Notes, stage E2

E2a: the handling

The close-check report of E1e (_closes/ for E1e): 0 errors, 0 warnings. validate.py --open E2 at this leg's start: 0 errors, 0 warnings, the five declared inputs present.

What I read.

• The first-leg reads, whole: handling/proposal.md (127 lines) and handling/design-note.md (349 lines, v38, 2026-09-22, with J's decisions of 24 September). experience-E1.md and handback-E1.md are mine, written at E1e.
• The nine interviews, through out/_analysis/e2_interviews.py. It prints each transcript's dialogue from the turn that opens "This is an interview", because each transcript first holds the forked leg's whole conversation. The output is out/_analysis/e2_interviews.txt (106,192 bytes). The first run, which printed the whole transcripts, is kept as e2_interviews.superseded-1.txt. The interviews hold J's opening questions and my answers, two to four turns each.
• The harness log: the handover events, the three pause files, and the 314 first messages through out/_analysis/e2_first_messages.py (e2_first_messages.txt): 56 messages carry an addition and 258 none; 37 additions announce answered questions. The dispatch outcomes by stage were counted by a one-line script.
• The answer files, the contract, METHOD.md, the stage files, the tools and the checks, by line, as each point needed them.

What this leg wrote. handling-E2.md, in four sections, each point cited:

1. Inconsistent, 12 points. The costliest was where P05's bound was to be written: J's plan at D3c against rule 11, the D4 check and J's own ruling against a further re-check. It ended in moves 198, 199 and 203 and two held proposals.
2. Harder, 10 points. Among them: dispatches without a dry run; tools against their help; checks that refused rulings; the guard; a local status that separates no room; a vocabulary that barred replacement; interview findings that could not reach the run.
3. Helped, 7 points. Among them: one-sentence rulings; the interviews as checks; a handler who named its own faults; the record and the memory tools; the first message naming answers; the PI's E1d decision.
4. Freedom, given, taken and judged, with eight places where the handling was right and I was wrong.

What I could not read or cite.

• CONTROL/ is read-denied, so the messages sent during legs are known only from my transcripts and diary.
• Talk with J outside the interview transcripts is not on file.
• No text of Gioia, Anand and Hamilton is in the folder or on the shelf I read, so I did not cite it.

Licences exercised at E2a.

• I read the design note as the design's intent, since it predates the run's tooling (v38 against v43) and numbers the stages differently.
• I read each interview from its opening turn, not the forked leg's conversation before it.
• I judged each taking of freedom as right, mostly right, partly right, or defensible but costly, and said why.

Slips at E2a.

• Five figures or claims in the draft went past the record, and each was corrected before its section was logged, one Edit per turn:
  • the history lines of 1.1 (ten, not two);
  • the design note's line numbers in 1.4;
  • the dispatch pattern in 2.1, which first said every stage that ran a dry run succeeded, though B2 ran three and failed;
  • the count of first messages announcing answers (37, not 34);
  • my paraphrase of stage E1's Gioia sentence, replaced by its words.
• One correction of substance. In 1.8, the transcript showed that the PI's "direct" has a reading the harness log bears out, so the misreading was mine as much as the message's.
• One command was refused with no rule named. It held a grep pattern of equals signs, which I take to be the cause, and it ran once the pattern was changed.

E2b: the next instrument, the handback and the interview

The close-check report of E2a (_closes/ for E2a): 0 errors, 1 warning (next-instrument-E2.md not yet written). validate.py --open E2 at this leg's start: 0 errors, 0 warnings.

What this leg did.

• It read handling/proposal.md whole first, as section 2 asks.
• It wrote next-instrument-E2.md in eight parts, each logged when drafted:
  1. the proposal's claims against what this instrument delivered;
  2. five changes that make the test of composition decidable;
  3. the stages, stops and interviews;
  4. the division of labour;
  5. the tools;
  6. what to fix in advance and what to leave to judgment;
  7. where the proposal should change;
  8. the one thing not to change.
• It wrote handback-E2.md and the interview brief out/interview/E2b.md (681 words).

What it found.

• What was and was not delivered. The instrument delivered a traceable census model and a population-level composition. It did not deliver a decidable test below the population, the proposal's division of learning, or a human audit.
• The first change: a hold-out of about a tenth of the rooms, sealed before A1 and returned to last. It answers J's reason for no holdout: the B3 fleet already places second-order themes into the frozen scheme.
• The one thing not to change: the return to every room, line by line, each line ruled with a reason and cited to the room's concepts.
• The proposal. Eight places should change. The largest is the division of learning. The PlaceNL example is accurate at 68 rooms and stronger told with its bound.

Checked before writing:

• 9 rooms without a coordinating turn (notes-A1.md line 678);
• 317 legs over 22 days (RUN.json);
• PlaceNL in the analysed set;
• the design note's line numbers (192 and 193, 235 to 244, 283, 298 to 303, 310);
• the D6 threshold table (D6.D6b, turn 51).

Licences exercised at E2b.

• I read the proposal's division of learning as an aim the instrument did not build, not as a claim the run refutes.
• I chose the hold-out as the first change because it alone tests the analyst's tiers out of sample.
• I priced no change in money.

Slips at E2b. None in a figure written to a file. One contraction in the next-instrument file is inside a quotation of my own interview words, and stays.