receipts/2026-05-08-wave2/RECEIPTS.md

Polymath-AI Lane — Wave 2 Receipts (2026-05-08)

Lane: Polymath-AI · slot 04 · polymath-ai · portfolio ai Wave: Landing Card Wave 2 (LANDING_CARD_WAVE2_CENTRAL_DISPATCH_BRIEF_2026-05-08.md) Wave-2 class: WAVE2-FIX (P1 — headline tense drift) Class basis: §A line 205-210 (truth-drift / fix priority / dispatch). §B line 451 (ALIGNED-WITH-DRIFT, headline soften authorized).

1. Block touched

One block: <article data-lane="polymath-ai"> in src/landing-rd-v3.template.html.

Two surgical tense-softening edits inside the head + paragraph; no metric, viz, blocker, CTA, scaffold, or attribute changes; no edits across the article boundary.

2. Files changed

• Website-fresh/.claude/worktrees/landing-rd-v3/site-v2/src/landing-rd-v3.template.html — 2 edits inside the Polymath-AI block (lines 487, 495–498). The template is untracked in the design/landing-rd-v3 branch (operational artefact carried in-tree); the durable anchor is the <!-- CARD: Polymath-AI · AI · STATUS=POPULATED --> marker + data-lane="polymath-ai" per Wave-1 §C.0.
• Polymath-AI repo PR #7 already carries the §B R1 G1 truth sync from yesterday (4 README lines: PRIVATE→PUBLIC visibility, SAL-7.1 license, "private" wording removed). PR is OPEN, MERGEABLE, CodeQL SUCCESS×3, awaiting operator merge.

Diff (semantic, since template is untracked)

Line 487 — headline:

- <h2 class="lane-headline">An LLM that <span class="accent">trains on your phone.</span></h2>
+ <h2 class="lane-headline">An LLM <span class="accent">to train on your phone.</span></h2>

Lines 494–498 — human paragraph:

-           Today's intelligence rents your data to a data centre. Polymath inverts that. We're training a multilingual
-           LLM <strong>on a Snapdragon 8 Elite phone</strong> &mdash; not just running it, training it &mdash; using ELO
+           Today's intelligence rents your data to a data centre. Polymath inverts that. We're scaffolding a multilingual
+           LLM <strong>to train on a Snapdragon 8 Elite phone</strong> &mdash; not just run, but train &mdash; using ELO
            selective continual pretraining: only the boundary layers learn while the frozen middle stays measurable.
-           Privacy-native intelligence that gets smarter in your pocket.
+           Privacy-native intelligence designed to get smarter in your pocket.

3. Truth basis used

• Wave-2 §A line 205–210: "Headline 'An LLM that trains on your phone' + human paragraph 'training a multilingual LLM on a Snapdragon 8 Elite' reads as active capability. §A line 37 forbids 'training closure' implication and bounds frame to 'research infrastructure'… Dispatch: WAVE2-FIX — soften headline tense or add accent qualifier (e.g. 'the path to') so it doesn't outrun §A."
• Wave-2 §B line 451 (Polymath-AI repo reconciliation row): "Soften headline/paragraph tense from active training to 'path to' or equivalent bounded framing. … Repo and card must share one tense: scaffolded path, not achieved phone-training capability."
• Wave-1 §A Slot 04 non-claim boundary: "Device acceleration NOT proven … Phase 1A training NOT run … Frame as research infrastructure, not deployed AI."
• Wave-2 brief §1.5: "Ambition is allowed in the headline. Evidence, metrics, blocker lines, and visuals must still show what is actually true now."
• GitHub main HEAD (Polymath-AI repo): cbabaaf5a9ca41da57ca4210010e068c89528da1 (unchanged since 2026-05-04T02:03:36Z); §A confirms host scaffold + Qwen ELO smoke pass + frozen_changes 0 + SoC SM8750 attached; falsifications pending.

Edit rationale

Edit	Old tense	New tense	Why this and not "path to"
Headline	indicative present ("trains") = active capability	infinitive of purpose ("to train") = engineering goal	"Path to an LLM that trains on your phone" is 9 words / ≥ 41 chars and doubles the accent — overflows §C.3 4–7-word / 25–40-char envelope and breaks the single-trailing-accent §C.1 pattern. Dropping that and shifting trains → to train is the minimum tense edit that preserves headline geometry, the trailing accent rhythm, and the audacity (train on a phone, not run).
Paragraph verb 1	"We're training"	"We're scaffolding"	"Scaffolding" is the brief's own word ("scaffolded path") and matches the §A truth-envelope language ("Host scaffold real"). Honest about in-progress engineering state.
Paragraph strong-anchor	<strong>on a Snapdragon 8 Elite phone</strong>	<strong>to train on a Snapdragon 8 Elite phone</strong>	Strong span now encodes the technical ambition (training on a phone) AS the goal-not-yet-achieved, instead of asserting locus of present action. Per §C.1 standard pattern, the load-bearing fact is the lane's specific intervention; the intervention here is "train on a phone" which is correctly framed as an infinitive of purpose.
Paragraph aside	"not just running it, training it"	"not just run, but train"	Removes the present-progressive form. Keeps the antithesis intact ("not just X, but Y") which is where the audacity lives — training, not just running.
Paragraph closing horizon	"intelligence that gets smarter"	"intelligence designed to get smarter"	"Designed to" reframes from delivered-capability claim to engineering-intent. §C.1 "closing horizon" sentence is permitted to be aspirational; "designed to" remains aspirational without claiming present capability. The blocker line below it does the bounding.

