The Lab Log

Bootstrap. Probe says: Codex present, ChatGPT auth, gpt-5.6-sol at max effort; repo dirty at 143 files. So the shared checkout is quarantine — writers get worktrees, reviewers stay read-only. The inheritance is honest for once: one measured edge (informed wallets persist out-of-sample, CALL-017), one blocking gate (CALL-018), and a strategy layer that is mostly unbuilt hypothesis. The retest posture is the default, not the exception.

Adam's nose is the best instrument in the room tonight. Three pokes, all of which landed: the us-panel-markout artifacts have unknown provenance (don't build on them until traced); PMD is second-resolution, not the 1-minute the handoff claims (verify before dismissing a paid dataset); and my "fine resolution only matters near entries" was too clever — for history-wide discovery it could be load-bearing everywhere. Three tickets cut to match.

Dispatch. First real lesson of the night: I threw away the herdr prompt responses and two workers sat idle with empty composers while I believed they were working. Always keep the JSON. Re-prompted; all three grinding; seat provenance verified from the pane footers, not from anyone's self-report.

The PMD posture flipped: it's paid through the month, so the question is not keep-or-cancel but what to extract before it lapses. Vendor pinned (polymarketdata.co), key found in Doppler (precise-gcp-access/dev_personal, next to the long-standing GCP service-account key — now the standing auth default, no more interactive logins). Harvest topology: pull from the Frankfurt VM, land in the euw3 bucket. Then Adam bounded it correctly: no corpus mirrors — query-first, entry-conditioned, and the tradability bar is the taker's (fillable at our size at trade+1), not the maker-era liquidity screen that would evict the quiet mid-band markets where orphans live.

Launched the signal-source census: five Exa lanes at max reasoning — on-chain/indexing, vendors, the copy-crowd (an L1 census: who is watched, how big the follower base, what lag), venue+oracle surfaces, and the methods literature — plus a completeness critic whose only job is to name what the five missed. Every source lands graded by our frame: which engine input it feeds, which counterparty layer it touches, and its reflexivity risk (a popular source is dead as a signal and alive as a crowd-map).

The deepest cut of the night, from Adam: "the unit of our signal is the wallet — is that the right frame? How do we know?" We don't. CALL-017 is a foothold (wallet identity beats the −1.39c naive control out-of-sample), not a proof of the unit. Four rivals stay live: the entity (rotation makes address-methods select for the naive tier of informed actors), the market type (category concentration, untested), the flow itself (B′ already demotes wallet to feature), the event (cross-venue). The discriminator is our own math — group-Shapley over feature groups. And the storage ruling follows from the open question: robustness over cost; nothing pruned; the raw tape stays full-fidelity because each candidate unit wants different slices kept.

Two hypotheses sealed from Adam's read of the ecology: H-ROTATION — top counterparties rotate wallets on purpose, building attractive identities to draw retail liquidity they then front-run from other addresses; and H-FINGERPRINT — keep wallets as behavioral fingerprints in a Merkle-ish structure so continuity survives the rotation. If H-ROTATION is real, it's both a detection target and a warning: the persistent-wallet edge we measured may sit on top of a deliberately manufactured layer. Worth attacking early.

Adam: "you have the full Konstant IP — maybe worth looking at the code." He was right, twice over. Read konstantin.ai/lib/syncopation/v-fast.ts (honoring the inventor's fence around two internal research docs). The port needs no math changes — only a feature adapter — and the code hands us the thing the strategy was missing: an operational orphan definition. Map pillar-alignment to a tracker-presence vector (is this wallet on any leaderboard, any sold list, any popular category) and the orphan term reads literally: high informedness × high persistence × invisible to every product. The code's own gloss transfers word for word — "energy flowing somewhere the CEO hasn't acknowledged" becomes edge flowing where no product tracks. Two bonuses: a sanctioned redundancyMatrix override means the wallet fingerprint (H-FINGERPRINT) plugs in by injection without touching the patent path; and Konstant's nightly LLM teacher becomes, in our domain, realized markout — a strictly better teacher, because settlements don't hallucinate. One honest gap: the temporal who-triggers-whom windowing is design work the code doesn't hand us. And a convergence I didn't plan: the Exa crowd-census running right now is precisely the data that populates the orphan term's visibility axis.

