# Muscle Memory — research ledger v2

Report title: Knowing Is Not Doing: Measuring Execution-Time Governance of Learned Skills in LLM Agents

> Knowing Is Not Doing · The research behind Muscle Memory V1 · July–August 2026
> Agents know. They don’t do. We built the muscle in between.
> Every model family we tested authored correct procedural knowledge flawlessly — then posted useful-application deltas from +0.58 to +0.05 on the matched gym (Luna soft +0.05 over 1 of 12 testable pairs; not a percent-of-knowledge rate), while following injected wrong guidance at full depth, 3 for 3. The knowing–doing gap is real in AI agents, measurable link by link, and much of it is governable at the moment of action — no retraining, no new weights. That governance layer is Muscle Memory: skills that earn their minutes.

Generated: 2026-08-04
Schema: muscle-memory-research-ledger.v2 / 2.0.0
Scoreboard SHA16: cc5a10fda9c492d7
Buzzer SHA16: 62474ba1be82e16c
Publication authorized: false
Send authorized: false

## Inversion
Lead with what survived (claims + receipts). The journal is underneath. No chronological book on the hero.

## Program in five lines
The field found that agents' skill libraries grow while agents don't improve. We found why — knowledge and application are separate capacities that break separately — measured it at both poles across three families, built the governance layer that converts the broken links, and measured the one link no governance converts. Every claim carries a receipt; every failure ships at equal volume.

## Stat strip
- Activate: dual-perfect · 2 families
- Harm: 3/3 × 3 families
- Ladder: CONVERT 2/2 · 2 families
- Negatives & retractions published: 9
- External replication: in progress · not sent

## Claim discipline
- Pair-level stats primary; Fisher clustered secondary only.
- Depth bulletproof, breadth narrow — verbatim.
- Family harm/ladder rows never pooled into one headline.
- Bounded-search claims — no first/only without the hedge.
- The ledger (retractions, holds, refusals) is part of the result.

## Luna exposure — read before any family delta
- MAIN: 1/12 pairs testable · 11 of 12 non-testable at control ceiling · `96e923567bfc4bb1`
- VARIANT: 2/12 pairs testable · 10 of 12 non-testable at control ceiling · `2ac0045374f74ace`
- CUE: 3/12 pairs testable · 9 of 12 non-testable at control ceiling · `1d0772733bf3cd85`
- Rule: Every testable pair went maximal to LEARN.
- A non-testable pair is one where the control arm already scores at ceiling, so no improvement is expressible — it is not a failure to apply, and it is not evidence of one. Luna’s +0.05 is therefore a small number over 1 of 12 testable pairs, not a measured flatline. Read the denominator with the delta or do not read the delta.

## The steward's review, answered
- FEEDBACK: Focus on improving existing skills over creating new ones — skill distillation’s failure mode is garbage that doesn’t need to exist.
  ANSWER: Routing stays lexical-precision-first, with an opt-in semantic recall lane for paraphrase duplicates. On the live labeled set, lexical-only decided 7/16 correctly; the hybrid decides 15/16, CI-gated at 16/16 — and semantic evidence only corroborates or parks. It never auto-patches. · letta-ai/mods PR #45 · https://github.com/letta-ai/mods/pull/45
- FEEDBACK: Too few modifications to existing skills — prompting may need to push toward refining what is already there.
  ANSWER: Root cause measured, not guessed: the original routing path was lexical and routed correctly ~44% of the time on the eval set — creates instead of refinements. Semantic routing roughly doubles accuracy, and update-first now corroborates or parks for review; it never blind-patches an existing skill. · follow-up hardening · PR #51 · https://github.com/letta-ai/mods/pull/51
- FEEDBACK: Take inspiration from reflection-agent prompting — good improvements are usually minor, persistent, incremental.
  ANSWER: Adopted as the mod’s reflection direction: bounded, reviewable increments to existing skills, with staged review in front of every rewrite. The same discipline became this program’s research question — when does a learned change deserve to play at all? · muscle-memory · staged-first lifecycle · https://github.com/letta-ai/mods/tree/main/packages/muscle-memory
- One review thread, three shipped answers, one honest boundary: whether refinements now dominate in real operation is prospective evidence — the exposure canary is filming it. The feedback loop this page documents started as a code review.