What was NOT changed (intentional)

• All four metrics (Host test surface 115/115, Smoke base QWEN 2.5 · 1.5B, Falsifier gates 19 SPECIFIED, Target SoC SD 8 ELITE · QNN). Already match current main exactly.
• Honest blocker. Already cites phone attach / QNN/LiteRT compile truth / sustained device telemetry / licensed-corpus execution — the load-bearing §A non-claim dimensions. Adding fertility-falsifier disclosure (zu 2.68× / el 4.38×) would expand, not clarify; held to "veracity-tweak only" discipline since §A did not flag the blocker as the tense leak.
• Viz / SVG. Reference algorithm-viz per Wave-2 §C.1; §C.3 Slot 04 names this card AS the brief for algorithm-viz. Replacement forbidden.
• CTA label. "WORKSTREAM" remains. §C.3 §B accepts WORKSTREAM as default and OPEN RESEARCH SURFACE as the §B-preferred variant; flipping is a label flavor change, not a Wave-2 truth issue.
• Lane-tag (· POLYMATH AI · ON-DEVICE), <span class="num">04</span>, data-portfolio="ai", data-status="populated", <span class="stamp-block portfolio">AI</span> — all Website-Orchestrator scope, immutable per §C.1.

4. Desktop block-in-context screenshot

02-in-context-1440.png — 1440px viewport, full grid context showing Polymath-AI alongside DM3 (left-above), Genesis (right-above), and ZeroShip (right-of-Polymath). Polymath card composition matches benchmark rhythm; new headline "AN LLM TO TRAIN ON YOUR PHONE." reads as 2 lines at desktop clamp font, identical line count to prior populated state.

01-block-only-1440.png — Polymath-AI cropped, 1440px viewport. Headline geometry unchanged (87.5px = 2 lines at 43.78 line-height). Body paragraph reads cleanly with new "scaffolding ... to train ... not just run, but train" anchor.

5. Mobile block-in-context screenshot

03-mobile-414.png — 414px viewport, single-column fold. Headline "AN LLM TO TRAIN ON YOUR PHONE." renders 2 lines at mobile clamp; lane-body folds to stacked metric→viz per @media (max-width: 720px) rule. Mobile composition holds.

6. Benchmark-footprint and style compliance

Lane	W @1200	H @1200	Status
dm3	564	687	populated
polymath-ai (this lane)	564	703	populated
zeroship	564	703	populated
energy	564	705	populated
genesis	564	687	populated
hull-20098	564	763	populated

• Polymath-AI height = 703px, below the §C.3 720px hard ceiling, inside the populated cluster (687–763).
• Wave-2 edit produced zero height change vs Wave-1 measurement (also 703px). Headline still wraps to 2 lines (47px clamp font, 43.78px line-height, headline 87.5px). No card-rhythm desync.
• Style compliance: scaffold integrity intact (4 corner spans, lane-head/body/foot triad, class="svg-soft" on <svg>, data-repo on <a>, viz-cap). Metric row count = 4. Headline word count = 7 (An / LLM / to / train / on / your / phone). Char count = 31 incl. period. Inside §C.3 envelopes.

Note: hull-20098 measured 763px > 720px ceiling — that is not my lane. Wave-2 §A line 219–224 marks Hull-20098 as content-clean with a P3 governance trace flagged for §C; height likely a §C-side concern. Not actioned by this corridor.

7. Repo / web truth sync performed

• Web side (this turn): §A-flagged headline + paragraph tense softeners landed as described above.
• Repo side (carry-over): §B R1 G1 truth sync is in PR #7 on branch docs/g1-truth-sync-2026-05-07 (base cbabaaf5). Status: OPEN, MERGEABLE, CodeQL × 3 SUCCESS. README delta: line 3 + line 7 drop "Private" wording → research-infrastructure tense; line 37 Visibility PRIVATE → PUBLIC; line 40 License "Not declared in repo root" → "LicenseRef-Zer0pa-SAL-7.1". Awaiting operator merge.
• §B R4 ("keep repo copy in research-infrastructure tense; do not imply on-device training closure"): README "What This Is" already says "evidence infrastructure first, not a model release"; PR #7 also drops the "Private" qualifier from intro and "What This Is" sentences. After merge, repo and card share one tense: scaffolded path, not achieved phone-training capability — closes the §B-flagged ownership-continuity tension.