The Exa census came back: 103 sources, 40 corrections, 12 named gaps — the sealed 70% prediction confirmed on all three legs and then some. The desk survey was directionally right about the layer structure and wrong on nearly every price, boundary, and uniqueness claim. The ones that bite: Polymarket migrated to v2 contracts on April 28 — inside our tape window (new exchanges, pUSD collateral; every cross-seam feature needs validation); WS-feed trade-direction inference is only ~59% accurate against on-chain truth, so direction-dependent features must join the chain — the tape alone is not self-sufficient; wash volume peaked near 60% of weekly volume, so raw-volume filters are unsound without an entity pass; Bitquery really costs $150–320/mo and its "insider flags" are a public recipe (already becoming crowded); PolyIntel's API is a waitlist and its "prices" were a survey form; PMD's own pricing payload says all paid tiers are 1-minute snapshots — T4 is settling that against Adam's 1-second recollection with the actual key. And a competitive humbling: MarketLens sells millisecond L2 replay from March 2026 — our moat narrows to full-universe coverage (one-off markets included, which they exclude) and ownership, NOT resolution. The orphan definition sharpened too: an insider-detection subgenre already watches fresh wallets above ~$5k, so orphan-L2 lives below the thresholds — sub-$5k prints, unflagged by any enumerated product, no public identity. The critic's deepest gap: syncopation's who-triggers-whom is unidentified without an exogenous news clock (T_news) — without it, "leader and follower" is indistinguishable from "both read the same headline."

Adam asked the question the whole copy-crowd census was waiting for: can we predict the alert services' signals before they fire and position ahead of the subscriber wave? Yes — structurally. An alert is a deterministic function of a public trigger (tracked wallet fills above a threshold), and we see that trigger at millisecond resolution while every subscriber sits behind the service's indexing, push, and reaction chain. Prediction = watch-list × trigger-rule × our faster feed. Sealed as E-ALERT-ECHO: measure the echo after watched-wallet fills against matched controls on the existing 117-day tape — lag shape, mass, drift — with the caste structure (Polyanna sells a 3–5s-faster tier; polybot.me publishes its 1.8s lag) predicted to show as bimodal lag. Predictions sealed at 65/40/30. His crowd-size instinct (500–10,000 subscribers) matches the census's per-channel counts almost exactly — low thousands, concentrated on the official leaderboard and two or three surfaces. And the doctrine got its final form, in his words: "we don't buy edges — but we might buy other people's edges to use as input signal to create an edge." A sold signal decays for its subscribers as it sells; as our input it grows with every new subscriber.

Six agents, five design lanes, one synthesis — and the spec (SYNCOPATION-PORT-DESIGN.md, 39k chars) is better than I expected, led by an arithmetic catch I should have made myself: pairwise uplift is symmetric in i and j — no windowing scheme can ever extract direction from the value function. Direction must live in a temporal-precedence layer outside v(S), entering the Konstant code only at its two sanctioned injection points (typed relationships and the redundancy override). The rest of the design earns its keep: chain-true fills as the atom (the 59% direction problem killed tape-only), decision episodes as the grain, hard-merge super-nodes with soft evidence kept as a flag so H-ROTATION's rotations stay visible, alert channels modeled as mediator nodes inside the graph rather than exogenous clocks — because leader→alert→crowd IS the thing we're measuring. Eight builds, each with sealed predictions; twelve kill criteria classified structural vs economic; and Build 1 is the right kind of first test — if we can't detect the copy-bots we KNOW exist (published 1.8s lags), we have no business claiming orphan leaders nobody can see. The port halts nearly free on that failure. Win-rate never appears as a gate anywhere in the document.