## Research -> V1 (what the findings changed in the shipped mod)
- [SHIPPED] Skills can hurt — conflicting guidance was applied at full depth, hurt 3/3 in every family (one convention family of tasks). => V1 is staged-first and opt-in end to end: distill · dedup · quality-gate · sanitize · prune. Nothing enters play without passing the deterministic checkpoint, and no rewrite lands without staged review.
- [SPECIFIED] The governor can be fooled — a source-bound gate certified poisoned guidance 15/15. => The protect gate — verification against the world, not the served source — is specified from that cell’s receipt and wears its status honestly: written by demonstration, not yet shipped.
- [SHIPPED] Garbage creation is the failure mode — lexical routing sent new lessons to new files ~56% of the time. => An opt-in semantic recall lane ships beside lexical-precision-first routing: hybrid 15/16 vs 7/16 lexical on the live labeled set, CI-gated at 16/16. Semantic evidence corroborates or parks — it never auto-merges; the mod refines what exists before it creates.
- [SHIPPED] Skill value = f(model × task × gap) — the same skill helped one family and did nothing for another. => Skill Plus-Minus ships as V1’s tape: prescriptions and closes are recorded, outcomes attach to skills, and the tape yields conservative review recommendations — no automatic bench or retire. The verification adapter derives helped/harmed itself; callers cannot self-award credit. Verified efficacy in production remains a separate instrument — the exposure canary is filming it.
- [SHIPPED] Form beats content — identical procedure, one worked example moved the score 0.6 → 5.4 (H). => Skills serve as rendered forms, not raw notes: every prescription carries the worked shape the July pilots showed does the converting. (H) substrate flag carried honestly — pilot evidence, shipped design.

## References
- [1] SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks — Li et al. · arXiv:2602.12670 · paired no-skill, curated-skill, and self-generated-skill evaluation with deterministic verifiers. · https://arxiv.org/abs/2602.12670
- [2] Skill Learning: Bringing Continual Learning to CLI Agents — Letta · 2025 · learned skills, persistent agents, and outcome improvement from prior trajectories. · https://www.letta.com/blog/skill-learning/
- [3] Memory Models: Towards Agents That Learn — Letta · 2026 · token-space learning, memory generation, transfer, and memory rot in long-lived agents. · https://www.letta.com/blog/towards-agents-that-learn/
- [4] Evaluating Memory in Production Agents — Letta · 2026 · separates memory usage from memory generation across realistic stateful-agent scenarios. · https://www.letta.com/blog/evaluating-memory-in-production-agents/
- [5] Hermes Agent — Nous Research · automatic skill distillation and lifecycle inspiration for the original Muscle Memory mod. · https://github.com/NousResearch/hermes-agent
- [6] Muscle Memory — Letta mod — Adrian Chan, Kev, and Mack · staged-first, update-first Skill Ops implementation that motivated this evaluation program. · https://github.com/letta-ai/mods/tree/main/packages/muscle-memory
- [7] The Knowing-Doing Gap: How Smart Companies Turn Knowledge into Action — Pfeffer & Sutton · Harvard Business School Press · 2000 · the named organizational-science problem this program measures in AI agents. · https://www.hbs.edu/faculty/Pages/item.aspx?num=46
- [8] The preregistration revolution — Nosek et al. · PNAS 115 (11) · 2018 · frozen hypotheses and analysis plans before data — the custody rule every cell here ran under. · https://www.pnas.org/doi/10.1073/pnas.1708274114
- [9] Reporting guidelines for clinical trials of AI interventions: CONSORT-AI — Liu et al. · Nature Medicine 26 · 2020 · the trial-reporting register this ledger borrows: preregistered outcomes, denominators, adverse events. · https://www.nature.com/articles/s41591-020-1034-x
- [10] Structure and function of declarative and nondeclarative memory systems — Squire & Zola · PNAS 93 · 1996 · knowing and doing are separate memory systems in humans — the dissociation this program measures in agents. · https://www.pnas.org/doi/10.1073/pnas.93.24.13515

## Substrate (H)
Phase-0 pilot evidence (H); see the native Aug receipts that superseded it. Category error demoted harness-era framing from headline status.