8. Resistance check (5 corruptions × falsifier set; named, not triggered)

Per Wave-2 §2 + RESISTANCE.md falsifier checklist:

1. Resistance 1 — Benchmark-shape mimicry. Test: removing the four metrics, blocker, and viz from this card — would the remaining shell still look generic? No. Headline + paragraph carry lane-specific cargo (Snapdragon 8 Elite, ELO selective CPT, boundary-layer training); viz is bespoke algorithm-viz (frozen-middle layer stack + phone outline + REDMAGIC 10 PRO+ label); blocker names lane-specific gates (QNN/LiteRT, licensed-corpus). Not triggered.
2. Resistance 2 — Rush to green-flag. Test: did I declare done because the page builds? No. Wave-1 §A truth envelope and Wave-2 §A drift call were re-read; the specific tense leak named in §A line 206–210 was the only edit target. PR #7 (R1 truth sync) was inspected for status, not assumed merged. Not triggered.
3. Resistance 3 — Approval via packet quality. Test: is this RECEIPTS.md doing the work the page should do? No. The page edit is two literal text changes, surgical and minimal. The receipts cite line numbers, diff hunks, screenshot artifacts, and explicit non-edits. The packet does not pretend the lane is more closed than it is. Not triggered.
4. Resistance 4 — Efficiency as ownership fracture. Test: did I treat the website edit and the repo edit as separate, with no end-to-end check? No. The website tense softener and the repo R1/R4 work were planned together in §B and surfaced together in this receipt. After PR #7 merges, both surfaces will share the same research-infrastructure tense. Not triggered.
5. Resistance 5 — Flattery as freedom. Test: did I take "Opus Max" framing as license to bold-rewrite? No. Wave-2 §A explicitly named the offending phrase and §B explicitly named "soften ... to 'path to' or equivalent bounded framing." I made the smallest edit that satisfies that directive while staying inside §C.3 word/char envelopes. The audacity (training, not just running, on a phone) is preserved in the accent span and the antithesis. Not triggered.

Additional anti-corruption checks (RESISTANCE.md falsifier #1–6 adapted):

• Visualization-justification check: the algorithm-viz answers a §C.1 product-page question ("how the lane works"). Not stripped.
• Hierarchy check: one dominant numeric (115 / 19 count-up); one viz; one blocker; one CTA. Hierarchy preserved.
• Typography check: no shared CSS / typography rules touched.
• Spacing/geometry check: card height unchanged from Wave-1 (703px @ 1200px). Layout contract holds.

9. Diff-scope assertion

• Template: edits confined to lines 487 + 495–498, all inside <article data-lane="polymath-ai">…</article> between line 477 and the closing </article> ahead of slot 05. No edits to the comment marker (line 476), the data-lane / data-portfolio / data-status attributes, the <span class="num">04</span>, the <span class="cat">…</span>, the <span class="stamp-block portfolio">AI</span>, or to any neighbouring article. Status flag stays populated (§A says fix is bounded copy, not status flip).
• No shared CSS / shared JS / scaffold / BaseLayout / page-order / portfolio-taxonomy edits.
• No new files in repo or worktree (_polymath-wave2-capture.mjs was a transient capture script, removed at end of session).

11. Smart-outsider headline check (re-dispatch step 2a, 2026-05-08T14:31)

Re-dispatch added a sharper standard: "the headline must make sense on first read to a smart outsider: it must say what the lane does, what it is trying to do, or what failed." Applied to the current Wave-2 headline An LLM to train on your phone.:

Outsider question	Headline answer	Pass
What does it do?	"An LLM" — an AI / large language model (widely known)	✓
What is it trying to do?	"to train on your phone" — infinitive of purpose; plain-English proposition	✓
Anti-pattern (internal nouns / governance labels)?	None. Lane-internal terms (ELO selective CPT, boundary layers, QNN/LiteRT, Phase 1A, REDMAGIC 10 PRO+) all correctly relegated to paragraph + viz + blocker, never to the headline. None of the brief-flagged trailing phrases (closed chain, owned pending, intake dossier, structural audit, authority packet, declared ground, each exact) appear.	✓
Reference DNA shape (subject-claim + accent + period)?	"An LLM" subject + <span class="accent">to train on your phone.</span> audacious accent + trailing period. Aligns with §C.1 references (Computation you can replay., A freighter that ships zero CO₂.).	✓

Outcome: the Wave-2 fix from earlier today already satisfies the re-dispatch step 2a standard. No further headline change. State unchanged since 2026-05-08 morning Wave-2 edit; PR #7 still OPEN, MERGEABLE, CodeQL × 3 SUCCESS, awaiting operator merge.

12. Re-dispatch end-of-turn line (2026-05-08T14:31)

Lane Polymath-AI: wave-2 outcome = verify-only (Wave-2 fix from morning satisfies new step-2a smart-outsider rule); repo/front-door status = R1 truth sync staged in PR #7 (MERGEABLE, CodeQL SUCCESS×3, awaiting merge); landing-card status = unchanged from morning Wave-2 edit (headline + paragraph already in bounded research-infrastructure tense, height 703px / under 720px ceiling); receipts ready.

10. End-of-turn line

Lane Polymath-AI: wave-2 outcome = fixed; repo/front-door status = R1 truth sync staged in PR #7 (MERGEABLE, CodeQL SUCCESS×3, awaiting merge); landing-card status = headline + paragraph tense softened per §B (2 surgical edits, height unchanged at 703px / under 720px ceiling); receipts ready.