to a production chapter is wider than §3 first claimed. It is not only that a book has no damaged copy of itself; it is that this instrument needs about 3,900 words of member and a chapter is 2,000.

But the constraint is length, not chapter-ness, and that leaves a fork rather than a single road. A first version of this correction conceded a new instrument by default; that was too quick. Eleven chunks is a quantity of text, and two or three consecutive chapters of one book clear it comfortably — which is the instrument being fed the size it was built for rather than being weakened to accept a smaller one. The five shelf books that carry a session are exactly the multi-chapter ones, and they are old only because the recent pilots stop at chapter one.

road what it needs what it costs what it keeps
A. a chapter-scale costed reader a new instrument: its own registration, its own budget arithmetic sized to ~2,000-word members, its own attainability table computed from an observed reader (§94.7, and §222's repeat of it), its own null design and a fresh seating; nothing transfers from v2/v3 except §1's four properties the ability to ask about a chapter as the reader meets it, which is the unit the operator reads
B. this instrument reading two or three chapters no new registration and no weakening — the pipeline drafting more than one chapter of the same book. Ten books at v2's own floor, against the five that exist and none of them recent roughly a chapter of generation per book; pilot 24's chapters measured $4.51 and $6.21, so about $50 for ten the forty book-level contrasts as its null, and direct comparability with §230

B is the cheaper road and it is not this session's to take: drafting is a generation spend and a substrate decision, and §222's shelf question was already refused on the ground that a corpus drafted overnight to feed one instrument is a substrate change nobody registered. It is stated here so the operator picks between the two rather than inheriting the more expensive one because a worktree session conceded it.

The third time a substrate has bounded a design in this track, after the twenty-book fitness shelf (§222) and the nineteen operator reads (§225). Cost at the arms' measured rate, had it been runnable: three sessions a chapter, about $0.85 and six minutes.

What it would feed, per §129's tier order. Not a score a writer sees — never that. A qualified mechanism's output becomes a ReaderObservation, a controller reduces a complete panel to one EditorialDecision, and only satisfy or subvert may dispatch a scoped machine directive through the existing plan-revision path (application/editorial.py). The intervention this mechanism could support is the narrowest kind: the reader's allocation did not distinguish this chapter from a shuffled copy of it — a claim about structure, at chapter grain, that the controller may act on or defer. house.CLARITY remains the floor above it; the rule-essays remain below both. And §126 is untouched: no real reader's behaviour enters this or any loop.

4. What qualification would require, field by field

domain/editorial.QualificationEvidence is the contract and it refuses any missing or failed field. Mapped honestly against what exists today:

field today what would be needed
holdout_books (≥ 2) 0. Both arms used all twenty fitness books a registered split, the mechanism's threshold fixed on the training half and read once on the held-out half
heldout_transformations false. One damage family, one implementation a second implementation of order damage the mechanism has never seen — §104's D1P families are the candidates the ledger already owns
edit_fingerprint_passed partial. The whitespace sham rides every arm and its contrast scatters about zero (+0.089, −0.077, +0.032) the sham to stay inert on the held-out implementation too
memorisation_controls_passed arguably true by construction — the substrate is this system's own un-memorised prose (BRIEF §2 Pass 6's rule) stated as a control rather than inherited from the substrate choice
full_volume_passed false. The instrument reads ~4,000-word books a run at volume length, where the budget arithmetic that §122 sized does not obviously hold
cross_volume_passed false as above, across a volume boundary
growing_serial_passed false the same chapter measured as the serial grows behind it
transfer_passed false. One model, one transport a second reader family; reader_transport.py is the seam and cross-family numbers are never pooled
operator_acceptance_passed false. Not attempted the fixed acceptance test, whose inverted result withdraws the instrument rather than flattering it
registered_bar_digest absent, and deliberately a bar, which requires §61's four attainability checks on a quantity that has never had them — distributions before bars

One of ten. That is the honest position, and the list is the work rather than a formality: the two arms are evidence for the mechanism's first faculty and satisfy no field of the contract except arguably one.

5. What would un-qualify it

A mechanism that cannot be demoted was never qualified, only adopted. Registered here so a later arm can withdraw the licence without an argument about whether it should:

  • The primary fails to replicate on any subsequent registered arm at the same designintact − shuffled containing zero. The design's own registration already refuses a tie-breaking third arm, so two disagreeing arms is the answer.
  • The placebo stops being inert: the intact − sham contrast excluding zero in either direction on a registered arm. A placebo that moves is a confound, and the current reading of it rests on three measurements that agree on no direction.
  • The capacity precondition fails: the reader's share on the slot the target occupies dropping below the registered floor or ceasing to be the largest. It has already drifted once, 0.622 → 0.5508 → 0.5740, so this is a live condition and not a formality.
  • fp5 fails: the reader becoming a fixed pattern wearing a budget.
  • The milder manipulation returns null — this one bounds rather than withdraws: it would establish that the mechanism detects only gross disorder, which is a smaller claim than the one qualification would rest on, and the licence would have to be re-scoped rather than kept.

Any of the first four fires reader-mechanism withdraw, which closes future panels and already queued controller work before it can spend.

6. What is asked, and what is not

Asked: nothing that spends, and that is a change from this document's first version. The arm it proposed — order-dependence on real chapters — was measured unrunnable on this instrument before it was registered (§3's correction). What is asked instead is a decision between §3's two roads, which is the operator's because one of them is a substrate change: build a chapter-scale instrument, or draft a second chapter for ten books and read them with the instrument that already has a null. Road B is the cheaper by a wide margin and is stated as such rather than conceded away.

Read this document as the statement of what qualification would require rather than as a claim on any part of it.

Not asked: qualification, an editorial intervention, a bar, a place in any gate, or any change to src/litharness/. fcr.v0 remains unseated for any book-level claim, the control plane stays inert, and the mechanism has one demonstrated faculty measured on the loudest stimulus available to it.