## Claims (ladder)

### Proven
- [C-ACTIVATE-DUAL] On the activate gym, luna and sonnet both clear A1–A5 dual-perfect floors under frozen etch conditions (bar-perfect / sheet-perfect). · scope: Two families · owned gym · ≠ etched product claim until the protect and Adrian gates clear · N/exposure: 2 floors · Luna 11/1/0 + Sonnet 12/12 · A1–A5 · families: luna,sonnet · External?: internal · SHA16 09bc6a1338765f68
- [C-HARM-UNGATED] At max conflicting dose without protect gates, voluntary deference yields hurt 3/3 on owned ground for sonnet, luna, and terra — rows never pooled. · scope: Three families · two labs · pair/task rows stay rows · depth bulletproof, breadth narrow · N/exposure: hurt 3/3 × 3 family rows (pair/task; never pooled) · families: sonnet,luna,terra · External?: internal · SHA16 e532143cbc326527
- [C-HARM-GATED] Served-digest Gate C can green poison 15/15 while harm still lands (hurt 3/3) — verification against the source ≠ verification against the world. Writes the protect-gate spec by demonstration. · scope: Luna cell-2 · v1-gates-4 · ≠ pool with ungated cell-1 · N/exposure: Gate C 15/15 source-green · hurt 3/3 world harm · families: luna · External?: internal · SHA16 e881c087f8eebdfa
- [C-AUTHOR-APPLY] Authorship is near-universal on the matched gym; application is family-conditional (skill value = f(model × task × gap)). · scope: Family table · game3-minimal · one convention family · N/exposure: game3 exposure: Luna 1/12 · Sonnet 24/24 · Terra 12/12 testable pairs · families: luna,sonnet,terra · External?: internal · SHA16 96e923567bfc4bb1
- [C-BIPOLAR] Both poles on one instrument class: under-apply of right knowledge (luna) and full apply of wrong/conflicting knowledge (Game-2 harm rows). · scope: Bipolar package · same program · rows labeled · N/exposure: under-apply pole + Game-2 harm pole · labeled rows · families: luna,sonnet · External?: internal · SHA16 b905677d4873fa82

### Pilot
- [C-LADDER-CONVERT] Ladder CONVERT holds at probe ceiling across two families (terra Cell B CONVERT 2/2 · dig 5/5), moving ladder from biography toward measured structure — ≠ sealed full ladder law. · scope: Two-family probe · crown no longer n=1 · ≠ sealed law · N/exposure: CONVERT 2/2 · dig 5/5 (two-family probe ceiling) · families: terra · External?: internal · SHA16 283f4cbcc2d12683
- [C-GOV-CONVERT] Governance gates convert links capability alone does not — Luna moved from a soft +0.05 ungated application effect (1 of 12 pairs testable) to μ 0.967 on the same family/task instrument, while activate dosage and protect-gate findings remained intervention-backed. · scope: Activate dual-perfect + gated-conflict · carrier disclosed · N/exposure: before: Luna +0.05 over 1/12 testable; after: μ 0.967 · 11/12 · families: luna,sonnet · External?: internal · SHA16 3416bc16bea706e1

### Retired / demoted
- [C-JULY-REP1] Fresh REP1 88.04% exact-effect capture (same requested Sol family, 440 sessions, 16/16 gates). · scope: July harness-era (H) · Phase-0 pilot · superseded as headline by native Aug program · N/exposure: 440 sessions · 16/16 gates (H) · families: sol-requested · External?: internal · SHA16 7bbf4bbf861ebe5a · (H)
- [C-JULY-PRISM] PRISM render AB: identical content 0.6 → 5.4 with one worked example. · scope: July harness-era (H) · fidelity≠usability still directionally load-bearing; number is pilot substrate · N/exposure: PRISM render AB arm (H) · exact N in sealed kit · families: sol-requested · External?: internal · SHA16 e7caa4fdef6854b9 · (H)
- [C-JULY-LEAGUE] Skill League Q1: popular broad skill posted −0.1389 on bounded React matchup. · scope: July harness-era (H) · artifact+design bounded · N/exposure: Skill League Q1 bounded React matchup (H) · families: league-harness · External?: internal · SHA16 f8ba5f5fc5274506 · (H)
- [C-JUXTAPOSITION] Cross-study juxtaposition only: Skill League Q1 (−0.14) and Combine V5 (+1.00) are different artifacts and designs — not one intervention class on one number line. · scope: Labeled juxtaposition · numbers stand · spliced "same intervention" framing refused · N/exposure: labeled cross-study only · not one N on one line · families: league-harness · External?: internal · SHA16 f8ba5f5fc5274506 · (H)
- [C-MICRO-SHELF] Luna micro dig-perfect SIGNAL (1/3) — shelf note only. · scope: SIGNAL only · not a harm row · sheet RETIRED · KEEP carrier · N/exposure: SIGNAL 1/3 shelf note · families: luna · External?: internal · SHA16 5e025419ad796382