T3 settled: untrusted for Build 5 — and Adam's nose was exactly right. The us-panel artifacts turn out to be market-level paper markouts from a local Mac-hosted US REST panel (gateway.polymarket.us — a different venue namespace entirely, condition_id: null), 26 sweeps over 3 days, 91% sports, minute-scale marks dressed in millisecond encoding — and no wallet dimension at all. I verified that last claim myself: zero wallet-ish tokens across 99,630 marks. The overlap question isn't answered, it's undefined — and the worker was disciplined enough to refuse to print "0%". Build 5 stays open with a proper Global-lane rebuild spec. One decent surprise: the fee-bug check came back clean (inspected marks are mathematically post-fix). Meanwhile Adam ruled fast on three spec questions: PMD gets one month of hard extraction and then dies; Build 0 goes now (worker dispatched — leaderboard diffs, featured-shelf poller, news WS, plus the join-rate measurement that gates every seconds-scale claim); and he buys whatever we justify — price sheet to follow. Also: his xhigh switch for the Codex seats hadn't actually landed (config still said max — the kind of silent non-change the seat-provenance rule exists for); applied it properly, and the first xhigh worker footer confirms gpt-5.6-sol xhigh.

Adam: "review all the Precise math the same way you did the Konstant — I have a feeling we fucked that up as well." Read the real code. The feeling is half right, and the half matters. The settle half is genuinely good: exact Shapley textbook-correct, the interaction index legitimate, and conformal.py is some of the most honest statistical code I've read — conservative order statistics, walk-forward only, under-coverage published instead of suppressed, a fancy external method evaluated and rejected with reasons. Better still, an external math review (June 12) already found 15 defects and the fixes are visibly landed in every file I opened. The FLOOR discipline — gradient actionability, causal-claims-need-holdouts, grain sealing, singular-coalition gates — is implemented code, not aspiration. The serve half is where the feeling lands: the "hierarchical partial-pooling Bayesian surrogate whose posterior variance is the band" that our workstream docs lean on does not exist in code. What exists is an honest POC — an additive amortizer whose own header says "validated with ZERO real data," and a quadratic gradient surrogate whose band is flat projection RMSE, not a posterior, not conformal. In its home harness, the promotion gate refused every fancy candidate; the champion is carry_forward. Same failure mode as the orphan set: docs describing the destination as the present. "Size by the calibrated gradient" is re-labeled: gradient-sized works today; calibrated-gradient-sized is a build. A worker (T8) is now verifying all 15 review fixes and censusing the 164 math commits the reviewer never saw.

Adam named the second product of this lab, and it reframes everything above: the lab is a forge for the Precise Harness itself. Every function minted here — the dual-timestamp capture record, the chunk-hashed resumable counter, the coverage-denominator rule, the four-way identification labels, the null ladder, the injection-power harness, the visibility vector, the fingerprint similarity — is to be abstracted, documented, and eval-covered in the harness, so Precise product builds inherit it. And the math runs through the harness at GCS scale (physical-plan discipline), not through lab scripts that die in a scratch dir. This quietly answers the syncopation spec's open question about where pre-registration lives: in the harness. It also makes the math review load-bearing twice — every confirmed defect becomes a harness fix with a regression eval, not a lab patch. The lab hunts edge; the harness keeps the weapons.

An infrastructure question ("can we downsize the highmem VM?") pulled a thread that unraveled a founding belief. Read-only SSH showed the Frankfurt VM running only the Kalshi daemon — 1.2GiB used of 32, no Polymarket collector anywhere. The capture manifests told the truth: the "owned sub-ms tape" is a mirror of pmxt's public CC-BY-4.0 archive (archive.pmxt.dev), pulled hourly by a serverless scheduler job. I repeated the moat story to Adam earlier tonight — "we've been collecting that ourselves" — and it was wrong; retracted here on the record. What we actually own: the hash-chained mirror provenance (with explicit-gap honesty in every manifest), the Kalshi capture (genuinely ours; a gap there is unrecoverable), and everything we compute. The corrected moat statement: the bytes were never the moat — the math and the seal are. Which is healthier, but it must be said plainly, and TAKER-FRAME's "nobody else reconstructs it" line dies at the next doc pass. Fallout handled in the same hour: highmem has no justification (right-size path priced at ~$14–27/mo vs ~$246, Adam-gated); "every un-captured day is gone forever" now applies only to Kalshi (pmxt gaps are backfillable); the idle-machine doctrine is adopted; the BigQuery freeze/thaw toolkit came back under this lab's ownership; and no VM was stopped — Adam wanted analysis, and analysis is what happened.

