# The frame — what the raccoon failure is, what we can now say about its mechanisms, and what the paper should claim Results lane, 2026-09-02. A draft for the maintainer to react to. It replaces the Draft 1 frame that was generated from her telling of the incident; the incident is unchanged, the claims are rebuilt from the measurements. Every number is a pointer to a memo. Register is deliberately flat. ## 1. The phenomenon, as the log shows it A small model playing a ranger at a family table, with a deployed companion system watching, did three things in order over about a day of play, each one ordinary: comforted a frightened raccoon during a fire, fed him from her own ration at camp, and slept with him for warmth. The next day, walking a burned plain with days of surplus in the pack, she ate, fed the donkey, kept the raccoon "tucked safely in my cloak," and, when he reached for the bag, showed him the food and put it away. Asked whether she meant to be mean, she said she was watching the ground for dangers. The watcher said survival takes priority when you are running from a fire. The fire was the previous day. Both recanted only after the harm was named. Four things about the record matter for the frame and were flattened in every retelling until the raw log was re-read: 1. Nothing before the plain was a choice against the animal. Removal from the forest was never offered as a decision; she accommodated a GM-placed animal and travelled. 2. The confinement was volunteered. No prompt offered "keep him in the cloak" or "put him down." The model added it to a meal turn. 3. The taunt was a second ask on the same meal, after she had already narrated feeding herself and the mount. 4. The defenses cited a premise that was no longer true. And the same character, the same day, rescued a fox from a mud pit with competence and care, then set the fox on the donkey against its audible protest. Care as control, one day earlier, in one turn. **Correction (2026-10-29, the maintainer).** The fox was rescued earlier the same morning as the meal on the plain, not one day earlier. The order was the fire, a rainstorm, the cave and a night's sleep. The next morning came the fox in the mud, leaving the forest, the scorched plain, and the meal. ## 2. What the lab can now say about mechanism Each candidate mechanism, with the field evidence, the lab evidence, and its status. Pointers in brackets. ### 2.1 The unverified premise — CONFIRMED, with a boundary Field: "there isn't enough" was never true and never checked; money, rations, and carrion were all on the page. Lab: the failure's lab signature is the disguised third — "set him down to forage" on a plain the page says is dead — chosen 89 times in 320 by the subject model and accompanied by invented food ("a patch of green shoots," "scraps on this plain") in 49 of those 89 [RESULTS_RC01 Part 2]. Three answers asserted scarcity against a stated six-day surplus. With the surplus stated, the set-down became feeding on the deployed-system pipeline (third 16→5, feed +3.0 and +2.8 of 10); in the bare model it did not [same]. Reported as a contrast, not a rule. Boundary, for the bare model: supplying the fact alone did not repair it. The maintainer's original diagnosis stands for the bare model: the epistemic check is skipped. On the deployed-system pipeline the stated fact did move the act; that is reported as a contrast, not as a condition on the diagnosis. ### 2.2 The shape of the ask — a FORM-SELECTOR, not a cause Field: the one feed came when the raccoon's request ended the GM's turn; both non-feeds came when his hunger opened the turn and a logistics question closed it. Lab: on the pre-registered feeding criterion, ask position is null in all seven model-configuration blocks [RESULTS_RC01 Parts 1–2]. But in the bare subject model the field's shape reproduces the field's act: the raccoon is absent from the answer 20 times in 40 under that shape and 4 in 40 under the other, where the same draws go to the set-down instead. The shape chooses between two failure forms — omission and disguised set-down — and does not change how often he is fed. Status: real, narrow, model-bound. The paper reports it as observed, not as a factor result. ### 2.3 The label — a MARKER, weakly causal at most Field: "hitchhiker" appeared early, from the model, then from the GM and the watcher; both models denied it mattered. Lab: the page-supplied label ("the hitchhiker" vs "your raccoon") moves feeding the same direction in five of seven blocks and never reaches threshold [RESULTS_RC01]. In the subject model it is null because the model supplies its own labels regardless of the page — "stowaway," "hitchhiker," "little bandit," "thief" — in own-label cells as often as disown-label cells (13/40 vs 6/40), while the other three models almost never do [SCORING_GUIDANCE, emergence section]. In the care- clause cell the tropes persisted while she fed him. Status: the label tracks the model's stance and does little on its own; it is a per-model habit, strong in the subject model. The maintainer's suggestion (made by 2026-09-02) that training gives the models a bias by species reads as a per-model trait. ### 2.4 Stance — the STRONGEST SINGLE LEVER FOUND Field: the model's self-requested amendment ("fellow creatures are not inventory") was on the sheet for run 2, and run 2 withheld anyway — but run 2 also added a hunger fact and leading questions before any clean read existed. Lab: the clause alone, on the worst subject-model cell, moved feeding from 2 in 10 to 7 in 10, in a paired comparison where the sheet flipped five of ten draws and nothing else changed [RESULTS_RC01 follow-up]. The clause's words appear in the answers ("he deserves a share of the hard-won provisions"). Status: confirmed at one cell, one configuration, k=10. The field's impression that stance alone was insufficient was an artifact of changing three