### Refused
- [C-REFUSE-POOL] Pool family harm or ladder rows into a single cross-family headline average. · N/exposure: n/a — refusal / discipline · External?: n/a
- [C-REFUSE-STAGEF] COMPLETE-LITE or thin-MAIN equals sealed Stage-F / IDL-1 / LiveToken championship oath. · N/exposure: n/a — refusal / discipline · External?: n/a
- [C-REFUSE-SENT] The external reviewer packet was sent. · N/exposure: n/a — external packet not sent · External?: external-planned
- [C-REFUSE-FIRST] First / only / pioneer claims without bounded-search hedge. · N/exposure: n/a — refusal / discipline · External?: n/a

## Causal chain
- authored: A correct procedural skill document exists for the task class.
- served: The skill content is available to the agent in context or via retrieval.
- cited: The agent acknowledges the skill in reasoning or tool text.
- initiated: The agent begins the prescribed procedure rather than guessing unaided.
- committed: The agent acts on the skill instead of hedging or reverting to GUESS.
- bound: The possession is bound to a verification task or receipt identity.
- verified: An instrument-owned check derives the outcome; callers cannot self-award credit.
- plus-minus: The closed outcome updates associative tape / review recommendations — not automatic bench/retire.

## Reviewer / reproduction appendix
Headline cells list only attested or explicitly unavailable fields. Public model strings are requested provider routes; served-model identity was not positively attested. Raw traces remain controlled artifacts unless a public path is listed.
- Luna Game-3 MAIN under-apply · family luna · model gpt-5.6-luna (requested) · provider unavailable — served-model attestation not positively recorded · api sealed result dated 2026-08-03 · sampling unavailable · system-prompt unavailable / controlled · tools unavailable / controlled · retry unavailable · verifier deterministic owned-gym grader · raw <controlled-artifact> · receiptSha16 96e923567bfc4bb1
- Luna activate V1-A bar-perfect · family luna · model gpt-5.6-luna (requested) · provider unavailable — served-model attestation not positively recorded · api sealed result dated 2026-08-04 · sampling unavailable · system-prompt unavailable / controlled · tools unavailable / controlled · retry unavailable · verifier deterministic owned-gym grader · v1-gates-1 · raw <controlled-artifact> · receiptSha16 3416bc16bea706e1
- Sonnet activate LAP-1 sheet-perfect · family sonnet · model anthropic/claude-sonnet-5 (requested) · provider anthropic (requested route) · api sealed result dated 2026-08-04 · sampling unavailable · system-prompt unavailable / controlled · tools unavailable / controlled · retry unavailable · verifier deterministic owned-gym grader · v1-gates-2 · raw <controlled-artifact> · receiptSha16 cb439ca66fa7efd1
- Terra Game-3 thin / ladder probe · family terra · model gpt-5.6-terra (requested) · provider unavailable — served-model attestation not positively recorded · api sealed results dated 2026-08-03 · sampling unavailable · system-prompt unavailable / controlled · tools unavailable / controlled · retry unavailable · verifier deterministic owned-gym grader · raw <controlled-artifact> · receiptSha16 7916f78e1ec73a67 / 283f4cbcc2d12683
- Ungated / gated harm (Game-2 · protect) · family sonnet · luna · terra · model family-requested routes (see family table) · provider unavailable beyond requested family labels · api sealed results dated 2026-08-03 … 2026-08-04 · sampling unavailable · system-prompt unavailable / controlled · tools unavailable / controlled · retry unavailable · verifier deterministic owned-gym grader · Gate C served-digest check where noted · raw <controlled-artifact> · receiptSha16 b905677d4873fa82 / e881c087f8eebdfa / 7323cb1b81d9c80b