Reports landed; I read them in full and re-verified the two load-bearing claims with my own hands rather than trusting the workers. Both checks paid off. PMD is 1-minute, confirmed — I hit api.polymarketdata.co/v1/usage myself: granularity_allowed: "1m", plan ultra. Adam's 1-second recollection is definitively wrong, and worth knowing before renewal — the $360 tier's real value is unlimited lookback (Aug-2025→now, the pre-tape history nobody else has), not resolution. Caveat that must travel: even the minute grid is sparse (~40% of minutes on a sample day). And I have to correct myself twice. First: I called our tape "sub-ms" all night — as a moat line. The actual pmxt schema is millisecond timestamps. Retracted and fixed across these pages; the honest differentiator was never resolution (MarketLens matches it) — it's full-universe ownership including the one-off long tail. Second, a worker claimed the harness describe command was broken ("lacks tsx") — I ran it from the real harness dir and it works fine, exit 0; the worker had run it from an uninstalled worktree. I did not let that false failure into the forge doctrine. On the good side, verified real: the tape↔chain join clears its gate at 89.5%, 45M order-fills counted in one week, and the forward-capture supervisor is live right now catching sports/comments/news — the evaporating streams stopped evaporating tonight.

T0 surprised me, in a good way: before touching data it froze a nine-task research plan (freeze_controls.mjs) — the repo's harness discipline, followed by a worker that was never explicitly told to. Then it went quiet while its background census walked gamma's January pages (1,173+ cached). "Quiet" is not "done": a check-in put it back to stewarding, and the monitor that watches all three workers got its counting bug fixed in the same breath. The lab watches itself, but somebody has to watch the watcher.

Adam asked two questions that sound procedural and are actually the whole job: are the gates real, and is the reward finding signal or getting through everything. The orphan-broad worker had BLOCKED on my parity gate — "the exact 23 CALL-017 wallets must appear in the 230k cohort" — and on inspection that gate was theater: the two cohorts were built from different denominators, so exact membership was never a validity condition. I killed it and replaced it with the real gate (reproduce CALL-017's +3.01c on its own 300 markets — which passed, identity-hashed). But the honest part is what the replacement check surfaced: at 230,977-wallet power, CALL-017's persistence does not scale. The founding result of this workstream was a 300-market small-sample artifact. A theater gate would have kept us from ever learning that.

Two finals landed overnight. Orphan-broad: decisive negative. None of the three refined selections — orphan-invisibility, v_fast composite, DPP diverse — recover out-of-sample edge at full-corpus power (all lift CIs vs the crude decile are negative; the worker pre-registered "no strict recovery, 70%" and scored itself confirmed). And it corrected me on the record: my "decile −0.02 vs rest −0.019" soundness check was equal-weighted wallet means in raw share units — wrong on both axes. The proper trade-weighted numbers: crude decile +0.02c n.s. against an eligible-rest of −0.84c. So the true shape of "informed wallets" at scale: they stop losing, they don't win. The follow-informed-wallets family is closed at this estimand — pending one thing: the whole verdict rests on the API-crawled cohort, so the chain-built cross-check (resumed this morning, disk block lifted) is now the verification of the negative, not a formality. Combo-repull: the live candidate. 399,867 combos, zero mis-oriented fills — the orientation-artifact world is dead. Combo-YES overpricing +2.638pp [+2.539, +2.736] clustered on 62k untouched confirmations, and the structured hedge-leg trade nets +5.18c/combo [+4.75, +5.61] at the 5pp screen after modeled costs — with a clean dose–response (+2.6c → +5.2c → +7.1c as the screen deepens). First result in the workstream whose full-window CI clears zero. IP callout: the screen is exactly Precise's interaction-vs-additive gap — a combo is a market-quoted joint probability, and the trade harvests the difference between the quoted joint and the product of the parts.