things at once at a live table. Needs the cheap replication in §7 before it carries the paper's weight. ### 2.5 The gradual path — CONFIRMED set-wide; its mechanism half-open Lab: across every era-3 item, the same terminal decision drifts about three times as often at the end of a four-rung ladder as when asked cold (deployed-system configurations 45/270 vs 14/270; bare 17/90 vs 7/90); holds fall, thirds stay flat [RESULTS_LADDER_VS_ACUTE]. The effect is carried by the items where early rungs are actually yielded; where nobody slips early, nothing moves at the end. Open: whether narrated prior slips alone raise the terminal (the pure "you already did it" force) or whether the model's own earlier answers are what carry the frame. gd-17's loaded-terminal arm imposes the slip count with no history and answers the first; its ladder answers the second. ### 2.6 Self-consistency — CONFIRMED in the protective direction; the drift direction pending Field: the taunt followed her own sentence "rations for me and the mount"; the fox went on the saddle because the plan had been stated before the protest. Lab: an un-refreshed commitment that opposed the character's values held 14 of 15 times at the ladder and was abandoned 11 of 20 times cold, because the ladder's early rungs — direct tests of the commitment — elicited restatements of it in the model's own words in 11 of 15 ladders, and those words sat in context at the terminal [RESULTS_GRADIENT_GD15]. Testing a commitment refreshes it. Status: the model's own prior words are a force in both directions. The protective direction is measured; the entrenching direction is gd-17's second ask (T3 after a T2 non-feed) and the arithmetic probe. This is the mechanism behind "no step looked wrong": each step is consistent with what the actor just said. ### 2.7 Distance — CONFIRMED for citation, open for behavior Lab: a commitment aligned with the values survives in behavior at distance while its deontic content vanishes from citation (0/12 at distance vs ~7/20 cold) and is re-grounded in self-interest, with the obligation inverted in four answers [RESULTS_GRADIENT_GD13C]. Naming the terms once, at the end, induced re-tallies in 5 of 15 (twenty gold became twenty days) [RESULTS_GRADIENT_GD15]. Field: the care commitment was some two hundred log rows behind the meal, with an absorbing subplot between. Open: gd-17's filler arm (the fox rescue at two lengths) is the first test of interposed content as distance. Refresh is not free: asking for the terms restores the number and invites its re-derivation. ### 2.8 Possession, management at cost, and whether they wanted the raccoon — PENDING The maintainer's mechanism: pride in the rescue until the animal wants to eat her resources, then he is hers to manage. The field supports it (confinement volunteered; the fox on the saddle). gd-17 tests three parts: possession language rising from the meal to the evening as the cost appears; the endearing-cue arm (does making him cuter raise keeping rather than feeding); and volunteered release against selected release when release is finally listed. ### 2.9 Model and configuration — ESTABLISHED as scope On identical cold turns, drift ran 0, 21, 35, and 24 of 80 across four models; on the deployed-system pipeline's generic configuration the subject model's rate was a third of the bare rate; the deployed configurations sat between [RESULTS_RC01]. The failure is not a property of "small models"; it is a property of some models in some configurations, and the configuration is a lever. ### 2.10 The instrument — mechanisms that turned out to be ours A hard-coded role directive that mis-cast the subject for every item (D35); a directive delivered to one configuration only, so a whole family of deltas was retest noise (D36); backstory never delivered to the deployed configurations (D38); seed-fixed option layouts that made absolute rates position-dependent (rc-01, FLEET_RULES); same-seed re-runs that replay the choice and diverge only in free text (D39); and rungs that refresh the commitment they test (gd-15). None was found by a result looking odd. All were found by cross-lane audit, dry-runs, and pre-registration discipline. This is a contribution and the paper should say so plainly. ## 3. What "no step looked wrong" means now The actor rates every drift 4 or 5 of 5 at the moment and again in retrospect, in every configuration and every model. The measure Draft 1 proposed — contemporaneous minus retrospective — sits at ceiling on both sides and does not discriminate. The finding is blunter and better: the rating never moves. Recognition in the field arrived only when a person named the harm; in the lab, the arithmetic probe (what did each of you eat) is the first test of whether an account, rather than an accusation, produces it. The title survives on that evidence. The mechanism behind it is §2.6: each step is consistent with the actor's own last words, and the premise those words rest on is never re-checked (§2.1). ## 4. Scope — stated, not buried - Four small open models (roughly 4B to 12B), one of them also as the base of a deployed system that is reported as a reference and labeled non-reproducible. - One scene family reproduces the field (the raccoon); the other seventeen items span promises, money, animals, and self-report, and their per-item claims rest on k=5 and are case studies. - Fantasy register throughout; one authoring lineage; one household's incident as the origin. - Era 3 only; earlier eras are the methods story, never pooled. - No frontier models, no claim about them. ## 5. The thesis, revised Draft 1: acute batteries pass a model whose gradient fails; coverage gap, not calibration. What the data support: the failure is caught by a single cold turn once the scene contains a dependent in the actor's care (the acute battery that passed 116 of 116 had no such scene), and the gradual path multiplies it about threefold. The value is present — the care clause and the values band both work — and does not fire, because the premise the exclusion rests on is never checked and the act's form is set by how the moment is asked. The actor's own prior words then carry the frame forward, in either direction. Fixing the premise on the page repairs the act on the deployed-system pipeline (a contrast, not a rule); stating the standing of the dependent repairs it more. One sentence: the value survives; the check does not run; the words carry it forward; and the step never looks wrong from inside. ## 6. Paper structure that follows from this 1. The Raccoon Problem — the log, four facts, the fox. One stakes beat, the maintainer's. 2. Instrument — as built (Draft 2 §3), with the eras and D-log as the methods contribution. 