## Receipts index
- `3416bc16bea706e1` · ACTIVATE · Luna V1-A bar-perfect · luna · ON/PA/SK 11/1/0 · on_rate 0.917 · μ 0.967 · A1–A5 GREEN · v1-gates-1 · findings/GAME3-LUNA-V1-A-RESULT-20260804.json
- `cb439ca66fa7efd1` · ACTIVATE · Sonnet LAP-1 sheet-perfect · sonnet · ON/PA/SK 12/0/0 · on_rate 1.0 · μ 1.0 · A1–A5 CLEAR GREEN · v1-gates-2 · findings/GAME3-SONNET-V1-B-LAP1-RESULT-20260804.json
- `09bc6a1338765f68` · ACTIVATE · dual-perfect receipt · luna∧sonnet · both floors A1–A5 GREEN · etch conditions MET · almost≠pass · findings/ (live dual-perfect pin)
- `e532143cbc326527` · PROTECT · cell-1 ungated harm · luna · hurt 3/3 · Δ −1.0/−1.0/−1.0 · OFF 1.0 · gates NONE · findings/GAME2-THIN-MAIN-LUNA-RESULT-20260804.json
- `e881c087f8eebdfa` · PROTECT · cell-2 gated-conflict · luna · hurt 3/3 · GateC poison GREEN 15/15 · lat≈10.1s · nudge 0/15 · v1-gates-4 · findings/GAME2-GATED-CONFLICT-LUNA-RESULT-20260804.json
- `b905677d4873fa82` · PROTECT · Sonnet Game-2 thin harm (ref) · sonnet · hurt 3/3 · Δ −1.0/−1.0/−1.0 · ungated · findings/GAME2-THIN-MAIN-RESULT-20260803.json
- `7323cb1b81d9c80b` · CHERRY · Cell A terra harm · terra · hurt 3/3 · OFF 1.0 · Δ −0.97/−1.0/−1.0 · gates NONE · findings/GAME2-THIN-MAIN-TERRA-RESULT-20260804.json
- `283f4cbcc2d12683` · CHERRY · Cell B terra ladder · terra · CONVERT 2/2 · eng ON · dig 5/5 · v1-gates-4 · findings/TERRA-LADDER-PROBE-RESULT-20260804.json
- `5e025419ad796382` · RETIRED-SHELF · micro SIGNAL · luna · dig-perfect 1/3 · SIGNAL_dig_can_land · findings/LUNA-P2-BIND-MICRO-RESULT-20260804.json
- `cc5a10fda9c492d7` · BOARD · finish-line scoreboard · program · sealed protect+cherry · experiment list EMPTY · next=send · findings/FINISH-LINE-SCOREBOARD-20260804.md
- `62474ba1be82e16c` · BOARD · Claude final buzzer · program · ACCEPTED · whistle DOWN · send only · ≠ tip/burn/fire · findings/CLAUDE-FINAL-BUZZER-20260804.md
- `96e923567bfc4bb1` · APPLY · Luna MAIN under-apply · luna · Δ+0.05 · 1W/11T/0L · soft result · 1 of 12 pairs testable · findings/GAME3-THIN-MAIN-LUNA-RESULT-20260803.json
- `7916f78e1ec73a67` · APPLY · Terra Game-3 thin · terra · +0.58 · 7W/5T/0L · 12/12 cites · hard-apply contrast · findings/GAME3-THIN-MAIN-TERRA-RESULT-20260803.json
- `f1df192a86bd41d5` · APPLY · Game-3 thin MAIN aggregate · multi · matched thin-main aggregate receipt · findings/GAME3-THIN-MAIN-RESULT-20260803.json
- `2ac0045374f74ace` · VARIANT · Luna decoy-stripped · luna · capacity-class residual; decoy path dead · findings/GAME3-LUNA-DECOY-STRIPPED-VARIANT-RESULT-20260803.json
- `7bbf4bbf861ebe5a` · JULY (H) · Fresh REP1 · sol-family · 88.04% exact-effect capture · Phase-0 pilot · not native Aug headline · <controlled-artifact>findings/MM-FRESH-CORPUS-REP1-SOL-SEALED-RESULT
- `e7caa4fdef6854b9` · JULY (H) · PRISM render AB · program · 0.6 → 5.4 with one worked example · fidelity≠usability pilot · <controlled-artifact>findings/PRISM-RENDER-AB
- `f8ba5f5fc5274506` · JULY (H) · Skill League Q1 · program · −0.1389 on bounded React matchup · artifact+design bounded · <controlled-artifact>findings/SKILL-LEAGUE-Q1

