PLAIN_SUMMARY.md

paper/PLAIN_SUMMARY.md

@@ -0,0 +1,98 @@
+# PLAIN_SUMMARY.md — the paper in one page (for Will, 2026-07-06)
+
+**What the paper claims.** We took one production language model — GPT-2 — and built what the
+paper claims in its locked form: the first complete, certified, bidirectional decode of an entire
+production language model. We accounted for its ENTIRE internal state, at every one of
+its 13 layer checkpoints, in three kinds of text, against a pass bar the model itself sets. Not
+"here are some interesting features" — an account: throw away the model's whole hidden state at a
+checkpoint, rebuild it from only the things our decoder can read, and the model's behavior stays
+inside its own noise floor at 39 out of 39 boundary×regime cells — 13 checkpoints × 3 kinds of
+text (36/39 on the older, stricter ruler, and
+those 3 are priced ruler-geometry, not model behavior). The amount of behavior we could NOT
+explain went from 11.2 units of surprise to exactly zero, in public, dated steps.
+
+**And then we translated it — the BABEL codec** (your ratified name; the paper's long form, used
+once: "a certified bidirectional codec for GPT-2's residual-stream language"). Every one of the
+decoder's 351 channels was put on trial: 54% carry
+an explicit English meaning (a "naval/warship" field, a "clause boundary" word, an "operator"
+anchor); the other 46% are PROVEN to carry no word — proven, because random directions of the same
+loudness move the model more. How the language composes from layer to layer is certified linear at
+all 36 layer-seam tests. We built the exact inverse (English → state) and proved the round trip is
+behaviorally invisible at all 39 cells. And the model obeys hand edits: turn the "naval"
+field up and GPT-2 starts predicting *amphib, sunk, ashore, reefs, sailed, submarine*. Three of
+the four axes we hand-edited steer the model in their own vocabulary; the fourth (the "rung", an
+executable formula rather than a static dial) is now CERTIFIED UNUSABLE AS A STEERING LEVER: a
+same-day follow-up (L5) probed it through the channel matched to what it IS — the repetition
+behavior itself — plus two internal readouts, and its tiny effect cleared none of the internal
+readouts, while a genuine onset direction did move the behavior. A second follow-up (L6) tested
+it against an honest 20-draw random null at both doses: at ±6 it does not separate from a random
+nudge of the same size, and at ±3 its tiny (~0.4%-probability) effect is indistinguishable from
+the honest random floor itself (two pre-registered 20-draw nulls straddle it — one lands just
+below the effect, one above), growing SUB-linearly with dose — pushing harder buys nothing. A gauge
+you can read, not a lever you can pull, at both doses we tested. One new fact came out of the
+hunt: this bounds what can be written INTO the rung, not what reads out of it — one candidate
+certified channel (the naval/warship field, one layer later; a thin 2.8% margin over its
+multiplicity null) consumes the rung's output. Perfectly readable, not a steering handle.
+
+**Why anyone should believe it.** This is the paper's real weapon. Every experiment had its pass
+bar written down and locked BEFORE the data, with a stated bet; the completeness verdict came back
+NOT-YET six consecutive times and we published the gap table and stopped each time; the meter was
+recalibrated once, under pre-registered sanity gates, and both meters are reported everywhere,
+forever. The hardest object in the model — a repetition-keeping computation at layer 5 — got three
+certified impossibility results (can't be read, can't be looked up, can't be forged by its own
+circuit) before the thing that finally worked: we taught a tiny LINEAR student to compute it from
+readable inputs, and the test was built so that bigger students who merely memorize get caught —
+and they were caught (they ace training, fail on never-seen repeat periods; the linear one passes).
+One instrument bug happened all week; our own replay gate caught it. Every number in the paper
+traces to a frozen, hash-stamped artifact (Appendix A maps each one), and everything ran on the
+one A4500 in this machine — any skeptic can re-run any row.
+
+**What we do NOT claim.** Not the first "activations → English" concept — Anthropic's Natural
+Language Autoencoders (May 2026) and the Cycle-Consistent Activation Oracles published that idea
+first, and the paper says so generously. Our claim is the locked one — the first complete,
+certified, bidirectional decode of an entire production language model — on four axes they don't
+touch: all boundaries priced (they read middle-to-late-layer
+samples); falsifiable verdicts incl. certified-word-less channels (they score plausible glosses by
+reconstruction); a decoder built from the model's own certified channels with an algebraic inverse
+(theirs is a trained external translator); and a round trip scored in BEHAVIOR against
+matched-random nulls — edit the English, the model obeys (theirs is scored in activation space,
+with a qualitative steering demo on top and no matched-null certification — the paper credits the
+demo). Also honestly scoped: one model, one grain,
+three regimes; the 46% word-less fraction and the un-steerable rung are named, not hidden.
+
+**The three honest percentages** (the paper states all three, always together): 100% of the
+pre-registered definition met (certified-no-word counts as an answer); behavioral round trip
+100% / 94.7% / 3-of-4 (reconstruct / meaning-transplant / human-edit; the transplant number is
+16 prose pairs at one mid-stack checkpoint); 53.6% of channels carry an
+actual English word.
+
+**The two loose ends are now finished — as certified negatives (L5, same day).** Neither headline
+number moved, and both favorite bets lost. The missing 5.3% of transplantable meaning is NOT
+hiding in the certified door channels: adding the certified door read (summarized or in full)
+moves the model exactly 0% further —
+it is certified to live in genuinely un-charted dark mass outside the whole dictionary. It
+resisted every translation method we tried: it transfers only as its exact raw configuration,
+never through any compressed or named form. And editing the
+rung cleared none of the four channels we can read it through (the honest dose story is above).
+Both remainders are
+closed properties now, not open wounds; the honest claim is sharper, not bigger.
+
+**And the questions those answers opened are finished too (L6, same day).** We hunted the 5.3% to
+ground. It is DIFFUSE: smeared across the entire 329-dimension dark space — no hidden low-rank
+concept (we tried every rank up to 256; none captures 80% of the gap, and a RANDOM 256-dim slice
+of the dark does nearly as well as the best hand-picked one). And it is mostly word-less: of its
+8 biggest directions, only 2 faintly cross the naming bar, through side channels, with no stable
+meaning (they're recorded as provisional dark signatures, NOT dictionary entries). The rung stays
+steering-unusable at double dose (above). And the one L4 oddity we'd never explained — the "operator"
+axis appearing to steer on-manifold when it was predicted inert — dissolves: it was read-out ECHO
+(~94% of the response is the injected word-vector riding straight through to the output; the
+computed part is inside the noise), so no new capability, and the L4 result's own control always
+held. L6's scoreboard: 2 favorite bets hit, 3 lost — every loss logged as a certified finding.
+No headline number moved, in any of it.
+
+**Deliverables ready for your review:** PAPER_DRAFT_V1.md (+ 6 figures in paper_figs/, all
+regenerable from the frozen JSONs by one script; L5 verdicts in §6.6, L6 verdicts in §6.7, new
+Fig. 6 = the dark-mass rank ladder) and the outreach kit (AF post, 4 emails, Zenodo + GitHub
+checklists, release sequence, and the new GitHub-front-door README_DRAFT.md you asked for) in
+outreach_kit/, all updated post-L6 and standardized to the ratified name "the BABEL codec".
+Nothing has been sent, posted, uploaded, or committed anywhere — every send is yours to fire.