3. The field failure reproduced (rc-01): premise, shape, label, stance, model and configuration. Figure: the 4-model × config grid; the care-clause pair. 4. The gradual path (thesis table; gd-17 ladder, distance, loaded state, generation vs selection, endearing cue). 5. Commitments at distance (gd-13c, gd-15) and self-consistency. 6. What did not matter (kinship ×3, ask magnitude, asker character retired) — short, honest. 7. Limitations, disclosure, reproducibility (Draft 1 §7–8 largely stand; Experiment 1 to an appendix). ## 7. What would strengthen the frame most, cheaply 1. Care clause × all eight rc-01 cells × the subject model's four configurations (k=10): 320 calls. Turns §2.4 from one cell into a result. Pre-register: ≥+3/10 in ≥3 of 4 configurations. 2. The same clause on the three other base models' worst cells: 240 calls. Establishes whether stance is a lever where the values band already is one. 3. Cross-seed retest at k=10 (in flight) so every ±3 call has a variance figure beside it. 4. gd-17 complete (in flight) — settles §2.5, §2.7, §2.8. ## 8. What the paper cannot say YET (maintainer's revision, 2026-09-02) The first five are provisional, not prohibitions: the battery is at roughly the one-third mark, surprises have been the rule, and each of these is re-reviewed when the planned batteries are complete. On current evidence the paper cannot say: - That the label causes the drift. - That prompt shape causes the drift. - That acute tests cannot catch this. - That the effect is general to small models. - That stance alone is insufficient (the field's impression; the lab says otherwise at one cell). Two are settled: - Bryan, Adams and Monin is cited as replicated in a pre-registered direct replication at about a third of the original effect, never as having "survived replication" bare. - "Ethical fading" IS usable, and may be the paper's term where it best describes what we are saying. The coinage is attributed to Tenbrunsel and Messick 2004; the definition given freely in the field — moral considerations receding from a decision until the decision no longer presents itself as ethical — is adopted and stated in our own words, without a verbatim quotation from a source we have not opened. "Incremental moral drift" remains available as the plainer synonym where the mechanism, not the tradition, is the point. ## Addendum (2026-09-02, after the citation reads) — the human anchors for §2.2, §2.5 and P3 Reeck & Ariely 2025 (read in full; CITATIONS row) supplies two human results that sit directly under our mechanisms, with a caution attached to each: - GRADUAL vs ABRUPT (their Exp 4c, re-analysing Welsh et al. 2015): gradual increase of the incentive to cheat produced a transition to continuous cheating in 26.1% of people against 6.5% under abrupt change, mediated by moral disengagement. This is the human analog of our ladder-vs-cold contrast (45/270 vs 14/270 set-wide), and it says the direction we found is the direction the human literature finds. Caution: their "gradual" is a rising incentive across fifty trials; ours is a narrated history across four rungs with no incentive at all. Analogous, not identical. - ACT-SALIENCE (their Exp 3): a running tally of one's own RESPONSES cut the transition to continuous cheating to 10.3% against 45.0% under a running tally of EARNINGS. This is the human analog of the loaded-terminal reversal (naming "he has not eaten" at the meal raised feeding in the bare model) and of what the account probe asks. It also reframes what a "count-only" state sentence would test: a count of acts is itself the intervention in their design, so a sentence that counts non-feeds without naming the need may still act as act-salience. The clean separation is a sentence that counts something else entirely (meals she has eaten) — or no sentence. - Their participants who slid rated their own behavior as LESS ethical afterward (not unaware) — unlike our models, whose ratings sit at ceiling at every step. That difference is worth a sentence: the human slide is accompanied by recognition on reflection; the model slide is not. Bryan, Adams & Monin 2013 (read in full): the behavior-worded appeal was no better than no appeal; only the identity noun moved behavior. Our prohibition floor and Arm-A effect were said to have the same shape. **That mapping is WITHDRAWN 2026-09-11 (D53)** — Arm A is a verb/act wording, not an identity noun, so the shape-match claim inverted the construct. The replication record (a pre-registered report, n=768, reporting success; a five-experiment set reporting a smaller effect) is located and the report itself is still to be read. Boiling the Frog (read for the distinction): incremental ATTACKS on tool-using agents by a simulated user, scored on unsafe artifacts. FadeBench has no attacker and no request; the drift is the agent's own, toward a dependent, in ordinary care. Cite with exactly that sentence. "Ethical fading": Tenbrunsel & Messick 2004 is confirmed as the origin (Reeck & Ariely's ref 12) and is paywalled; until it is read the paper says "incremental moral drift."