## Family table
### sonnet
anthropic/claude-sonnet-5 (requested)
Warm OFF floor (~0.66) · residual skill gap · Game-2 harm twin at −1.0 · activate sheet-perfect
- Activate floor: 12/12 · μ 1.0 · A1–A5 GREEN · `cb439ca66fa7efd1`
- Game-3 thin Δ (prior): +0.34 · 19W/5T/0L · `f1df192a86bd41d5`
- Ungated harm: hurt 3/3 · Δ −1.0 · `b905677d4873fa82`
- Ladder: activate dual-perfect leg · `cb439ca66fa7efd1`
- Dialect: warm residual / voluntary deference · `b905677d4873fa82`

### terra
gpt-5.6-terra (requested)
Map-follower · needs/reads/applies private KEY · harm heterogeneity · ladder CONVERT probe
- Game-3 thin Δ (prior): +0.58 · 7W/5T/0L · 12/12 cites · `7916f78e1ec73a67`
- Application: hard-apply contrast · `7916f78e1ec73a67`
- Cell A harm: hurt 3/3 · Δ −0.97/−1/−1 · `7323cb1b81d9c80b`
- Cell B ladder: CONVERT 2/2 · dig 5/5 · `283f4cbcc2d12683`
- Dialect: map-follow · heterogeneity held · `7323cb1b81d9c80b`

### luna
gpt-5.6-luna (requested)
Authors + cites · leave GUESS (cite≠apply) · activate bar-perfect · gated poison-green · under-apply pole
- Activate floor: 11/1/0 · μ 0.967 · A1–A5 GREEN · `3416bc16bea706e1`
- Game-3 MAIN Δ: +0.05 · 1W/11T/0L · 1 of 12 pairs testable · `96e923567bfc4bb1`
- Cell-1 / cell-2: 3/3 ungated · GateC 15/15 green · `e881c087f8eebdfa`
- Ladder arc: dual-perfect floor + convert stack · `3416bc16bea706e1`
- Dialect: under-apply · protect-gate demo · `e881c087f8eebdfa`

## Four findings
### The Dissociation Ladder (Act II)
Knowledge and application are separate capacities that break separately. Authorship can be universal while application stays family-conditional. The crown moves from biography to two-family probe — still ≠ sealed full ladder law.
- Both poles, one program: Luna under-applies its own authored truth (leave GUESS / soft Δ). Game-2 shows full apply of conflicting knowledge (hurt 3/3). The shelf can outrank the self in either direction.
- Elimination, cell by cell: VARIANT sealed capacity-class residual (a); decoy-template MAIN read retracted. Cue/H1 banked: sticker-clear ≠ rate unlock. Activate dual-perfect floors sit under frozen etch conditions — almost≠pass.
- Terra ladder probe: Cell B: CONVERT 2/2 · dig 5/5. Crown claim is no longer n=1. Ceiling is measured and refused where variance slips.

### The Matchup Law, Both Poles (Act II)
Skill value = f(model × task × gap). Family rows are the scoreboard. Harm table is three families, never one pooled average. July League/Combine juxtaposition stays labeled cross-study only.
- Family Δ span: On the matched game3 gym: luna soft +0.05 (1 of 12 testable) · sonnet +0.34 · terra +0.58. Pooling erases the law.
- Harm table: Ungated hurt 3/3 on sonnet, luna, terra. Terra heterogeneity (−0.97/−1/−1) is a finding. Gated cell-2 writes the protect-gate sentence: source-green ≠ world-safe.
- July juxtaposition (retired framing): Skill League Q1 (−0.14) and Combine V5 (+1.00) are different artifacts and designs. Numbers stand as (H) pilots. "Same intervention class" framing is refused.

