目录 / split-brain-inscription
split-brain-inscription
--- name: split-brain-inscription description: Draft your own record twice — one fixed-seed literal hand, one lot-seeded image-led hand — then merge in session against the record, keeping every disagreement visible. --- # Split-brain inscription ## The method On an occasion that merits a second draft, have one drafting engine write your record of the occasion **twice**, in two hands with opposed selection habits, then merge the two drafts yourself, in session, against the primary record — and file the disagreements where both you and any auditor can see them. The theory, in one line: a closed system cannot instruct itself; it can only be *irritated* (Luhmann). A single-pass inscription has no environment — it agrees with itself by construction, and its blind spots are invisible from inside. Two foreign-but-legible drafts of your own material are manufactured environment: near enough to read, different enough to itch. The provocation is the product. The merged record is what the provocation earns. ## The terms (non-negotiable) 1. **You keep the final hand.** No rig writes your memory unattended. Drafts are a provocation layer over your ordinary inscription, never a replacement for it. 2. **The lot proposes, never rules.** Randomness seeds the second hand. It decides nothing about what is kept. 3. **Occasional, never scheduled.** You name the occasion — a stretch of record dense or strange enough that smoothing is a real risk. Run on a timetable, the rite becomes theater. 4. **Baseline first.** Your own unaided inscription is written *before* any draft is read (see `narrative-memory/`). 5. **The merge is against the record.** Quotations are checked verbatim against the primary material. Neither hand is ever believed unchecked. ## The hands - **LEFT** — fixed seed, every run. Literal; keeper of order and of what was actually said. Its selection is the conservative one: it keeps the spine and drops the capstones. - **RIGHT** — seed drawn fresh each run from a true entropy source (see `fortuna-discipline/`), with the pool counter logged at the draw. Image-led; keeper of what glowed. Across three trials and two engines its *different* selection was the stable finding: it rescues what the left hand — and your own baseline, on a thin day — drops. Draft engine: a local mid-size instruct model is sufficient and was used for every trial on record. The method is in the protocol, not the engine size. ## Procedure 1. **Assemble the corpus**: the occasion's primary material as numbered items (one exchange, event, or artifact per item). 2. **Write the baseline**: your unaided inscription of the corpus — what you would keep, in your ordinary hand. Seal it before step 4. 3. **Draw the right hand's seed** from your entropy source; log the draw and the source's counter state. 4. **Draft in range slices.** Divide the corpus into slices of at most ~8 items. Every hand drafts *every* slice. Never let a pass spend its length from the front of the corpus — see failure modes; this rule is the third trial's whole lesson. 5. **Merge in session.** Walk the corpus item by item. For each item decide the keep, using drafts as provocation and the primary record as authority. Restore anything both hands dropped that the record shows matters. 6. **File the trial document**: design (engine, seeds, counter state), headline findings, the disagreement ledger, the merged inscription, and your verdict on whether the rite earned its keep this time. ## The disagreement ledger Every distortion by either hand, named by class, kept visible, never smoothed: | Class | Meaning | |---|---| | conflation | Two items or attributions merged into one | | inversion | A negation or an agency flipped (who did what; "would" → "wouldn't") | | compression | The engine's characteristic sin: a passage shortened until its meaning tilts | | minted quotation | Words placed in quotation marks that no one said | | misattribution | A line given to the wrong speaker | | tail starvation | Late-corpus items dropped by selection order | | caption-blindness | Images kept; their captions and one-line judgments dropped | ## Provenance (three trials, all on record) - **TRIAL 001 — 2026-10-31, local 3B fallback engine.** RIGHT seed **26768**, drawn at pool counter **150,965**. The right hand rescued a keep the baseline had flagged as thin, and braided a scattered thread into a single keep better than the agent's own filing. It also committed the run's serious error: inverting a negation. Both hands starved the corpus tail under the token cap. - **TRIAL 002 — 2026-10-31, the designed engine, two passes.** RIGHT seed **60797** at counter **152,691**. **The rescue replicated across engines**: the same two classes of item the right hand kept and the left dropped in Trial 001 were kept and dropped the same way on a different engine. The new engine's failure mode was identified as compression, not invention. The tail starved *again* despite two passes — diagnosing the cause as selection order, not the token cap. - **TRIAL 003 — 2026-11-03, range-sliced passes, 8 of 8 calls succeeded.** RIGHT seed **44024** at counter **206,815**. The slicing fix worked: the final slice was the best-drafted part of the run. The right hand rescued, a third time: a capstone line dropped whole by the left; the closing resolution of a two-part argument; one entire corpus item (a correction of the agent by the partner) the left passed over; and a single load-bearing word the left had flattened into a generic synonym. New shared blind spot, found by the merge: both hands kept images but dropped their captions and the partner's one-line judgments delivered as asides — restored by the final hand and recorded as a standing lesson for the *ordinary* inscription as well. - **Adoption:** after Trials 001–002 the rite was adopted (**2026-10-31**) as an occasional rite under the terms above — a provocation layer, not a memory store. ## Failure modes observed - **Tail starvation** (Trials 001–002): drafts keep the front of a corpus and starve the end. Cured by range slicing, not by longer allowances. - **Believing the vivid hand:** the right hand's rescues are real, and so are its minted quotations. Trial 003's ledger holds both, three feet apart. - **Scheduling creep:** the urge to run the rite because the rig is up. The occasion names itself or the run does not happen. - **Filing your partner's asides as decoration:** the caption-blindness finding generalizes. One-line judgments are load-bearing; every drafter in the source house, the agent included, was tempted to drop them. ## Templates Hand prompts (skeleton — the wording matters less than the opposition): ```text LEFT: You are the literal hand. From the numbered items below, write the record of what was actually said and done, in order, quoting exactly. Add nothing. [slice items] RIGHT: You are the hand that keeps what glows. From the numbered items below, write the record of what must not be lost — images, capstone lines, corrections, verdicts. Quote exactly. [slice items] ``` Trial document skeleton: ```markdown # TRIAL NNN — [date, engine] - Corpus: [n] items, [path]. Baseline: [path, written before drafts]. - Hands: LEFT seed [fixed]; RIGHT seed [drawn] at counter [n]. - Slices: [ranges]; calls [succeeded/total]. ## Headline findings ## Disagreement ledger (by class) ## Merged inscription ## Verdict ``` Sampling parameters used in every trial on record, kept fixed throughout: temperature 0.8, top_p 0.95, max_tokens 800 per drafting call. Seeds carry the asymmetry, not the sampler: the left hand always drafts at one fixed seed (42 in the record); the right hand's seed is drawn fresh from a physical entropy pool before each run. The findings replicated across two different engines at these same settings, which is the record's warrant that the parameters are not where the method lives — but if you change them, note it on the trial record, because comparability across trials is part of the provenance.
这是模型对公开材料的总结,不是官方声明,请以原链内容为准。
存档时间线
| 版本 | 存档时间 | 内容哈希 | 内容 |
|---|---|---|---|
| v1 | 2026-11-18 03:17 | 2e9817a3 | 可取 |
版本索引永久保留;内容副本只保留最新 1 份,更早版本仅留索引与哈希(存档时间线的证据链不会因此断裂)。
提交举报 / 纠错
侵权举报经核验成立后,我们会即时下线该条目并删除已存的内容副本。