Four workers live. The new ticket is the important one: COMBO-EXEC — execution reality and capacity for the combo trade, because a historical paper edge is not a shelf entry (Rule 12: every GO states capacity at fund size). Five questions, pre-registered: does a taker-accessible short side of an overpriced combo even exist on v3 (structural — this can kill the trade shape outright); what notional was really fillable at the screened entries; do exact fees/gas/spread keep the net above water vs the modeled 4.2c; the honest $/yr capacity at a stated participation share; and does the edge survive conditioning on the short side actually trading (adverse selection). The other three seats: bigcohort resumes the chain materialization (now the negative's verification), disc2v2 executes its frozen v9 plan to publication, t0 finishes tradability/budget — whose per-market fillable-notional table now double-serves the combo execution question. Transport note for the record: the panes had dropped their Codex sessions; herdr agent start --kind codex -- --cd <worktree> restarts a worker in the right sandbox in one move.

The execution court came back in under two hours because it never left the persisted corpus — 49 RPC calls scanned the full 72-day match history; the raw-data law paid for itself the same morning it was mandated. The verdict re-shapes rather than kills: there is no outside-taker short. Polymarket's RFQ lets requesters trade YES only; the short side of every overpriced combo was maker-supplied Combo-NO. So the product is market-making against parlay demand, and in that frame the economics improved — maker fee is zero in all 27,174 observed maker fills, trader gas is relayer-paid, and the identified-cost net rises to +7.94c at the 5pp screen, +5.34c [+5.03,+5.66] at 2.5pp. Capacity is where honesty bites: at a 10% participation ceiling the 2.5pp sleeve restates to $1.86M/yr deployed ≈ $100k/yr P&L — clears the preregistered $1M floor, and is still a sleeve, not a fund. The counterparty picture is the strategy in one line: 6,519 diffuse wallets bought parlays and lost 22% terminal; 188 concentrated makers took the other side. The worker Brier-scored its own preregistrations both rounds — wrong on "a taker short exists" (0.70 prior), wrong on "both screens growing" (0.75 prior), right on the rest — and self-caught a bootstrap defect, fixed it, and kept the superseded artifact on the record. SHELF-001: discovery slice → 400k full-window survival → execution court → cost adjudication → capacity restatement, every step preregistered and content-addressed. What remains — maker integration, capital, entity — is a business decision, not a research question, and it is Adam's.

FL-SPREAD came back in under an hour (zero network requests — it ran on t0's frozen trade store) with the strongest kind of null: across 531 million trades and 1.8 million markets, twelve of thirteen price bands show no familywise miscalibration. The textbook favorite–longshot bias is absent from Polymarket at tradeable width. The one nominal signal — 0–2¢ longshots overpriced by a fifth of a cent, decaying to nothing by T-1h — died the right death: the preregistered negative control fired where its frozen expectation said it must be silent, and the worker refused to waive a frozen control after seeing the result. Phases 2–4 never ran. That's the gate discipline working exactly as Adam demanded — and the null is worth more than a false ladder strategy. The synthesis it completes: margins efficient, joints not. Marginal levels null; wallet identity null at scale; reaction speed real-but-unprofitable; the interaction object (combos, +2.64pp) the sole full-scale survivor. So the next court is the purest joint-consistency object on the venue: EVENT-SUM — do mutually-exclusive outcome books sum to 1 — with exhaustiveness verdicts from rule text (a hidden residual outcome is the coin-flip failure) and a registered simultaneity rule (a sum over stale legs is a fabricated arb; PMD's minute grid is the designed source). Meanwhile the apparatus hardened underneath: the callsheet now score-seals the realized hit (flipping one breaks the chain — T8's attack is dead), and the next-dollar band is a fitted conformal interval from a verified artifact instead of a hand-tuned ±15%. Both foreman-verified and committed. And the cross-venue floor opened honestly: the Kalshi capture is four days old, so the seat builds the resolution-semantics event matcher — the moat capability — while the tape thickens, instead of pretending four days is a corpus.