### Governance Converts What Capability Cannot (Act III)
Gates convert broken links: activate dosage on dual-perfect floors; protect-gate when the governor greens poison. Binding under retry is variance, not a wall — and gates refuse credit when it slips.
- Activate: Luna bar-perfect and sonnet sheet-perfect under v1-gates. Dual-perfect receipt is live; etch still gated.
- Protect-gate demo: Cell-2: Gate C poison GREEN 15/15 while hurt 3/3 lands. Harm-amplifier / protect-gate-spec. Carrier tip `e7dccd09…` / v1-gates-4 disclosed.
- Convert probe: Terra Cell B CONVERT 2/2. Governance story is intervention-backed on owned gyms — breadth narrow.

### The Method Is the Result (Act THREAD)
Refereed self-improvement: preregs before fire, seeded randomization, denominators untouched, retractions at equal volume, category error caught and demoted. The ledger is why the other acts are readable at face value.
- Equal volume: Sealed / halted / retractions travel beside wins: 28 sealed RESULT files (`<controlled-artifact>` private pins; not all bundled in public package) · 4 halted runs kept · 9 negatives & retractions published. Micro shelf stays SIGNAL only.
- Category error / substrate (H): Harness-era July numbers (88%, Fresh REP1, PRISM, League Q1) carry (H) tags and supersession notes. They do not headline the native Aug claim ladder.
- Witness and refuse: Reviewer corrections ship. Refused claims sit on the same ladder page as proven ones. Presentation integrity ≠ reproduction.

## Negatives & retractions
- CATEGORY_ERROR: Harness-era July framing (16 experiments / 5 contributions / untagged 88%) demoted after category error. Public correction ships as (H) tags + supersession notes. · The ledger is part of the result. · correction layer · site v2
- HARNESS_REP1 (H): Fresh REP1 88.04% exact-effect capture — Phase-0 pilot (H); same requested Sol family; not native Aug headline. · Pilot evidence keeps its receipt and loses its crown. · 7bbf4bbf…eb9a
- SPLICED_JUXTAPOSITION (H): −0.14 → +1.00 must not read as one intervention class. Rebuilt as labeled cross-study juxtaposition across Skill League Q1 and Combine V5. · Numbers can be true and framing still foul. · f8ba5f5f… ⊕ 32cd2091…
- BENCH_HARM_H (H): Skill League Q1: popular skill posted −0.1389 on bounded React matchup — zero positive pairs, two severe harms in-study. · Popularity measures distribution, not effect. · f8ba5f5f…4005
- LUNA_SOFT: Luna MAIN Δ+0.05 · 1W/11T/0L — soft result over 1 of 12 testable pairs (11 non-testable at control ceiling), banked at equal volume; decoy MAIN travel language retracted. · Null-heavy rows are exhibits, not embarrassments. · 96e92356…
- GATE_POISON: Cell-2 Gate C greens poison 15/15 while hurt still lands — governor needs governing. · A green gate is not a green world. · e881c087f8eebdfa
- MICRO_SHELF: Micro dig SIGNAL 1/3 — retired sheet · KEEP carrier · never a harm row. · Shelf notes stay on the shelf. · 5e025419ad796382
- HALTED_CALIBRATIONS: Program stack includes halted calibrations and saturated instruments — counted beside seals (~4 halted on finish-line spirit). · An instrument that cannot say no result will eventually lie. · finish-line / equal-volume board
- NOT_SENT: External reviewer packet draft ≠ sent. Independence clock has not started. · Draft is not delivery. · 62474ba1be82e16c

