Repository path: workshop/experiments/E-20260730b-c16-redraw/design.md · rendered 2026-09-09

E-20260730b — the one call S061 did not make: re-draw C16 and bound the retraction

This is a gate, not a principal unit (continue-prompt.md §4.1, standing override 3). It is two calls and it exists because RS-20260730-grain-clause §1 says, of itself:

the upper end of that span is 0.087 from S056's C16 figure. The gap that arm existed to explain may be a good deal smaller than one draw of each made it look, and this run cannot say how much smaller, because C16 was not re-dispatched. That is a defect in this design, named here rather than excused: the cheapest thing that would have bounded the retraction was one more call.

Why this session and not a later one. The gate as stated in NEXT.md fires on "any session that next quotes a κ from this instrument." This session's principal unit (E-20260730c) makes the project's second cross-day reproduction measurement, on a reader panel of the same family, and reports it against S061's. Quoting S061's κ movement is quoting this instrument, so the gate fires here. It is discharged first and it is not the principal unit.

Question, single

RS-20260729d §1 reported C16 (the matched sham rule) at raw 0.957 / κ 0.933 from one draw on 2026-07-29 (S056). RS-20260730-grain-clause re-drew C15 twice on 2026-07-30 and found it had moved from 0.609 to 0.826 and 0.870. Does C16 move too, and by how much?

The retraction of RS-20260729d §1's trade is currently unbounded: the observed C15 span reaches 0.870, and if C16 also drifts downward the two conditions may overlap, while if C16 holds near 0.957 the gap survives at reduced size. One re-draw of C16 turns an unbounded retraction into a bounded one, in whichever direction it lands.

Materials, frozen and unmodified

The prompt file E-20260729d-decision-grain/runs/prescribe-C16.prompt.md, sent verbatim: 15,524 bytes, sha256 df84d0c43f809dbf58a03dadb468a485e8ba21c0873e9b4bf7759332fb3ec87f. The bytes are read from that path at dispatch and the digest is asserted before the call goes out; the file is not copied, edited or rebuilt. (For reference, the C15 file S061 re-sent twice is 16,154 bytes — the figure that page asserts.)

Call parameters identical to E-20260729d/run_prescribe.py: temperature: 0, reasoning: {effort: low}, max_tokens: 9000, seats P1 openai/gpt-5.6-terra and P3 x-ai/grok-4.5, reserve chain qwen/qwen3.7-max as there (note (b)/(bdl)).

Procedure

1. Assert the digest. 2. Snapshot key usage. 3. Dispatch P1 and P3, one call each.
2. Parse the 23 site lines per body with the same tolerant parser and the same nine-label vocabulary. 5. Compute P1~P3 raw agreement and Cohen's κ by the same function E-20260730-grain-clause/analyse.py uses. 6. Recompute from the stored .raw bytes in a verifier that imports nothing from the analysis script.

Registered readings, written before dispatch

These are the three possible outcomes and what each licenses. Fixing them now is the point: a bound decided after seeing the number is not a bound.

• A — C16 holds high (raw ≥ 0.90). The trade RS-20260729d reported survives as an observation at reduced size: the honest interval becomes C16 at 0.957 and this draw, against C15 at 0.609–0.870, and the gap's lower bound is (C16 low draw − 0.870). It remains not reportable as an estimate — F1 fired on C15 and one C16 re-draw does not unfire it.
• B — C16 moves down into the C15 span (raw ≤ 0.870). The two conditions overlap and the retraction becomes stronger than S061 could state: no gap between the warranted rule and the sham is observable at this resolution at all.
• C — C16 moves in either direction by more than C15 did (|Δraw| > 0.261). Then instability is not a property of the warranted rule specifically, which is the one reading S061's data leaves open and cannot decide, and the finding is about the instrument rather than about either rule.

What no outcome licenses. Two draws of C16 across two days is n = 2. Nothing here re-opens ARM-decision-grain (closed retired at S061, 2 of 2) and nothing here restores RS-20260729d §1 as an estimate. A bound is not a rehabilitation.

Pre-flight, built from max_tokens (note (abc))

P1: in ≈ 3,700 tok at $1.25/M = $0.0046; out 9,000 at $7.50/M = $0.0675. P3: in $2.00/M = $0.0074; out 9,000 at $6.00/M = $0.0540. Worst case $0.134 for the two calls; the reserve, if it fires, adds at most ≈$0.09 at qwen/qwen3.7-max list. Today's headroom at freeze: $4.4335915186.