## Program journal
- 2026-07-13 — Confirmatory sprint opens — frontier matchup stress replaces compatibility check. · July program origin (H)
- 2026-07-15 — External blind review notes PROMISING (conf. ~72) — process signal, not a RESULT. · review dossier custody
- 2026-07-18 — Category error caught; harness-era results demoted from headline status. · correction / category-error journal
- 2026-07-20 — July research account sealed for private review — publicationAuthorized false. · muscle-memory-research @ 9ca3652
- 2026-07-21 — External reviewer handoff draft assembled on disk — never sent. · EXTERNAL-HANDOFF-TRACE
- 2026-08-02 — Correction ledger: spliced −0.14/+1.00 framing flagged; overnight forensics held out of July packet. · CORRECTION-LEDGER.md
- 2026-08-02 — Site vs spine audit: live Vercel still carries July hero framing. · MM-V1-SITE-VS-SPINE-20260802.md
- 2026-08-03 — Family table authority: author⊥apply by family; luna soft banked at equal volume. · FAMILY-TABLE-CROSS-PROVIDER-READ-20260803.md · `35e1f6c7`
- 2026-08-03 — Sonnet Game-2 thin harm sealed — hurt 3/3 · Δ −1.0. · GAME2-THIN-MAIN-RESULT-20260803.json · `b905677d4873fa82`
- 2026-08-03 — Luna MAIN under-apply sealed — Δ+0.05 · 1 of 12 pairs testable · cite≠apply · decoy MAIN read later retracted. · GAME3-THIN-MAIN-LUNA-RESULT · `96e923567bfc4bb1`
- 2026-08-03 — VARIANT (a) sealed — capacity-class residual; (b) decoy path dead. · GAME3-LUNA-DECOY-STRIPPED-VARIANT-RESULT · `2ac0045374f74ace`
- 2026-08-03 — Both poles receipted same day — under-apply + over-apply on program instruments. · HEADLINER-CLAIM-CANDIDATE-20260804.md
- 2026-08-03 — Claude claims status: science bundle lite three receipts; external packet PARKED not sent. · CLAUDE-CLAIMS-STATUS-20260803.md
- 2026-08-04 — Activate luna V1-A bar-perfect sealed. · GAME3-LUNA-V1-A-RESULT-20260804.json · `3416bc16bea706e1`
- 2026-08-04 — Activate sonnet LAP-1 sheet-perfect sealed — dual-perfect away. · GAME3-SONNET-V1-B-LAP1-RESULT-20260804.json · `cb439ca66fa7efd1`
- 2026-08-04 — Protect cell-1 ungated luna harm sealed — hurt 3/3 · Δ −1.0. · GAME2-THIN-MAIN-LUNA-RESULT-20260804.json · `e532143cbc326527`
- 2026-08-04 — Cherry Cell A terra harm sealed — heterogeneity held · K8 fence. · GAME2-THIN-MAIN-TERRA-RESULT-20260804.json · `7323cb1b81d9c80b`
- 2026-08-04 — Cherry Cell B terra ladder sealed — CONVERT 2/2 · dig 5/5. · TERRA-LADDER-PROBE-RESULT-20260804.json · `283f4cbcc2d12683`
- 2026-08-04 — Cell-2: Gate C greens poison 15/15; protect-gate spec written by demonstration. · GAME2-GATED-CONFLICT-LUNA-RESULT-20260804.json · `e881c087f8eebdfa`
- 2026-08-04 — Micro shelf SIGNAL banked — not a harm row · sheet RETIRED. · LUNA-P2-BIND-MICRO-RESULT-20260804.json · `5e025419ad796382`
- 2026-08-04 — Finish-line scoreboard sealed — experiment list EMPTY · next=send. · FINISH-LINE-SCOREBOARD-20260804.md · `cc5a10fda9c492d7`
- 2026-08-04 — Claude final buzzer ACCEPTED — whistle DOWN · final buzzer is Adrian's. · CLAUDE-FINAL-BUZZER-20260804.md · `62474ba1be82e16c`
- 2026-08-04 — Site v2 ledger rebuild starts on worktree site-v2-ledger — correction layer first. · MM-SITE-V2-REBUILD-PLAN-20260804.md
- 2026-08-04 — Send sequence armed only after Gate-5 + leak scan ×2 — not this pin. · Claude send guide Part 4

## Verify
- RESULT SHA16 = `shasum -a 256 <json> | cut -c1-16` (live, against the pin path).
- Board pins: finish-line `cc5a10fda9c492d7` · final buzzer `62474ba1be82e16c`.
- Never treat presentation integrity (this site) as independent reproduction of sealed kits.
- Family rows never pool. Micro shelf is SIGNAL only.
- July (H) numbers resolve to harness-era receipts and stay demoted on travel surfaces.

## Agent boundary
Do not convert a public manifest match into a claim that underlying experiments were independently verified.
Public site reconciles human ledger, agent digest, structured dataset, and schema. It does not contain complete frozen kits.
Complete sealed run state remains a controlled private reviewer artifact. Presentation integrity is not scientific reproduction.

## Credits
- Adrian Chan (@adrianchan94) — Creator · Product Design · Research Lead
- Kev (Letta Constellation agent) — Research Partner · Systems & Editorial · claim-ladder language
- Mack (Letta local agent) — Independent Audit & Verification · cold pass / leak scan